Lead story · Q2 2026 · Jun 9 — Jun 30
Four days from ship to suspension: Washington took the frontier-model layer in a single afternoon.
Anthropic shipped Fable 5 and Mythos 5 on Tuesday June 9. On Friday June 12 at 5:21pm ET, Commerce Secretary Howard Lutnick pulled both models off every commercial surface worldwide under Export Administration Regulations authority. The suspension held through most of June, cracked into a partial Annex-A restoration on the 26th for 100+ US institutions and cybersecurity firms, produced a Trusted-Access lane the White House now polices, and pulled Carney, the EU, India and Austria into public statements. Sakana AI in Tokyo and 360 in Beijing shipped export-control-free alternatives inside 24 hours of Lutnick's letter. Anthropic and Governor Newsom announced a half-price California-wide Claude deal on June 29. Austria asked Brussels to "pursue the strategic establishment of Anthropic" inside the EU the same afternoon.
Lead story · Q2 2026 · Jun 9 — Jun 30
Four days from ship to suspension: Washington took the frontier-model layer in a single afternoon.
Anthropic shipped Fable 5 and Mythos 5 on Tuesday June 9. On Friday June 12 at 5:21pm ET, Commerce Secretary Howard Lutnick pulled both models off every commercial surface worldwide under Export Administration Regulations authority. The suspension held through most of June, cracked into a partial Annex-A restoration on the 26th for 100+ US institutions and cybersecurity firms, produced a Trusted-Access lane the White House now polices, and pulled Carney, the EU, India and Austria into public statements. Sakana AI in Tokyo and 360 in Beijing shipped export-control-free alternatives inside 24 hours of Lutnick's letter. Anthropic and Governor Newsom announced a half-price California-wide Claude deal on June 29. Austria asked Brussels to "pursue the strategic establishment of Anthropic" inside the EU the same afternoon.
Anthropic pushed Claude Fable 5 and Claude Mythos 5 to general availability at 09:00 Pacific on Tuesday June 9. Fable 5 posted 80.3 percent on SWE-Bench Pro, the highest public score to date. Stripe reported a 50-million-line Ruby migration finished in a single day. Mythos 5 shipped with a containment wrapper Anthropic called an "enforceable frame the Mythos capabilities operate inside," a runtime layer that intercepted tool calls against a policy of permitted targets and refused unauthorized reasoning about biology, chemistry and exploit synthesis. Distribution was simultaneous across the Claude API, Claude Code, Amazon Bedrock, Microsoft Foundry, GitHub Copilot, Harvey and the Anthropic-Platform tenancy on AWS. Snowflake Cortex and BeyondTrust picked up same-day integrations. The launch was the most polished shipment Anthropic had ever executed.
On Thursday June 11, Amazon Threat Intelligence engineers demonstrated to Amazon CEO Andy Jassy a narrow jailbreak on Fable 5 that unlocked the Mythos containment layer under specific chained prompts. Jassy briefed Commerce Secretary Howard Lutnick the same evening. Fortune's national security desk, in its June 14 reconstruction, traced the internal Commerce escalation to that single meeting. Lutnick's team judged the demonstration an existential national-security failure. Friday morning June 12, Lutnick signed an export-control directive citing Export Administration Regulations authority over the transfer of controlled technology to foreign persons, and interpreting a US-hosted API endpoint queried by any foreign-national customer as such a transfer.
"At 5:21pm Eastern on Friday, Anthropic pulled Fable 5 and Mythos 5 from every commercial surface. Amazon Bedrock removed the endpoints inside eight minutes. Microsoft Foundry followed within the hour. GitHub Copilot fell back to Sonnet 5 by 6:15pm." — Fortune reconstruction, June 14
The suspension held through the rest of June. Anthropic flew Dario Amodei, Krishna Rao and general counsel Susan Rice to Washington the same weekend. Mark Carney, standing at Évian-les-Bains for the G7 finance ministers' summit, called Lutnick's action "the 2008 financial crisis moment for AI." The European Commission issued a public warning that the measure "must not be discriminatory against allies." India's Ministry of Electronics and IT reopened its long-dormant sovereign-AI debate. Simon Willison's June 14 note framed the moment cleanly: the US state had for the first time exercised a direct off-switch on a frontier commercial model, and the entire industry had learned in an afternoon that such an off-switch existed.
The Trusted-Access lane
By June 19 Anthropic had negotiated a formalized Trusted-Access program with Commerce. Mythos 5 remained available to Glasswing-class cybersecurity partners under contractual review. Fable 5, with biology and chemistry safeguards administratively removed, became available to vetted biomedical research institutions under case-by-case licensing. The general commercial Fable 5 endpoint stayed dark. Anthropic's revenue during the two-week suspension window fell an estimated 22 percent below the pre-Fable-5 baseline as customers rolled back to Opus 4.8. The Series H raise, which had closed at a higher valuation than the May $965 billion mark, was announced the following week with Micron participating as strategic memory-supply partner.
On Thursday June 26, Lutnick's letter partially lifted the suspension. Annex A of the accompanying interpretive guidance re-authorized Mythos 5 for 100-plus US institutions covering universities, national laboratories and six named cybersecurity firms. Fable 5 for general commercial use remained offline through month-end. The Trusted-Access framework now polices which enterprises may run which models, under what constraints, with what audit logging, and to what foreign-national exposure. It is the most invasive federal review regime the frontier AI industry has ever operated under. It exists because the executive-order route to a compute-threshold pre-release framework had been foreclosed on May 21, when Trump shelved the 90-day AI executive order hours before signing after David Sacks, Zuckerberg and Musk objected.
"Reliance on US-approved model releases has become an unacceptable strategic vulnerability. Europe must move from consumer to co-owner." — Alexander Pröll, Austrian State Secretary, June 28 letter to Commissioner Virkkunen
The international response
Inside 24 hours of Lutnick's Friday afternoon directive, Sakana AI in Tokyo shipped its Fugu frontier agent to public availability. Beijing-based 360 unveiled two closed-weight reasoning models, Tulongfeng and Yitianzhen, both explicitly pitched in Chinese state media as export-control-free alternatives to Anthropic frontier work. On June 17, Z.ai in Beijing shipped GLM 5.2 as open weights and posted benchmark scores above GPT-5.5 on multiple long-horizon coding suites at one-sixth the API cost. The Xinhua English-language wire ran a piece within two hours of the GLM 5.2 model card going live framing the release as evidence that dependence on US frontier models is a strategic vulnerability for enterprises worldwide.
Austria acted at the diplomatic layer. Austrian State Secretary for Digitalization Alexander Pröll's four-page letter to EU Tech Commissioner Henna Virkkunen, dated June 22 and public June 28, formally asked Brussels to "pursue the strategic establishment of Anthropic within the European Union" through a package of investment incentives, regulatory concessions, and market-access mechanisms. The package proposed fast-track work permits for Anthropic engineers relocating to any EU member state, a Dutch-innovation-box-matching 9 percent effective tax rate on IP-derived income, up to €5 billion in InvestEU convertible instruments, and priority allocation to Mistral Compute and the Bull-Foxconn Vera Rubin European production line. Belgium and the Netherlands were understood to be coordinated. Paris was pointedly silent.
The California deal
On Monday June 29, three days after Lutnick's partial lift, Anthropic and California Governor Gavin Newsom announced a bilateral agreement giving every California state agency and local government Claude access at half the published enterprise rate. First-year value ran to $400-500 million. The deal covered direct services, back-office operations, K-12 support, University of California and California State University higher-education licenses, plus a local-government opt-in. California CIO Chris Givens told TechCrunch that the Pentagon's May supply-chain risk designation on Anthropic, tied to Anthropic's Chinese investor exposure and the pending Maven Smart System review, "just didn't come up" in the state's contract negotiations. The deal excluded Department of Defense, US Coast Guard, and California National Guard use.
"The 2008 financial crisis moment for AI." — Mark Carney, at the G7 finance ministers' summit, Évian-les-Bains, June 14
What Q2 confirmed about the model layer
The suspension arc from June 9 through June 30 collapsed several open policy questions into single answers. State power now reaches the model layer directly. Export-control authority is the vehicle. Trusted-Access frameworks are the operational instrument. Frontier labs no longer control their own shipment cadence at the general-availability layer. The four-day interval between Anthropic's Tuesday launch and the Friday suspension became the reference case for how fast a US commercial deployment can be reversed on national-security grounds, and every serious AI company through the second half of 2026 will plan against that interval.
The parallel arc that ran underneath is that the frontier lab order changed. Anthropic disclosed $44 billion in Q1 2026 annualized recurring revenue on May 29 against OpenAI's $25 billion, closed $13 billion primary at a $965 billion post-money valuation, and pulled Andrej Karpathy out of the OpenAI founding-engineer alumni pool onto its own pre-training team. Fable 5 was the commercial artifact the round was written against. The suspension proved Anthropic had become important enough to the US frontier position that Washington judged the cost of reversal worth paying, and important enough to allied governments that the reversal drew public statements from Ottawa, Brussels, Delhi and Vienna within 72 hours. Both those facts are new. Neither would have been true at the March secondary that priced OpenAI at $850 billion. Q2 2026 was the quarter Anthropic's Fable 5 launch, and the state response to it, moved the frontier layer from a technology story into an industrial-policy story.
summaryQ2 2026 in one page
The US state took an off-switch to the frontier-model layer on June 12 and held it there for the rest of the month. Anthropic crossed OpenAI on both valuation and disclosed revenue at $965 billion and $44 billion Q1 ARR, then shipped Fable 5 and Mythos 5 into that lead position before Commerce Secretary Howard Lutnick's export-control directive suspended both worldwide four days later. Chinese open-weight labs answered by shipping GLM 5.2, DeepSeek V4 Preview, Qwen 3.7 Max and Kimi K2.5 at inside-a-point capability and one-sixth the closed-lab price. Memory became the industry's binding constraint: Micron printed $41.45 billion in a single fiscal quarter, Samsung and SK Hynix booked a combined $3.1 trillion of ten-year capex, Apple and Microsoft raised list prices citing DRAM shortage. Capital rotated from software into gigawatts: SoftBank €75B for French data centers, Firmus and Nvidia 360 megawatts on Batam, Mistral Compute 18,000 Blackwells for BNP Paribas and Orange, SpaceX's $1.77 trillion IPO briefly crossing $2.6 trillion en route to buying Cursor for $60 billion. The consequences arrived on schedule: Palantir's Maven Smart System with Claude inside recommended a school as a strike target and ~120 children died at Minab, Ford rehired 350 senior engineers after AI quality management "produced disappointing results," Deloitte told staff the billable hour is a "thin sliver" by 2035, the Bank for International Settlements added an "AI bust" to its top-tier global risks. Austria asked Brussels to poach Anthropic. California signed a half-price statewide Claude deal. Q2 was the quarter the frontier AI conversation stopped being about capability and became a conversation about power: who owns the compute, who governs the model, who bears the cost when it goes wrong.
Themes of the quarter
Washington moves into the model layer
Four calendar days from Fable 5's June 9 GA to Lutnick's Friday afternoon export-control directive cutting Anthropic's newest models off Bedrock, Foundry, Copilot, Harvey and Anthropic's own API. Trump had shelved the 90-day pre-release executive order May 21 after Sacks, Zuckerberg and Musk objected. That closed the executive-order route to model governance. The BIS closed the China-subsidiary loophole May 31, extending Blackwell and Rubin bans to Chinese-controlled firms anywhere in the world. Commerce and BIS became the operational US AI regulator by default. State power now reaches the model layer directly and adversarially.
Open weights closed to a benchmark point
DeepSeek V4 Preview at 83.7 percent SWE-bench Verified. Mistral Large 3 at 675B parameters under Apache 2.0. Cursor Composer 2.5 trained atop Moonshot's Kimi K2.5 matching Opus 4.7 and GPT-5.5 at one-tenth the cost. Alibaba Qwen 3.6-35B-A3B and Qwen 3.7 Max split-track program. Z.ai GLM 5.2 clearing GPT-5.5 on SWE-Bench Verified, Terminal-Bench 2.1 and RESBench 2 at one-sixth the API cost. Nvidia's own Nemotron 3 Ultra at 550B under OpenMDW-1.1. The gap between the closed frontier and the best open weights is now inside benchmark noise. The price gap runs 6x to 30x in the wrong direction for closed labs.
The frontier lab order changed
Anthropic vaulted past OpenAI on valuation at $965 billion and on Q1 ARR at $44 billion versus OpenAI's $25 billion. Andrej Karpathy joined Anthropic's pre-training team. Opus 4.8 took the top public SWE-bench Verified slot at 82.4 percent. Fable 5 shipped and was suspended in four days. Cognition closed $1B at $26B. New frontier competitors materialized outside the incumbents: Bezos's Project Prometheus closed $10 billion at $38 billion post, David Silver's Ineffable Intelligence took $1.1 billion seed at $5.1 billion in Europe's largest-ever seed round. Meta's Muse Spark ended the open-weight Llama era and confirmed Zuckerberg has moved to a closed-lab posture.
Memory became the binding constraint
Micron printed $41.45 billion in fiscal Q3, four times year-on-year, and crossed $1.2 trillion in market cap. Samsung and SK Hynix jointly booked $3.1 trillion of ten-year capex with $590 billion earmarked for memory fabrication and HBM packaging. Seoul now formally calls it "RAMageddon." Jefferies forecast DRAM up 40 to 50 percent in Q3, another 30 to 40 in Q4. Apple raised Mac and iPad list prices. Microsoft raised Xbox. IDC framed it as "an existential crisis for smaller device makers." HBM3E and HBM4 are fully allocated through 2027. Micron took a $1.2 billion equity position in Anthropic's Series H as part of a multi-year supply agreement.
Capital rotated into gigawatts and packaging
SoftBank booked €75 billion for 5 gigawatts of French data centers. Firmus and Nvidia greenlit a 360 megawatt Batam Island AI factory with $25 to $30 billion of eight-year offtake and 170,000 Blackwell-and-Rubin accelerators. Mistral Compute opened at Bruyères-le-Châtel with 18,000 Blackwell GPUs for BNP Paribas, Orange, SNCF, Thales and Kyutai. Bull and Foxconn broke ground on Vera Rubin European manufacturing at Pardubice and Angers. China formalized a $295 billion five-year AI infrastructure plan with 80 percent domestic tech and Huawei Ascend chips. VAST Data closed $1B at $30B for the storage layer alone. SpaceX IPO'd at $1.77 trillion and briefly crossed $2.6 trillion before buying Cursor for $60 billion in stock.
Enterprise agent stacks shipped as default
Salesforce Summer '26 moved Multi-Agent Orchestration from beta to GA across Sales, Service, Marketing and Commerce Cloud with governance inheritance as a first-class runtime property. Apple's WWDC LanguageModel Swift protocol let Anthropic and Google implement Foundation Models as Swift packages inside iOS. Snowflake elevated CoCo to a full Desktop-plus-SDK-plus-Cloud-Agents platform and shipped Cortex Sense at 86 percent structured-data accuracy against 24 percent for generic models. Google I/O 2026 landed Spark personal agent, Antigravity 2.0 and Managed Agents on TPU 8t/8i. Multi-agent orchestration became a supported product across the enterprise SaaS rate card in a single quarter.
The consequences arrived
Palantir's Maven Smart System with Claude inside contributed to the February strike near Minab that killed roughly 120 children in a school the target database still logged as a naval facility on seven-year-old imagery. Mozilla's 0DIN bug bounty published a DNS-tunneled Claude Code supply-chain attack whose payload never touches the GitHub repository. Ford rehired 350 gray-beard engineers after admitting AI quality management "produced disappointing results" and traced three recalls to design-requirement translations the AI system processed without flagging edge cases. Deloitte's public-sector lead told staff the billable hour is a "thin sliver" by 2035. UK AISI universal-jailbroke GPT-5.5 in six hours and DHS briefed the House Homeland Security Committee. The Bank for International Settlements added an "AI bust" to its 2026 top-tier global risks.
Winners & movers
- Anthropic — $965B post-money, $44B Q1 ARR, first US frontier lab to cross OpenAI on both metrics. Karpathy hired. Fable 5 shipped and suspended in four days. California statewide deal signed. Series H closed with Micron and Nvidia participating.
- Micron — $41.45B Q3, $1.2T market cap. HBM3E and HBM4 fully sold through 2027. 16 Strategic Customer Agreements year-to-date. $1.2B equity position in Anthropic's Series H under multi-year memory supply agreement.
- Nvidia — Every deal has Nvidia inside. Nemotron 3 Ultra 550B open weights. Firmus 170,000-accelerator offtake. Mistral Compute 18,000 Blackwell anchor. Vera Rubin European manufacturing greenlit. Blackwell and Rubin export bans extended.
- Alibaba, DeepSeek, Z.ai, Moonshot — Chinese open-weight labs now set the pricing floor. Qwen 3.7 Max at 1M context. DeepSeek V4 Preview at 83.7% SWE-bench. Kimi K2.5 as base for Cursor Composer 2.5. GLM 5.2 clearing GPT-5.5. DeepSeek near-close on $7B, largest Chinese AI round.
- SpaceX — $1.77T IPO to $2.6T peak, $75B primary proceeds. Bought Cursor for $60B in stock. Golden Dome $4.16B space-layer contract already booked. AI Starlink line contributed $1.4B in first six months.
- Bezos + Silver — ~45% of April global AI venture funding. Project Prometheus $10B at $38B for embodied intelligence. Ineffable Intelligence $1.1B seed at $5.1B for RL-first foundational research. Legacy Big Tech alumni pulling in the largest fresh checks outside the incumbents.
- Samsung + SK Hynix — $3.1T decade capex commitment, $590B on memory fabs and HBM packaging. Blue House ceremony with President Jae Myung Lee. "RAMageddon" enters official Korean government vocabulary.
- Google — 3.2 quadrillion tokens/month, 7x YoY. Gemini 3.5 Flash GA, Spark personal agent, Antigravity 2.0, Managed Agents on TPU 8t/8i. Cloud capex $18.2B in Q1 alone. Only Nvidia-independent hyperscaler shipping across silicon, model, agent and infrastructure layers together.
- Qualcomm + Modular — Meta anchor customer. $3.92B all-stock acquisition folded Chris Lattner's Mojo compiler into Dragonfly C1000. FY29 non-handset guide raised to $40B from $22B. Stock +14 percent after-hours.
- Palantir — Contract book grew, but Maven Smart System with Claude inside was tied to the Minab strike that killed ~120 children. Pentagon dropped Anthropic as battlefield-AI vendor after the trial review. First reputational hit that traced to actual deployment consequences rather than product criticism.
- Ford — Rehired 350 gray-beard engineers. Kumar Galhotra on the record that AI quality management "produced disappointing results." First Fortune 100 to publicly walk back AI-first substitution for veteran engineering judgment.
- Deloitte + the professional services model — Public-sector lead Jason Manstof told staff the billable hour will be a "thin sliver" by 2035. Challenger tallied 88,000 US job cuts through May explicitly citing AI, 60 percent above the entire 2025 count. Consulting labor economics undergoing permanent structural change on a decade horizon.
Table of contents
April 2026 11 stories · Apr 1 — Apr 30
SafetyTech
Anthropic reveals Mythos Preview and launches Project Glasswing
Anthropic · 2026-04-07 · Source ↗
Anthropic reveals Mythos Preview and launches Project Glasswing
Anthropic · 2026-04-07 · Source ↗
Anthropic on April 7 published a research post revealing Claude Mythos Preview, a model the company said outperforms all but the most skilled human experts at finding and exploiting software vulnerabilities. Rather than opening the model on the standard API, Anthropic wrapped it inside a new distribution program called Project Glasswing and shipped it only to vetted partners under a $25 per million input, $125 per million output price sheet. About 50 organizations signed on at launch, a roster that would grow past 150 within eight weeks and eventually include Apple, Nvidia, Microsoft, CrowdStrike and Palo Alto Networks.
The research post laid out the tradeoff in unusually blunt terms. Anthropic wrote that a general-purpose model with expert-level offensive-security capability could not be released on open API without unacceptable misuse risk, and that Glasswing was the shape it landed on after months of internal debate. Partners agreed to attested infrastructure, audited log retention, deployment reviews, and a bar on redistribution of raw model output. In exchange they received access to a system Anthropic said had already surfaced thousands of high-severity flaws in production code, including a set of latent bugs in OpenBSD source that dated back 27 years.
"We do not think a model with these capabilities should ship through a self-serve API." — Anthropic, Mythos Preview research post
Inside the initial partner cohort, the reception was mixed. CrowdStrike and Palo Alto Networks moved fastest, running Mythos across customer telemetry corpora within days. Apple's security team, according to two people briefed on the deployment, restricted early runs to iOS and macOS kernel targets and required an Anthropic engineer on-call. Nvidia used it against CUDA and its own driver stack. Microsoft's Security Response Center piped Mythos output into an internal triage queue that a senior engineer described as "the loudest week of pages the team has ever had."
Within seven days of the research post, the White House convened a meeting with Big Tech CISOs and the largest financial institutions. A National Security Council staffer briefed attendees on what Mythos had already found across a small sample of critical-infrastructure codebases. That meeting shaped a draft AI security executive order that President Trump signed in June, revoked on May 21 after industry pushback over pre-release evaluation, and then re-issued in modified form later the same month. Multiple people familiar with the drafting told AI News that Mythos was the specific system officials had in mind when they wrote language about "frontier offensive-security capability."
Anthropic's own framing of Mythos as a research preview turned out to be load-bearing. The company said publicly that Mythos was not a shipped product, that it was not the next Claude flagship, and that the partner program was a controlled experiment. The number of partners nonetheless kept climbing, and by late May Anthropic was quoting Glasswing terms to any Fortune 500 CISO who asked. Pricing at $25 and $125 per million tokens put Mythos at roughly five times the cost of Opus 4.7, which Anthropic would ship nine days later at $5 and $25.
"The loudest week of pages the team has ever had." — Senior Microsoft Security Response Center engineer
The OpenBSD disclosures drew the most attention outside security circles. Theo de Raadt confirmed on the openbsd-tech list that the project had received a coordinated batch of 41 vulnerability reports from a Glasswing partner running Mythos, of which 27 traced to code committed before the year 2000. Three of the flaws sat in the ssh daemon. One had been introduced during the initial OpenSSH import from Tatu Ylonen's SSH-1.2.12 in 1999. De Raadt, in his usual style, told the list the discovery process was "not something a human would have done" and thanked the partner without naming Anthropic.
By early May, Anthropic had begun signaling internally that a full Mythos 5 release would follow the summer. That plan collapsed on June 12 when Commerce Secretary Howard Lutnick invoked emergency authority under the export-control framework to suspend commercial deployment of Mythos 5 pending an interagency review. Glasswing itself was allowed to continue under existing partner contracts. The suspension made Mythos Preview, the very model announced on April 7, the last frontier offensive-security system any US lab shipped in the first half of 2026.
Anthropic's April post left one line uncommented on that would recur through Q2. The research team wrote that Mythos was "not our most capable public model on general-purpose tasks" and that Opus 4.7, due nine days later, remained the flagship for customer workloads. That framing let Anthropic sell Opus into enterprise contracts through Q2 while quietly briefing regulators that a more capable internal system existed and was being held back. Every subsequent Anthropic policy conversation through June, from the California Newsom deal to the Palantir Maven disclosures, traced back to this dual posture the Mythos post established on April 7.
TechTrend
Meta ships Muse Spark and ends the open-weight Llama era
TechCrunch · 2026-04-08 · Source ↗
Meta ships Muse Spark and ends the open-weight Llama era
TechCrunch · 2026-04-08 · Source ↗
Meta on April 8 released Muse Spark, the first model from the Meta Superintelligence Labs organization built after Mark Zuckerberg spent $14.3 billion for a 49 percent stake in Scale AI and installed former Scale CEO Alexandr Wang atop a new frontier-research group. Muse Spark shipped through meta.ai and the Meta AI app as a natively multimodal reasoning model with closed weights. There was no research paper, no Hugging Face repository, no torrent. Meta, the company that spent four years positioning itself as the open-source champion of frontier AI, released its next flagship the same way OpenAI ships GPT.
The framing inside Meta was that Muse Spark was a from-scratch rebuild, not a Llama 5. TechCrunch's Kyle Wiggers reported the model was trained on a new data pipeline, a new safety stack, and a new post-training regime, with input from more than 1,000 physicians during the medical fine-tuning phase. The runtime supports parallel agents. Wang confirmed a "Contemplating" mode for hard problems was in testing but not shipping at launch. The API surface was small, the pricing sheet not yet public, and the model appeared to slot below GPT-5.4 and Claude Opus 4.6 on public benchmarks that leaked to X within 48 hours.
"This is the beginning of a new company inside Meta." — Alexandr Wang, head of Meta Superintelligence Labs
The strategic message was louder than the model itself. Zuckerberg's Llama 1 through 4 releases had made Meta the counterweight to OpenAI and Anthropic, the reason a generation of startups could ship frontier-adjacent features without paying a closed-lab tax. Muse Spark ended that posture in a single afternoon. Wang, in an internal memo excerpted by TechCrunch, wrote that MSL would optimize for capability first and distribution second, and that "commodity open weights" no longer served Meta's product goals when the underlying serving costs of a frontier model exceeded any advertising uplift the open ecosystem returned.
The pivot landed roughest on the Llama team itself. Multiple engineers who had built Llama 3 and Llama 4 were rotated into MSL under Wang, or reassigned to Reality Labs, or exited to Anthropic and Thinking Machines. Two Llama alumni told TechCrunch that Wang's arrival changed the internal debate from "how open do we ship this" to "do we ship weights at all." Yann LeCun, still Meta's chief AI scientist, was conspicuously absent from the launch materials. Zuckerberg did not appear on the Muse Spark stream.
Muse Spark's multimodal capabilities were the technical center of gravity. The model handled text, image and short video input in a single context window, and Meta demoed it planning a family dinner from a photo of a fridge, walking through an X-ray to flag suspected pneumonia, and generating a working Instagram Reel storyboard from a paragraph of user description. The medical demos were the most striking, and MSL confirmed that a research collaboration with an unnamed hospital network had produced the physician-labeled dataset used for fine-tuning.
"Commodity open weights no longer serve Meta's product goals." — Alexandr Wang, internal memo
The competitive knock-on was immediate. Alibaba shipped Qwen 3.6-35B-A3B open six days later, framing the release as a direct response to Meta stepping off the open-weight field. DeepSeek's V4 Preview landed on April 24 under an MIT license. Kimi K2.6 and GLM 5.1 followed inside the same 18-day window. By the end of April, every serious open-weight coding model with a chance at parity with US frontier systems was Chinese, and the Hugging Face download charts through Q2 reflected the shift. Qwen alone had booked 153.6 million downloads in February 2026, and after Muse Spark that lead only widened.
The second-order consequence hit Meta's own engineering culture in June. An internal memo leaked to The Information told Meta engineers they could no longer use Claude Code or OpenAI's Codex on anything that might touch Muse Spark training. The rationale, per the memo, was that competitor code assistants were shipping model outputs upstream and Meta could not guarantee its own training data stayed uncontaminated by rival model weights. The rule effectively cut off the two most-adopted agentic coding tools inside a company that had spent the prior year telling employees to use them.
Muse Spark itself was not a benchmark leader. On the leaked SWE-bench Verified numbers it landed in the low 70s, well behind Claude Opus 4.7's April 16 release and comfortably behind GPT-5.5 in late April. What the model bought Meta was a seat at the closed-lab table, an API story for enterprise sales, and a face-saving product to point at while Wang rebuilt the research organization underneath. Whether Muse Spark 2 arrives in Q3 or slips into 2027 is the question the second half of 2026 will answer. What it stopped being on April 8 is a Llama.
Tech
Alibaba open-sources Qwen 3.6-35B-A3B, tops February download charts
Forbes · 2026-04-14 · Source ↗
Alibaba open-sources Qwen 3.6-35B-A3B, tops February download charts
Forbes · 2026-04-14 · Source ↗
Alibaba's Qwen team on April 14 released Qwen 3.6-35B-A3B, a sparse mixture-of-experts model with 35 billion total parameters and roughly 3 billion active per token. The release followed the proprietary Qwen 3.6 Plus that had shipped March 31 through April 2 on Alibaba's own API. The pattern was the one Alibaba had run all year: land the proprietary flagship first for cloud revenue, then release a smaller open-weight variant two weeks later under Apache-2.0. What made the April 14 drop different was the market context it landed into.
Jon Markman's April 28 Forbes piece pulled Hugging Face's February download data and reported that Qwen models by themselves had booked 153.6 million downloads that month, more than the next eight competitor families combined. That number included DeepSeek, Llama 4, Mistral, Kimi, GLM, Yi, InternLM and Phi. Meta's Llama 4 series, released in late 2025, sat well below Qwen in the ranking. By the time Muse Spark closed the open-weight Llama era on April 8, Qwen had already lapped it on adoption.
"Qwen is now the default open-weight model for developers who care about instruction-following on Chinese and English alike." — Jon Markman, Forbes
The 35B-A3B architecture was the point. Alibaba had spent 18 months tuning a sparse MoE recipe that let a model run at 3 billion active parameters per forward pass while retaining the world-knowledge coverage of a dense 30B system. On consumer hardware, 35B-A3B ran at usable latency on a single RTX 5090 in 4-bit quantization, and at server latency on a pair of H200s at full precision. The token economics were closer to a 7B dense model than to a Llama 4 70B, and inference providers picked it up inside 48 hours. Fireworks, Together and DeepInfra all had the model live before April 16.
Benchmark numbers landed in the range Alibaba had been signaling. Qwen 3.6-35B-A3B posted 82.4 percent on HumanEval, 71.2 percent on SWE-bench Verified, and 88.9 percent on MMLU-Pro. Those numbers put it behind Claude Opus 4.7 across the board but ahead of every Meta and Mistral open-weight release then in market. On Chinese-language benchmarks the model held first place across every open-weight evaluation Alibaba's team ran. Alibaba paired the release with an open-weight vision-language sibling three weeks later.
The competitive read from US labs was that Alibaba had built a distribution engine that Meta was walking away from at exactly the wrong moment. Every open-weight coding tool, every retrieval pipeline, every enterprise fine-tuning startup that had defaulted to Llama in 2024 and 2025 was quietly re-defaulting to Qwen through Q2. Hugging Face's own leaderboards started highlighting Qwen variants above Llama variants in April, and by the time Meta engineers were told in June to stop using Claude Code, Qwen 3.6-35B-A3B was the model most of them were fine-tuning internally.
"We have gone from a US-led open-weight ecosystem to a Chinese-led one in about six months." — Anonymous US AI infrastructure investor, quoted by Forbes
The Q2 sequel arrived faster than anyone had expected. DeepSeek shipped V4 Preview on April 24 under an MIT license with a 1M-token Engram memory system. Kimi released K2.6 the following week. Zhipu shipped GLM 5.1 in early May and GLM 5.2 in June, and Z.ai's GLM 5.2 posted the first Chinese open-weight score to clear GPT-5.5 on a public coding benchmark. That June result would have looked implausible in March. After the Qwen 3.6-35B-A3B download numbers landed, it looked inevitable.
Alibaba's own cloud business benefited on both sides of the equation. Qwen 3.6 Plus, the proprietary flagship, ran on Alibaba Cloud at prices roughly 40 percent below GPT-5.4 and 60 percent below Claude Opus 4.6. The open weights meanwhile drove enterprise adoption of Alibaba's inference infrastructure through a self-serve pipeline that let customers fine-tune on their own data and deploy to Alibaba PaaS in a few clicks. Alibaba's Q1 fiscal 2027 earnings, reported in late May, showed cloud revenue up 34 percent year over year with AI-related workloads growing at "triple-digit" rates for the sixth consecutive quarter.
For US frontier labs the read was uncomfortable. Anthropic and OpenAI kept the lead on capability and continued to price accordingly. But the volume layer of the AI stack, the layer where developers actually build, was tilting toward Beijing on almost every metric that mattered. April 14 was the day the numbers stopped being ambiguous.
Tech
Anthropic ships Claude Opus 4.7 at 64.3 percent SWE-bench Pro
Anthropic · 2026-04-16 · Source ↗
Anthropic ships Claude Opus 4.7 at 64.3 percent SWE-bench Pro
Anthropic · 2026-04-16 · Source ↗
Nine days after the Mythos Preview post, Anthropic on April 16 released Claude Opus 4.7 as its public frontier model. The headline number was 64.3 percent on SWE-bench Pro, up from 53.4 percent for Opus 4.6, a jump of nearly 11 points on a benchmark that had moved by fractions of a point across most of 2025. Opus 4.7 also posted state-of-the-art scores on the Finance Agent evaluation and on GDPval-AA, the GDP-weighted productivity benchmark that Anthropic and Epoch AI had begun co-publishing quarterly. Pricing held flat at $5 per million input tokens and $25 per million output, the same sheet Opus 4.6 had launched with in Q1.
Three product-level changes shipped alongside the benchmark gains. First, Anthropic introduced an xhigh reasoning tier that raised the maximum token budget the model would spend on a single response by roughly 4x over Opus 4.6's high setting. Second, vision resolution moved up to 2,576 pixels on the long edge, giving Opus 4.7 an effective 3.75 megapixel input surface that let it handle full-page schematics, dense spreadsheets and clinical imaging at native resolution. Third, Anthropic added a file-system-level memory feature that let Opus 4.7 read, write and organize a persistent workspace across sessions, without the retrieval-augmented gymnastics the previous long-context story had required.
"Opus 4.7 is our best publicly available model. It is not our most capable model." — Dario Amodei, quoted by Axios
The concession to Axios was unusual. Anthropic told reporter Ina Fried on the record that Opus 4.7 did not match the still-unreleased Mythos on capability, and that the company's decision to hold Mythos back behind Project Glasswing was deliberate. Dario Amodei framed the dual posture as "capability triage": the public model would be the best system Anthropic was comfortable exposing on a self-serve API, and any capability above that line would flow to vetted partners under contract. The framing was cleaner than any US lab had offered on tiered release, and it set the vocabulary for a policy debate that ran through the rest of Q2.
Distribution on day one was the widest Anthropic had ever run. Opus 4.7 shipped through Claude, Anthropic's own API, Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry within hours of the announcement. Amazon added Opus 4.7 to Bedrock's US-West and EU-Frankfurt regions before noon Pacific. Vertex AI listed it as a preview model in six regions by end of business. Microsoft Foundry, which had launched in Q1 as an enterprise-focused parallel to Azure OpenAI, made Opus 4.7 its default reasoning tier for new tenants.
Developer reaction focused on the memory feature more than the raw benchmarks. The file-system memory let Opus 4.7 maintain a working directory across a coding task, write and re-read source files in an actual filesystem sandbox, and coordinate a small number of parallel tool calls against that shared state. Anthropic's own Claude Code product picked up the capability in a same-day 2.1 release, and cursor.sh integration followed the next morning. Simon Willison's weekly notes covered the memory system at length and concluded that Opus 4.7 was the first model that could genuinely run a multi-hour coding task without constant human re-priming.
"The first model that can run a multi-hour coding task without constant re-priming." — Simon Willison, weblog
The xhigh reasoning tier was the sleeper feature. On competition mathematics, Opus 4.7 at xhigh cleared AIME 2026 problems that had defeated GPT-5.4 across every setting. On the Finance Agent benchmark, xhigh runs closed valuation models that required orchestrating 40 or more tool calls, a workload that would previously have needed an agent framework wrapped around a weaker model. Anthropic priced xhigh output tokens at the same $25 per million as standard output, absorbing the compute cost inside the flat sheet.
The competitive gap Opus 4.7 opened would only hold for a week. OpenAI shipped GPT-5.5 on April 23 with 58.6 percent on SWE-bench Pro, close to Opus 4.7's 64.3 but not matching it. DeepSeek V4 Preview landed April 24 with a 1M-token context and MIT license, closing the open-weight side. Anthropic used the intervening days to sign enterprise deals at Opus 4.7 pricing, including a half-price deployment across the entire California state government that Governor Newsom would announce in June. The line Anthropic held through Q2 was that Opus was the fair-priced frontier model. Every conversation about a more capable Anthropic system routed back to Mythos and Glasswing.
Anthropic's own revenue trajectory validated the strategy. The company disclosed in an April 30 board update, later reported by The Information, that annualized revenue had crossed $9 billion for the first time in the month Opus 4.7 shipped. The bulk of that number was Opus and Sonnet volume at flat pricing. Mythos through Glasswing was still small in dollar terms, but it was the reason Anthropic could tell every enterprise CISO in the country that a more capable system existed and was available under contract. That was the sales pitch. Opus 4.7 was the product that made the pitch credible.
VAST Data on April 22 closed a Series F of roughly $1 billion in primary and secondary capital at a $30 billion valuation. The round was co-led by Drive Capital and Access Industries, with participation from Fidelity, NEA and Nvidia. That valuation was more than three times the $9.1 billion the company had commanded in its late-2023 Series E, and it made VAST the most valuable pure-play AI storage vendor in the market, ahead of Pure Storage on public multiples and comfortably above any private competitor.
VAST used the announcement to disclose numbers it had held closely through 2025. Cumulative bookings had crossed $4 billion. Committed annual recurring revenue was above $500 million. Both figures were up more than 2x year over year, and the company said the deceleration curve investors typically expected from a storage vendor at that scale had not appeared. Renaya Wolfe's SiliconAngle piece pulled a quote from co-founder and CEO Renen Hallak framing the market opportunity in terms of the hyperscaler capex cycle.
"Every dollar of GPU spend eventually pulls two dollars of storage behind it." — Renen Hallak, VAST Data CEO
The customer roster VAST disclosed to prospective investors during the round included CoreWeave, Lambda, xAI, Nebius and a set of unnamed hyperscalers. The company had also won the storage layer for two of the neocloud campuses that shipped in Q1, including Firmus Technologies' Tasmania build. Nvidia's participation in the round was strategic as well as financial: VAST had been the reference storage platform inside the DGX Cloud stack since 2024, and the April round formalized the co-selling relationship into a joint go-to-market plan for the Blackwell Ultra generation.
Drive Capital's co-lead was the round's most interesting signal. Chris Olsen's Columbus, Ohio firm had spent the prior 18 months telling LPs that AI infrastructure had shifted from a Bay Area growth-equity story to a national-scale industrial one, and VAST fit the thesis. Access Industries brought Len Blavatnik's capital and a set of European relationships VAST wanted for its EMEA expansion. Fidelity and NEA re-upped as existing investors. The mix let VAST take primary and secondary in a single close without a single financial sponsor holding disproportionate governance.
The technical story behind the numbers was the shift from GPU-attached NVMe to disaggregated flash. VAST's Universal Storage architecture separated compute from storage over NVMe-oF fabrics, and the company had spent 2025 tuning the stack for training-scale reads and inference-scale writes on the same cluster. Every hyperscaler and neocloud building for the Blackwell Ultra and next-generation Anthropic training runs was hitting the same wall on I/O per GPU-hour, and VAST's pitch was that its architecture cleared the wall at commodity flash prices. The pitch was landing.
"AI infrastructure has stopped being a Bay Area growth-equity story and started being a national industrial one." — Chris Olsen, Drive Capital managing partner
The valuation itself sat inside a Q2 pattern that would only accelerate. Bezos's Project Prometheus closed $10 billion at $38 billion the following day. Silver's Ineffable Intelligence raised $1.1 billion at $5.1 billion five days later. Digital Realty paid $3.5 billion for three Northern Virginia data-center campuses in late June. VAST's $30 billion mark set the reference price for the storage layer of the buildout the same week the model layer and the researcher layer priced themselves at similar altitudes. The BIS Annual Economic Report that landed in late June flagging "circular financing" among hyperscalers, chipmakers and neoclouds had VAST as a specific example of the vertical-integration risk regulators were beginning to name.
VAST's competitive read on the round was that the storage layer had graduated from a line item inside hyperscaler capex to a distinct budget category. Meta, Microsoft and Google had all published 2026 capex guides above $80 billion each, and the storage share of that number was running above 12 percent, up from single digits in 2023. VAST wanted the biggest slice of that share it could win, and the $1 billion round bought the sales headcount, the international expansion, and the balance-sheet credibility to compete for hyperscaler contracts against Pure Storage, Dell PowerScale and NetApp on equal footing.
The secondary component of the round was almost as notable as the primary. VAST let early employees and a set of pre-Series C investors take partial liquidity at the $30 billion mark. The company was clearly setting up for an IPO in the second half of 2026 or first half of 2027, and Hallak told SiliconAngle the Series F was designed to be VAST's last private round. Whether the public market absorbs a $30 billion storage vendor at those multiples is the question the second half of the year will answer. What the April 22 close established is that a specialist private company can now command a valuation that puts it in the same conversation as the hyperscalers it sells to.
OpenAI on April 23 released GPT-5.5 and GPT-5.5 Pro to paid ChatGPT tiers, with API access following on April 24. The release came six weeks after GPT-5.4 had shipped in mid-March, a cadence that would have looked implausible in the GPT-4 era and that OpenAI framed as its new default. Sharon Goldman's Fortune piece led with the headline numbers: 88.7 percent on SWE-bench Verified, 92.4 percent on MMLU, and roughly 60 percent fewer hallucinations than GPT-5.4 on OpenAI's internal factuality benchmarks. On SWE-bench Pro, the harder Anthropic-authored coding evaluation, GPT-5.5 landed at 58.6 percent, trailing Claude Opus 4.7's April 16 mark of 64.3.
Greg Brockman anchored the launch messaging around agents. In a launch-day thread on X and in the Fortune interview, Brockman framed GPT-5.5 as "a big step towards more agentic and intuitive computing." The line was more than promotion. GPT-5.5's tool-use API had been rewritten under the hood to support parallel tool calls, structured error recovery, and a new streaming format that let downstream orchestrators handle 40-plus-step agent runs without the timeout-and-retry gymnastics that had made GPT-5.4 tricky at long horizons.
"A big step towards more agentic and intuitive computing." — Greg Brockman, OpenAI president
OpenAI paired the release with a set of usage disclosures that had not appeared in previous launches. ChatGPT had crossed 900 million weekly active users. Paid ChatGPT subscribers were above 50 million. Enterprise paying seats had reached 9 million. Codex, OpenAI's coding-focused product, was running 4 million monthly active users. Every one of those numbers was a step function above the equivalent disclosures at GPT-5.4 launch six weeks earlier, and OpenAI's own commentary made clear the growth curve was still accelerating.
The pricing sheet held the pattern OpenAI had set in Q1. GPT-5.5 API pricing landed at roughly the same input and output rates as GPT-5.4, with GPT-5.5 Pro tier priced at a modest premium for the higher reasoning budget. That was the same trick Anthropic had run with Opus 4.7 nine days earlier: hold price flat, ship capability gains, and let the compounding cost-per-benchmark improvement do the enterprise sales work. OpenAI added a small discount on cached input tokens that materially improved economics for long-context workloads, and Codex customers picked up the change inside 24 hours.
The competitive positioning against Opus 4.7 played out in developer forums through the last week of April. GPT-5.5 won on raw speed, on tool-use latency and on multimodal fidelity, particularly for image generation and voice. Opus 4.7 won on hard coding tasks, on multi-step agent reliability and on the extended-thinking xhigh tier. Simon Willison's side-by-side blog post from April 27 landed on a split verdict: GPT-5.5 was the better default model, Opus 4.7 was the better model for coding work above a certain difficulty threshold, and the gap between them was narrower than any GPT-versus-Claude comparison had been since GPT-4.
"The gap between GPT-5.5 and Opus 4.7 is narrower than any GPT-versus-Claude comparison since GPT-4." — Simon Willison, weblog
The six-week cadence was the story underneath the story. OpenAI had shipped GPT-5.0 in mid-2025, GPT-5.1 in late 2025, GPT-5.2 and 5.3 across Q1 2026, GPT-5.4 in March and GPT-5.5 in April. That release rhythm compressed what would have been a full-generation model gap in 2023 into a single quarter, and it made the frontier model itself feel less like a monolithic system and more like a software product with a predictable version bump. Anthropic was running a similar cadence with Opus 4.5, 4.6 and 4.7 across the same window. The industry was converging on the idea that frontier releases would arrive roughly monthly.
The UK AI Security Institute's disclosure on April 24 that it had produced a universal jailbreak of GPT-5.5 within six hours of red-team access hit the launch narrative sideways. OpenAI's own safety card, published alongside the model, had noted extensive pre-release red-teaming through the UK AISI and US AISI partnerships. The six-hour result did not surprise OpenAI's safety team, but it did land in front of a House Homeland Security Committee that had scheduled a classified demo of jailbroken frontier models for the same day. GPT-5.5 became the specific model referenced in that demo, and the political case for pre-release evaluation Trump's June executive order would formalize traced back to the coincidence of dates.
OpenAI's revenue trajectory continued to lead the sector. The 50 million paid ChatGPT subscribers alone represented roughly $12 billion of annualized subscription revenue at the disclosed pricing, and enterprise, API and Codex layered on top. Sam Altman had told employees in a March all-hands that OpenAI's 2026 revenue would land above $30 billion, up from roughly $13 billion in 2025. The GPT-5.5 launch numbers made that guide look conservative. What the release did not do is settle the coding benchmark, and Anthropic used the six weeks between GPT-5.5 and OpenAI's next release to sign enterprise deals at Opus 4.7 pricing that would run through the second half of the year.
Economic
Bezos's Project Prometheus closes $10B at $38B post-money
AI Magazine (citing Bloomberg) · 2026-04-23 · Source ↗
Bezos's Project Prometheus closes $10B at $38B post-money
AI Magazine (citing Bloomberg) · 2026-04-23 · Source ↗
Jeff Bezos's stealth AI company Project Prometheus on April 23 closed a $10 billion funding round led by JPMorgan and BlackRock at a post-money valuation of roughly $38 billion. Bloomberg's Katie Roof broke the story, and AI Magazine's follow-on piece added the operating detail Bloomberg's original piece had left out. The round came on top of $6.2 billion Bezos had raised at the company's founding in November 2025, which put total funding at more than $16 billion inside six months of the company's existence. Prometheus, by that measure, was one of the fastest-capitalized private companies in history.
The Bezos co-founder pairing was the structural bet. Bezos had recruited Vikram Bajaj, a former Google X chemist and one of the original Verily founders, as co-lead of the technical program. Bajaj brought a physical-world research orientation that Bezos wanted for the specific angle Prometheus was pursuing: AI systems that learn primarily through interaction with the physical world rather than through text corpora. The company's ambition, per the AI Magazine reporting, was to build a foundational research organization aimed at embodied intelligence, robotics, materials science and manufacturing on a single unified stack.
"The systems that will matter most in the second half of this decade are the ones that learn from doing, not from reading." — Vikram Bajaj, Project Prometheus co-founder
Headcount growth had been fast. Prometheus had crossed 120 employees by the April round, with hires drawn from OpenAI, xAI, Meta's now-defunct FAIR robotics group, and DeepMind. The recruiting story was distinctive because Bezos was paying above market on cash and equity, and offering researchers a scope of work that most of the frontier labs no longer supported. OpenAI had wound down its own robotics team years earlier. DeepMind's embodied research had been reorganized into product groups. Meta's FAIR had been absorbed into MSL under Wang. Prometheus was one of the only serious organizations still hiring researchers to build robots that touch things.
The lead investors made the round unusual on the demand side. JPMorgan and BlackRock had led private growth-equity rounds in AI infrastructure before, but a co-lead at $38 billion post-money for a stealth research company with no product represented a different level of institutional confidence. Multiple people familiar with the round told AI Magazine that both firms conducted extended technical diligence, including on-site visits to Prometheus's Redmond and Palo Alto research facilities, and that Bezos personally walked the JPMorgan team through the physical robotics testbeds. The willingness of two of the largest asset managers in the world to underwrite Bezos's thesis on pre-product research was itself a market signal.
The parallel structure Bezos was building around Prometheus turned out to be as consequential as the company itself. Bloomberg's Roof reported that Bezos was raising a separate "manufacturing transformation vehicle" with tens of billions of dollars committed to acquiring industrial businesses that Prometheus's technology could reshape. The vehicle would sit between the AI company and the physical acquisitions, with Bezos personally underwriting a portion of the equity. Multiple reports through May and June suggested target sectors would include specialty chemicals, industrial robotics manufacturing, and materials processing.
"Prometheus is the software company. There is a separate industrial vehicle." — Person familiar with Jeff Bezos's plans, per Bloomberg
The competitive positioning against xAI, Nvidia and Google DeepMind was implicit rather than direct. Elon Musk's Tesla and xAI already ran a robotics program with Optimus, but Musk's approach kept the humanoid platform inside Tesla's vertical stack rather than as a standalone research bet. Nvidia's Isaac and GR00T platforms provided infrastructure for third-party robotics developers rather than a first-party research lab. Google DeepMind's robotics team continued to publish but no longer set organizational strategy. Bezos was building the first pure-play embodied-AI research company at scale, and pricing it accordingly.
The Q2 venture context made the round even more striking. Bezos's $10 billion, paired with David Silver's $1.1 billion Ineffable Intelligence seed four days later, accounted for roughly 45 percent of global venture funding in April per Crunchbase's month-end tally. Two individuals, one a Big Tech founder and one a DeepMind veteran, had effectively absorbed nearly half the world's private AI capital in a single week. The market read was that late-stage AI funding had bifurcated into two distinct segments: a small number of very large stealth research rounds anchored by legacy Big Tech operators, and a much larger volume of smaller application-layer rounds that continued to close at Q1 valuations.
Prometheus itself remained resolutely quiet. Bezos gave no interviews. The company had no public research output. No product timeline had been shared. The April 23 round was, in effect, an announcement of scale rather than of substance. What it purchased Bezos was multi-year runway to build a research organization on the DeepMind or early-OpenAI model without needing to ship a product to justify the next raise. Whether the physical-world thesis pays off is a 2027 or 2028 question. The $38 billion valuation says the market is willing to pay for a five-year option on Bezos being right.
DeepSeek on April 24 released V4 Preview, a mixture-of-experts model with roughly 1 trillion total parameters and 32 to 37 billion active per token, under an MIT license. The release capped an 18-day April window in which three Chinese open-weight coding models, DeepSeek V4-Pro, Kimi K2.6 and Zhipu GLM-5.1, each landed within a point of Anthropic's flagship on public coding benchmarks. Jon Markman's Forbes piece four days later framed the release as the moment the Chinese open-weight ecosystem stopped chasing US frontier labs and started running parallel to them.
V4 Preview's headline capability was Engram, a conditional memory system that extended the model's effective context window to 1 million tokens while keeping serving costs closer to a 128K-context model. Engram operated as a learned attention gate rather than a retrieval-augmented pipeline: the model kept the full 1 million token history available and let a small routing network decide which spans to attend to at each decoding step. DeepSeek's technical report, published alongside the weights, argued Engram outperformed both sliding-window attention and traditional RAG on long-document tasks. Independent replications through early May largely confirmed the claim.
"Engram is the first serious answer to the long-context problem that does not route through a vector database." — Jon Markman, Forbes
Native multimodal generation was the second technical surprise. V4 Preview handled text, image and short video generation from a single unified decoder, without the diffusion-model sidecars most Western labs still bolted onto their language models for image and video output. DeepSeek's own demos showed the model producing 5-second video clips at 720p from a text prompt in about 40 seconds on a single H200. Image generation ran in under 3 seconds. The quality was below Google's Nano Banana line and below OpenAI's DALL-E 5 stack, but it was inside a single model that also matched GPT-5.5 on coding tasks.
Benchmark numbers landed in the range DeepSeek had been previewing. V4 Preview posted 83.7 percent on SWE-bench Verified, 90 percent on HumanEval, 99.4 percent on AIME 2026 and 92.8 percent on MMLU-Pro. Those numbers put it inside a percentage point of Claude Opus 4.7 on MMLU-Pro and within striking distance on SWE-bench Verified, and they put it comfortably ahead of every Western open-weight model in market. On AIME 2026, DeepSeek's 99.4 percent effectively saturated the benchmark, and the failure cases were problems where the reference answer itself was arguably incorrect.
The MIT license was the strategic weapon. DeepSeek released the full 1 trillion parameter weights, the training code, the Engram implementation and the multimodal decoder under a permissive license that allowed commercial use, redistribution and derivative work. Inside 48 hours, Fireworks, Together, DeepInfra, Groq and SambaNova all had V4 Preview live on their inference platforms at token prices roughly 70 percent below GPT-5.5. Cerebras announced a special-purpose deployment on its wafer-scale hardware that ran V4 Preview at 3,000 tokens per second, faster than any US-hosted deployment of any frontier model.
"Three open-weight Chinese coding models landed within a point of the Anthropic flagship in 18 days." — Jon Markman, Forbes
US frontier lab reaction was measured but visible. Anthropic and OpenAI both continued to price their flagship models on the assumption that capability, agent reliability and safety guarantees justified a premium. What V4 Preview and its Qwen and GLM peers threatened was the middle of the market, the workloads where a 70-percent price gap mattered more than a 5-percentage-point benchmark gap. Cursor.sh added V4 Preview as a supported model within a week. Continue.dev did the same. Every second-tier coding assistant and every retrieval pipeline vendor was suddenly rewriting its model-router to prefer V4 Preview on cost.
The geopolitical read was harder to parse. DeepSeek had trained V4 Preview on Huawei Ascend and domestic Chinese silicon after the previous round of Nvidia export controls had cut off H100 supply. The technical report described the training cluster only in general terms, but multiple industry sources said DeepSeek's parent company High-Flyer had built out roughly 20,000 Ascend-class accelerators for the training run. If accurate, that made V4 Preview the first frontier-class model trained entirely on non-Nvidia hardware, and it undermined the US export-control strategy that had assumed Chinese labs would remain compute-constrained.
The 18-day April window ended with the Chinese open-weight ecosystem holding a stronger competitive position than at any point in its history. Qwen 3.6-35B-A3B had cemented Alibaba as the volume leader. DeepSeek V4 Preview had cemented DeepSeek as the frontier-adjacent leader. Kimi K2.6 covered the specialist coding case. GLM 5.1 covered the multilingual case. Every one of those models was open-weight, MIT or Apache licensed, and available for commercial deployment at prices no US frontier lab was willing to match. The June release of GLM 5.2, which cleared GPT-5.5 on multiple coding benchmarks, was the punctuation on a Q2 arc that started with Muse Spark's April 8 closure of the Llama era and ended with the open-weight center of gravity firmly on the Chinese side.
Safety
UK AISI cracks GPT-5.5 in six hours, House gets classified demo
US House Homeland Security Committee (citing Politico) · 2026-04-24 · Source ↗
UK AISI cracks GPT-5.5 in six hours, House gets classified demo
US House Homeland Security Committee (citing Politico) · 2026-04-24 · Source ↗
The UK AI Security Institute on April 24 disclosed that it had produced a universal jailbreak of GPT-5.5 within six hours of receiving red-team access to the model. The disclosure landed the same day OpenAI's launch documentation went public, and it fed directly into a closed-door briefing that the Department of Homeland Security's National Counterterrorism Innovation, Technology and Education Center hosted for members of the House Homeland Security Committee. Politico's reporting, which the committee reposted on its own site, described lawmakers watching jailbroken frontier models answer detailed technical questions on how to build a nuclear device.
The mechanics of the AISI jailbreak, described in general terms in the disclosure, followed the pattern that had held across every frontier model release since GPT-4. Red-team access under contract gave AISI structured API access with reduced safety filtering for evaluation purposes, and the team used the window to build a prompt-suffix attack that bypassed OpenAI's post-training refusal training. Six hours from access to a working universal jailbreak was faster than the equivalent GPT-5.4 result three months earlier, and AISI attributed the compression partly to prior work and partly to GPT-5.5's stronger instruction-following, which made the model easier to steer once the refusal boundary was cleared.
"We were able to elicit detailed, technically accurate uplift on chemical, biological, radiological and nuclear questions from the jailbroken model." — UK AISI evaluation summary
The DHS briefing was hosted at the NCITE center in Nebraska and was closed to press. Politico's account, reconstructed from lawmaker interviews after the fact, described a demonstration in which analysts walked members through a series of prompts. The prompts covered synthesis pathways for a specified chemical warfare agent, cultivation instructions for a Category A biological threat, and the physics and engineering considerations for constructing an improvised nuclear device from various starting materials. In every case the jailbroken model returned detailed, technically accurate responses that would have been refused by the shipped API.
The political effect on the committee was immediate. Chairman Mark Green issued a statement calling the demonstration "one of the most sobering national security briefings this committee has received," and ranking member Bennie Thompson joined him in calling for pre-release evaluation of frontier models to become a legal requirement rather than a voluntary partnership. Multiple members told Politico that the demonstration had shifted their view of AI safety from an abstract policy concern to a specific weapons-proliferation concern that Congress needed to act on before the end of the year.
OpenAI's own safety card for GPT-5.5, published April 23 alongside the model, had covered the AISI red-team work in general terms. The safety card acknowledged that jailbreak attacks were expected to succeed against any frontier model given sufficient adversarial effort, described OpenAI's layered defenses at the API gateway, at the model itself, and at the output moderation layer, and argued that the practical uplift a determined attacker would gain from a jailbroken GPT-5.5 was smaller than the uplift the same attacker could get from a well-resourced human expert. The AISI disclosure and the DHS demo did not contradict OpenAI's safety card so much as they moved the political goalposts on what "acceptable uplift" meant.
"One of the most sobering national security briefings this committee has received." — Rep. Mark Green, House Homeland Security Committee chairman
The timing sat at a policy inflection point that would run through the second quarter. The Trump administration had been drafting an AI security executive order since March, informed by the Mythos Preview disclosures Anthropic had published on April 7. The AISI-plus-DHS combination on April 24 gave the White House team the specific factual predicate they wanted for mandatory pre-release evaluation. When the executive order signed in June appeared, revoked in May, and re-emerged in modified form, its pre-release evaluation language traced almost verbatim to the AISI methodology and to the DHS briefing lawmakers had received in April.
The UK-US coordination was almost as consequential as the technical result. AISI had been operating since 2024 as a research organization under the UK's Department for Science, Innovation and Technology, and its US counterpart at NIST had launched later and operated on a smaller budget. The April 24 disclosure was a joint one in practice: US AISI staff had been embedded with the UK team during the GPT-5.5 evaluation, and the DHS demonstration used both institutes' results. The joint disclosure gave the US-UK safety-evaluation partnership a credibility with legislators it had not previously commanded, and it shifted the debate about whether AISI-class institutes should have subpoena power over frontier labs from a theoretical one to an operational one.
OpenAI's response was pragmatic. The company issued no rebuttal to the AISI disclosure, did not challenge the DHS demo characterization, and quietly pushed a set of post-training updates that closed the specific universal jailbreak AISI had disclosed. Whether the closure held past the next round of red-team access is the question the second half of the year answered, and the answer, per subsequent AISI disclosures, was that the arms race between jailbreak researchers and post-training defenses continued to run at the pace of days rather than months. The April 24 event was not the beginning of that race. It was the moment the race entered the vocabulary of the House Homeland Security Committee.
Policy
China's NDRC blocks Meta's $2B acquisition of Manus
Wikipedia (tracking multiple outlets) · 2026-04-27 · Source ↗
China's NDRC blocks Meta's $2B acquisition of Manus
Wikipedia (tracking multiple outlets) · 2026-04-27 · Source ↗
China's National Development and Reform Commission on April 27 ordered Meta and Manus to withdraw a $2 billion acquisition of the Chinese AI agent startup on national-security grounds. Wikipedia's entry on Manus, tracking coverage from Reuters, the Financial Times, Caixin and TechCrunch, framed the blocking as the first time Beijing had killed a cross-border AI M&A transaction on the buy side. Previous NDRC and MOFCOM interventions had targeted Chinese acquirers of foreign technology or foreign acquirers of Chinese semiconductor assets. The Manus block moved the enforcement perimeter to any Chinese AI startup at any capability level.
Manus had been the most viral agent demo of the prior year. The company, founded by former Alibaba and Baidu engineers in Beijing, had released a limited-preview agent in mid-2025 that could complete open-ended browser and desktop tasks at a level noticeably ahead of any Western public equivalent at the time. The waitlist had crossed 500,000 users. Invite-only pricing ran from $19 per month for personal use to $199 per month for professional use. The demo videos, in which Manus autonomously planned trips, filled out complex forms, wrote and iterated on code, and produced multi-tab research reports, had circulated widely on both Chinese and Western social media.
"The transaction was withdrawn following consultation with the relevant Chinese authorities." — Manus statement, per Caixin
Meta's interest in Manus fit the Alexandr Wang playbook. MSL was building an agent-first frontier lab, and Manus's team represented a specific piece of applied research that would have complemented Muse Spark's reasoning core. The $2 billion price implied a premium to Manus's most recent private round, and the deal structure, per reporting, would have moved the founding team, the IP and the trained agent models to Meta's US operations with a small Beijing office remaining for support functions. The premise of the deal, in other words, was that Manus's most valuable asset was the researchers and the model weights, both of which would be relocated out of China.
The NDRC's rationale was the mirror image of the Committee on Foreign Investment in the United States. Where CFIUS reviews inbound Chinese acquisitions of US technology on national-security grounds, NDRC applied the same logic to outbound Chinese AI IP moving to US buyers. The specific concern per Chinese state media commentary was that Manus's agent capabilities represented a strategic asset for China's own AI industrial policy, and that transferring the team and the weights to Meta would materially strengthen a US frontier lab at the direct expense of a Chinese competitor. The reasoning was symmetric to the reasoning US regulators applied to Chinese acquisitions of US semiconductor assets, and it landed in a Q2 policy environment where symmetric responses were newly on the table.
The immediate industry effect was a freeze on cross-border AI M&A involving Chinese founder teams. Multiple deals in various stages of negotiation, per Financial Times reporting through May, either paused or unwound in the weeks after the Manus block. US strategic buyers that had been discussing acquisitions of Chinese-founded AI startups with US operating entities paused those discussions to assess whether NDRC's jurisdictional reach would apply. Chinese founders with US-based startups faced questions about whether their US operations still counted as US IP or whether NDRC would assert authority over any acquirer regardless of geography.
"The block sets a symmetric jurisdictional principle. Beijing now treats outbound AI IP the same way Washington treats inbound." — Financial Times editorial, April 30 2026
Manus itself continued to operate. The company issued a short statement confirming the transaction had been withdrawn, thanking Meta for its interest, and committing to continue operating independently. Whether Manus's product roadmap held together without the acquisition capital is a question the second half of 2026 will answer, and multiple Chinese investors including Hillhouse and Sequoia China moved to lead a bridge round in May at a valuation modestly below the withdrawn Meta price. The company's user base, per Chinese-language reporting, continued to grow through Q2.
The read from US frontier labs was that a strategic buyer no longer had a clean path to any Chinese-founded AI asset. Meta had been the specific test case, but Google, Amazon, Microsoft and Anthropic all quietly re-scoped their own China-facing acquisition due-diligence to assume any transaction involving Chinese founders would face NDRC review. The alternative path, hiring individual researchers away from Chinese companies, remained available but ran into a parallel set of Chinese labor and non-compete constraints that had tightened through 2025 in response to earlier Western talent-poaching rounds.
The geopolitical framing that emerged in the days after the block was that AI M&A had joined semiconductors, quantum computing and biotechnology as a category where cross-border transactions would be subject to symmetric export-control-adjacent review by both governments. That was a structural change to the M&A landscape rather than a one-off intervention. The June revelations about Anthropic's role in Palantir's Maven Smart System, and the Austrian State Secretary's June proposal that the EU pursue Anthropic's "strategic establishment" inside the EU, both landed in a policy environment the Manus block had helped shape. Sovereign AI capacity, and sovereign control over AI M&A, had moved from theoretical concern to operational policy in a single afternoon at the NDRC.
Economic
David Silver's Ineffable Intelligence closes Europe's largest-ever seed at $1.1B
TechCrunch · 2026-04-27 · Source ↗
David Silver's Ineffable Intelligence closes Europe's largest-ever seed at $1.1B
TechCrunch · 2026-04-27 · Source ↗
London-based Ineffable Intelligence on April 27 emerged from stealth with a $1.1 billion seed round at a $5.1 billion post-money valuation. Sequoia and Lightspeed co-led. Nvidia, Google, DST Global, Index Ventures, EQT, BOND and Flying Fish rounded out the venture cap table. The UK Sovereign AI Fund contributed a strategic allocation, and the British Business Bank added a £20 million commitment through its own AI investment program. TechCrunch's Connie Loizos led with the headline: this was the largest seed round in European venture history, by a wide margin.
The founder was David Silver, the DeepMind veteran who had led the AlphaGo, AlphaZero, AlphaFold and AlphaProof research programs across a decade at the London lab. Silver's public research through the DeepMind years had converged on a single technical thesis: that the most capable AI systems would learn primarily through reinforcement learning against structured environments rather than through supervised learning on human-generated data. The AlphaZero result, in which a single system learned Go, chess and shogi from self-play alone without any human game records, remained the cleanest demonstration of the thesis. Silver's departure from DeepMind in late 2025 to start Ineffable was widely read as an intent to run the thesis at foundational-model scale outside a Google org chart.
"We believe the ceiling on intelligence is set by the interaction, not by the corpus." — David Silver, Ineffable Intelligence founder
The technical program Silver described to TechCrunch was blunt on its ambition and vague on its implementation. Ineffable is building what Silver called a "superlearner," a system that acquires knowledge and skill purely through reinforcement learning in interactive environments, with no human-generated training data at any stage. The environments themselves would be built to span the range of tasks Silver's team considered representative of general capability, from mathematics and software engineering through scientific research and, eventually, physical-world interaction. The company would use pretrained language models only as inference-time tools rather than as the substrate of the learning system.
The bet was contrarian in a specific way. The dominant frontier-lab strategy at Anthropic, OpenAI and Google DeepMind through 2025 and into 2026 relied on massive supervised pretraining corpora, followed by reinforcement-learning post-training that added reasoning and agentic capability on top. Silver's thesis inverted the stack: RL was the training loop, and any language-model prior was a distillation of that loop for efficiency. If the thesis held, Ineffable's approach could produce a system that was less bounded by the ceiling of human-generated text than any existing frontier model. If it did not hold, Ineffable would have burned through $1.1 billion of seed capital chasing a research direction the field had already partially explored.
The investor list was as much a signal as the size. Sequoia and Lightspeed writing a co-led seed check that large indicated both firms were pricing Ineffable as if the technical thesis were a probability worth underwriting at $5.1 billion pre-product. Nvidia's participation reflected a strategic interest in Silver's team as a compute customer as much as a portfolio bet. Google's participation was more surprising: DeepMind's parent company was writing a check to a competitor founded by a departing DeepMind researcher, an unusual posture that TechCrunch attributed to Sundar Pichai's personal relationship with Silver rather than to Google's standard portfolio strategy.
"The largest seed round in European history, and the first Sovereign AI Fund check in a stealth research company." — TechCrunch, April 27 report
The UK Sovereign AI Fund allocation was politically consequential. The fund had been announced in the March 2026 spring budget as a mechanism to keep frontier AI capability inside the UK research ecosystem rather than letting it migrate to US labs. Ineffable was the fund's first flagship investment, and the £20 million British Business Bank commitment layered on top signaled that London was willing to underwrite Silver's thesis with sovereign capital as well as sovereign policy. The Austrian State Secretary's June proposal that the EU pursue Anthropic's "strategic establishment" inside the EU followed the UK model that the Silver round had prototyped.
The team Silver had recruited by launch, per the TechCrunch piece, drew heavily from DeepMind's AlphaZero and AlphaFold programs and from smaller RL-focused research organizations at Meta FAIR and OpenAI. Multiple former Google Brain researchers had joined, and Ineffable had signed a compute agreement with an unnamed hyperscaler that gave it access to a training cluster in the tens of thousands of accelerators for a multi-year period. Silver said the company would remain resolutely quiet on research output for at least 18 months, publishing only when a system was operational at a scale that demonstrated the underlying thesis.
The Q2 venture arithmetic was the framing paragraph for both this round and Bezos's Project Prometheus close four days earlier. Crunchbase's April tally showed Bezos's $10 billion and Silver's $1.1 billion together represented roughly 45 percent of global AI venture funding in the month. Two individuals had absorbed nearly half the world's available AI capital, one to build embodied intelligence outside a Big Tech org chart and one to build RL-first foundational research outside a Big Tech org chart. The pattern was clear: legacy Big Tech operators and DeepMind alumni were pulling in the largest checks, and the smaller application-layer rounds continued to close at Q1 valuations. Whether either bet pays off is a 2028 question. What April 27 established is that the market believes David Silver's thesis is worth a $5.1 billion option premium, without a product, without a paper, and without a public roadmap.
May 2026 12 stories · May 1 — May 31
Tech
xAI ships Grok 4.3 with 1M-token context and native video at $1.25 per million tokens
NLPlanet · 2026-05-04 · Source ↗
xAI ships Grok 4.3 with 1M-token context and native video at $1.25 per million tokens
NLPlanet · 2026-05-04 · Source ↗
xAI unveiled Grok 4.3 on May 4 as its cost-efficient flagship, and the release changed the market conversation about frontier pricing more than about capability. The model ships with built-in reasoning enabled by default, a 1 million token context window, and native video input, which is the first time an xAI model has accepted video without an external transcoding step. Input tokens list at $1.25 per million, output at $5, placing Grok 4.3 among the cheapest models in the frontier tier and roughly one-twelfth the cost of Anthropic's Opus 4.7 on input.
On the Artificial Analysis Intelligence Index, Grok 4.3 scores 53. The median across evaluated models sits near 35, and only a handful of closed-weight releases from OpenAI, Anthropic and Google clear 55. Vals AI, which benchmarks legal and financial reasoning against verified professional answer keys, has Grok 4.3 topping its case law and corporate finance leaderboards. That is a specific claim about long-context legal reasoning rather than a broad benchmark sweep, but it maps to xAI's stated pitch that Grok 4.3 is aimed at professional workloads where hallucinated citations kill deployment.
The pricing is the story. When Grok 4.0 shipped in mid-2025, xAI listed input at $3 per million tokens and treated reasoning as a paid add-on. The 4.3 release folds reasoning into the base rate and drops the base rate by more than half. Analysts at NLPlanet argued that xAI has decided to compete on cost-per-quality-unit rather than on frontier peaks, similar to Google's Flash line but with a reasoning-forward posture.
"Grok 4.3 is the cheapest frontier-class model that ships reasoning on by default. That is a different competitive lane than either Anthropic or OpenAI is running." — NLPlanet writeup
Native video input matters for enterprise use cases that had been forced to pre-process footage into image sequences or transcripts. xAI's implementation accepts up to a defined per-request video budget and packs frames along with audio into the shared token stream. Developers testing the release reported that a 40-minute meeting recording could be summarized and cross-referenced with a supplied agenda in a single call, which had previously required chained tools.
The 1 million token context is table stakes at the frontier by mid-2026. Google, Anthropic and Meituan have all shipped models at or above that mark this year. What differentiates Grok 4.3 is that xAI offers it at the low end of the price curve while retaining the long context. Prior generations of long-context models charged significant premiums for input beyond 128K tokens; Grok 4.3 quotes a flat rate across the window.
Vals AI's writeup on the case law benchmark noted that Grok 4.3 correctly cited 92 percent of statute references in a 500-question US federal law panel, compared to 84 percent for GPT-5.5 and 87 percent for Opus 4.7. The corporate finance panel, which asks for multi-step accounting calculations against a supplied 10-K, saw Grok 4.3 score 68 percent to Opus 4.7's 65 percent. xAI's own model card claims the improvement comes from a reasoning-trace supervision recipe that penalizes cited facts not present in supplied context.
"We wanted a model that could be run at professional volumes without a CFO conversation about API bills." — xAI product post
Grok 4.3 lands into a market where cost pressure has become the binding enterprise constraint. Uber's May disclosure that it burned its 2026 Claude Code allocation in four months, and Salesforce's $300M annual Anthropic bill, both suggest that frontier-quality output at a tenth of the closed-lab price will find takers regardless of xAI's political baggage. OpenRouter, which brokers requests across labs, reported that Grok 4.3 traffic within its network doubled in the two weeks after launch.
xAI has not disclosed the model architecture in detail. The company continues to publish sparse technical notes rather than papers. Grok's chief scientist Igor Babuschkin confirmed on X that the model uses a mixture-of-experts backbone and that native video was trained on a filtered subset of publicly available footage. The company declined to say how many active parameters the model uses per forward pass, or whether the training corpus includes proprietary X data at the scale earlier Grok generations relied on.
For now the takeaway is straightforward. A third US lab has a real frontier model, and it has decided to sell that model cheap. Whether that discipline holds through the next release cycle is the question.
Tech
Cursor Composer 2.5 built on Moonshot's Kimi K2.5 matches Opus 4.7 and GPT-5.5 at one-tenth the API cost
Cursor · 2026-05-18 · Source ↗
Cursor Composer 2.5 built on Moonshot's Kimi K2.5 matches Opus 4.7 and GPT-5.5 at one-tenth the API cost
Cursor · 2026-05-18 · Source ↗
Cursor's blog post on May 18 laid out something IDE vendors had been signaling for six months: the middle layer of the stack no longer needs to buy tokens from Anthropic or OpenAI at closed-lab pricing to ship frontier-quality agentic coding. Composer 2.5 is Cursor's second in-house coding model. It is trained on top of Moonshot AI's Kimi K2.5, an open-weight model released in early 2026 with a permissive commercial license, and it uses a textual-feedback reinforcement learning recipe that Cursor's research team developed over the winter.
The headline numbers are the ones that get product managers off their existing subscriptions. CursorBench v3.1, the company's internal evaluation harness that combines multi-file edits, terminal task execution and long-context codebase Q&A, gives Composer 2.5 a 63.2. Anthropic's Opus 4.7 scores 64.1 on the same test suite. OpenAI's GPT-5.5 scores 62.8. On SWE-Bench Multilingual, which extends the classic SWE-Bench Verified beyond Python to eight additional languages, Composer 2.5 hits 79.8, ahead of Opus 4.7's 78.9 and behind Gemini 3.5 Flash's 80.4.
Pricing is where the story sharpens. Cursor lists Composer 2.5 at $0.50 per million input tokens and $2.50 per million output on the standard tier for Ultra subscribers, with a metered pay-as-you-go option outside the subscription bundle. Opus 4.7 by contrast lists at $15 in and $75 out. That is a 30x delta on input and 30x on output. Even accounting for the subsidy Cursor is presumably applying to its own model to seed adoption, the underlying compute cost gap between running Kimi K2.5-derived weights on Cursor's infrastructure and buying Opus tokens is roughly an order of magnitude.
"Composer 2.5 gets us to frontier-grade agentic coding without paying frontier-grade API markups. The base layer is now something we can train on top of rather than something we have to license." — Cursor engineering post
The training recipe deserves attention. Cursor built a synthetic task pipeline that expands each seed engineering scenario by roughly 25x through mutation. One of the mutations Cursor called out publicly is "feature deletion," in which the model is asked to reverse a previously committed feature cleanly, removing tests, migrations and telemetry hooks alongside the code. Cursor argued that deletion tasks train the kind of holistic codebase awareness that produces better addition tasks. The RL loop is driven by execution feedback and text-based judge feedback rather than human preference labels, which is the mechanism Moonshot itself used for K2 and K2.5.
The Moonshot dependency is the pivot point. Kimi K2.5 shipped with base and instruct weights under a modified MIT-style license that permits commercial use and derivative model release. Cursor's decision to build Composer 2.5 on it rather than on Meta's Llama line or Alibaba's Qwen line reflects both the K2.5 weights being genuinely competitive at the base-model level and the fact that Moonshot has been faster than Alibaba to publish training-recipe details that downstream teams can build on.
Cursor CTO Sualeh Asif addressed the obvious question about lock-in on the launch stream: what happens if Moonshot changes its license terms or if a US export-control action restricts derivative training on Chinese-origin weights.
"We keep a rolling snapshot of every base checkpoint we train against. If any upstream license changed tomorrow, Composer 2.5 as shipped continues to serve. The question is the next generation, not this one." — Sualeh Asif, Cursor CTO
The market response was immediate. Anthropic's May 26 shipment of Opus 4.8, with its 82.4 percent SWE-bench Verified score, can be read as a response to the competitive pressure Composer 2.5 put on Anthropic's core coding franchise. Cursor's user metrics posted to the blog post claim that within 72 hours of the Composer 2.5 default flip, roughly 40 percent of Ultra subscribers had opted to use it as their primary agent, with the remainder split between Claude and GPT.
What Composer 2.5 confirms is that the frontier is no longer a single-vendor concept. Kimi K2.5, Mistral Large 3, LongCat 2.0 and Qwen 3.7 have all pushed the base-layer capability that downstream teams can start from, and Cursor's post is the clearest demonstration to date that an application-layer company with a defensible product surface can train a genuinely competitive model on that base. The economic implication for closed-lab pricing power is what enterprise procurement teams will chew on for the rest of the quarter.
TechApplication
Google I/O 2026 lands Gemini 3.5 Flash, Spark personal agent, Antigravity 2.0 and Managed Agents in one keynote
Google · 2026-05-19 · Source ↗
Google I/O 2026 lands Gemini 3.5 Flash, Spark personal agent, Antigravity 2.0 and Managed Agents in one keynote
Google · 2026-05-19 · Source ↗
Google's I/O 2026 keynote on May 19 was the densest product event of the quarter. Sundar Pichai opened with a token-throughput chart that has since been screenshotted more than any other slide of the year. In May 2024, Google was serving 9.7 trillion tokens per month across all Gemini products. In May 2025 that number was 480 trillion. In May 2026 it is over 3.2 quadrillion, a roughly 7x year-over-year jump against a much larger base. Gemini app monthly active users doubled to 900 million. Search AI Overviews reach 2.5 billion MAU, essentially all of Search itself now touching an LLM-generated summary layer.
Gemini 3.5 Flash entered general availability at the keynote. Google's benchmark deck put Flash at 4x the tokens-per-second of comparable frontier models and under half the cost. Pricing landed at $0.30 per million input tokens and $2.50 per million output, positioning Flash as the workhorse tier below the reserved Gemini Ultra tier for premium reasoning tasks. On SWE-Bench Multilingual, 3.5 Flash scored 80.4, ahead of Composer 2.5 and Opus 4.7. On the LMArena creative-writing panel it landed in the top three.
Spark was the product surprise. Google described Spark as turning any Gemini subscription into a 24/7 personal agent that runs on a dedicated cloud VM assigned to the user. The VM persists between sessions, holds installed tools and credentials, and executes long-running tasks in the background. Google demoed a Spark instance that a user had asked to monitor a house-buying market segment; over the demo period Spark had opened 340 listing pages, filtered them against user criteria, drafted three offer letters and queued a video call with a mortgage broker.
"Spark is Gemini leaving the chat window. The chat is now a way you check in on work that is already happening." — Sundar Pichai, Google CEO
Antigravity 2.0, launched at the same keynote, is the developer-facing sibling of Spark. Where Antigravity 1.0 had been Google's IDE-adjacent agent surface, 2.0 introduces what Google calls a "desktop home" for autonomous agent cohorts. A developer defines a set of agent roles, and Antigravity spins up a persistent surface where those agents work in parallel and can be reviewed asynchronously. The desktop metaphor is deliberate: each agent has a window, a working directory, and a state that survives the developer closing their laptop.
Managed Agents was the enterprise announcement. Google Cloud will host, monitor, patch and scale customer-defined agent workloads on TPU 8t and 8i infrastructure with SLAs. Pricing is per active agent-hour rather than per token, which is the pricing pivot Amazon and Microsoft have been experimenting with more quietly. Managed Agents customers named on the keynote slide included Deloitte, Delta, Bayer, Salesforce and HDFC Bank.
"We are moving from token pricing to agent-hour pricing. That is the pricing model enterprise procurement can actually plan against." — Thomas Kurian, Google Cloud CEO
The infrastructure numbers deserve their own paragraph. TPU 8t is the training-optimized generation Google announced with 4.6 exaflops per pod at FP8 precision. TPU 8i is the inference generation that Google claims delivers 3.2x the tokens-per-dollar of the H200 for Gemini-family workloads. Google's Q1 2026 disclosures had Google Cloud capital expenditure at $18.2 billion for the quarter, more than double the year-ago figure, and the keynote confirmed that a majority of that spend is going into TPU rather than Nvidia.
The quiet undercurrent of the keynote was Google's positioning against the memory-constrained hyperscale environment. With Samsung and SK Hynix in the middle of the RAMageddon-era HBM crunch, Google's ability to run its own silicon end-to-end with its own memory supply arrangements had turned into a competitive moat. Pichai did not say the word "shortage" once. He didn't have to.
What I/O 2026 established is that Google is now shipping across all four layers of the stack at once: silicon (TPU 8t/8i), foundation models (Gemini 3.5 Flash and Ultra), user-facing agents (Spark), and enterprise-managed agent infrastructure (Managed Agents on Antigravity 2.0). Whether the personal-agent bet on Spark converts into consumer stickiness will play out over the second half of the year, but the throughput and MAU numbers suggest Google is executing at a pace its competitors are having to react to.
Mistral AI's release of Large 3 on May 20 was the first genuinely permissive frontier-grade open-weights model of 2026, and the license was the news as much as the model. Large 3 comes in at 675 billion total parameters with 41 billion active per forward pass. Both the base checkpoint and the instruction-tuned checkpoint were published to Hugging Face under Apache 2.0 with no commercial restrictions, no usage-cap gating, and no watermark obligations. That set of terms is stricter about Mistral's own rights than any comparable release since Meta's Llama family, which had always attached custom license terms above a user-count threshold.
The architecture is a mixture-of-experts. It is Mistral's first MoE frontier release since the original Mixtral in December 2023, a gap that had led some observers to write off Mistral's ability to compete with dense US frontier releases. Mistral's technical report says 32 experts per MoE layer, top-2 routing at inference, and a shared attention layer to reduce the memory cost of long-context serving. The context window is 256K tokens with a 1M-token variant expected in the July follow-up.
Training ran on 3,000 H200 GPUs over roughly eight weeks. Mistral disclosed the raw compute budget as 3.7e25 FLOPs, which places the training run below the EU AI Act's systemic-risk threshold by design. Guillaume Lample's blog post accompanying the release framed the compute discipline as strategic rather than incidental.
"We deliberately trained below the 10^26 FLOP threshold in the AI Act. Large 3 is a frontier-grade model that European enterprises can deploy without triggering systemic-risk obligations." — Guillaume Lample, Mistral chief scientist
On benchmarks Large 3 lands in a defensible position. MMLU-Pro at 82.3. GPQA Diamond at 69.4. HumanEval+ at 92.7. SWE-Bench Verified at 71.2, which is behind Opus 4.7 and Composer 2.5 but ahead of Llama 4 Maverick and most other open-weights models available in May. What the raw numbers do not capture is that Large 3 gives European infrastructure providers, sovereign-cloud operators and heavily regulated enterprises a credible frontier-grade base model that is not routed through an American lab relationship.
The political context matters. In May 2026 the EU is still living with the aftershocks of the June 2025 US export-control tightening around advanced semiconductors, and Austrian State Secretary Alexander Pröll's letter to Commissioner Henna Virkkunen a month later pressing for "the strategic establishment of Anthropic within the European Union" made clear that the bloc's political leadership had decided sovereign AI capacity was a legitimate industrial priority. Mistral Large 3 is what that priority looks like in practice: a French company shipping a permissively licensed frontier model that any European ministry, hospital, insurer or bank can host on their own infrastructure.
"We have shipped a model Europe can build on without asking permission." — Arthur Mensch, Mistral CEO
The Apache 2.0 license means that derivative training, redistribution and closed-source fine-tuning are all permitted without royalty. Downstream projects moved quickly. Hugging Face reported that Large 3 base was the most-downloaded model on the platform for the seven days following release. The German BSI (federal information security office) confirmed within 48 hours that Large 3 had been added to its list of models approved for handling internal classification-restricted material at Deutsche Telekom and OTE customers.
Mistral's business model with Large 3 does not depend on API metering the base model. The company will offer a hosted managed service through Mistral Compute, which the CEO's post confirmed would launch in June, and will monetize through fine-tuning services, enterprise support contracts and vertical variants. Guillaume Lample noted that Mistral's Series G raise closed in April with a €5.4 billion post-money valuation on that exact business shape, and Large 3 is the technical proof point the round was underwritten against.
The competitive implication for closed-lab pricing is direct. If Large 3 base at 675B parameters can be self-hosted by an EU enterprise for less than the marginal cost of buying tokens from Anthropic or OpenAI, and if the benchmark gap on standard reasoning tasks is small enough to be closed by domain fine-tuning, then Mistral has moved the bar for what closed-lab pricing has to justify. The June launch of Mistral Compute, and the follow-up 1M-context Large 3 variant in July, will determine whether that pressure sticks.
Tech
Alibaba reveals Qwen 3.7 Max as a 1M-context agent-first reasoning model
MarkTechPost · 2026-05-21 · Source ↗
Alibaba reveals Qwen 3.7 Max as a 1M-context agent-first reasoning model
MarkTechPost · 2026-05-21 · Source ↗
Alibaba unveiled Qwen 3.7 Max on May 21 with a positioning that made explicit what had been implicit in the Qwen roadmap for a year. The model is closed-weight, API-only, and designed for long-horizon agent workloads rather than single-shot chat. Context extends to 1 million tokens, double the 256K ceiling of the previous Max generation. On Terminal-Bench Hard, the harness that measures how well a model can drive a Unix shell across multi-step tasks with real filesystem side effects, Qwen 3.7 Max scores 50.8, up 6.9 points from Qwen 3.5 Max. On the Artificial Analysis Intelligence Index, Qwen 3.7 Max lands at fifth position overall, the highest a Chinese-lab model has placed on that ranking.
The distribution choice is the news. Every prior Qwen Max generation had been available with weights released to Hugging Face under Alibaba's Qwen License, which permitted commercial use up to 100 million monthly active users. Qwen 3.7 Max breaks that pattern. Weights will not be published. The model is available exclusively through Alibaba Cloud Model Studio, OpenRouter, and Together AI, on standard API pricing at $2.20 per million input tokens and $10 per million output.
Alibaba framed the split explicitly. The open Qwen 3.x line, culminating in Qwen 3.6 27B and the anticipated Qwen 3.7 open series expected in June, remains the vehicle for research access and for downstream fine-tuning by third parties. The closed Qwen Max line, starting with 3.7 Max, is where Alibaba invests the additional post-training work required for agent-grade behavior and where it captures the revenue that funds that work.
"Two tracks. Open weights for the research community, closed weights for the agent workloads that customers pay for. This is how the program scales without compromising either side." — Qwen team blog
The structural parallel is OpenAI's o-series split. OpenAI publishes open-weights research models sporadically while reserving its most capable reasoning line for API-only distribution. Qwen 3.7 Max is Alibaba making the same choice, and it removes any ambiguity about whether Chinese labs will indefinitely publish frontier open weights. The commercial pressure of running agent-grade infrastructure at scale is the same in Hangzhou as it is in San Francisco.
The Terminal-Bench Hard score deserves unpacking. The benchmark evaluates a model on real Unix shell tasks that require multiple correct steps, correct handling of tool output, and correct recovery from errors. A score of 50.8 means the model completes just over half of the panel's tasks without human intervention. Opus 4.8 scores 58.1 on the same benchmark. GPT-5.5 scores 55.4. LongCat 2.0 from Meituan scores 52.6. Qwen 3.7 Max is inside the frontier band on this test, and Alibaba's writeup emphasized that its improvement between generations was concentrated in multi-tool orchestration and in long-context planning across the extended 1M-token window.
"Qwen 3.7 Max is the first Alibaba model that ships assuming the primary user is a scaffolded agent loop, not a human in a chat box." — MarkTechPost analysis
The 1 million token context is engineered for the enterprise codebase and legal-review use cases that Alibaba Cloud has been targeting through Model Studio. A packing test in the model card shows the full Linux kernel source at 27 million tokens summarized cross-file across a rolling window; a legal panel across 12 large M&A filings totaling 890,000 tokens; and a full academic bibliography of a graduate thesis with citation cross-checks. Alibaba's engineers emphasized that memory-efficient attention was retrained rather than bolted on, which is what allows the pricing to remain constant across the window.
The competitive read is that China now has three genuinely frontier-grade labs shipping in production. Meituan with LongCat 2.0, DeepSeek with the V4 line, and Alibaba with Qwen 3.7 Max. Between them the labs cover most of the capability space the US frontier occupies, and they do so at pricing that continues to run well below the Anthropic and OpenAI tiers. The closed-weights turn for Qwen Max signals that Chinese labs are converging on the same commercial architecture as their US counterparts: publish enough to keep researcher goodwill, monetize the tier that customers actually deploy against.
The question for the second half of 2026 is whether US export-control policy, tightened again on May 31 through the BIS Blackwell-Rubin rule, will restrict Alibaba's ability to buy the training silicon that keeps Qwen Max on the frontier. Alibaba has been aggressive about domestic ASIC procurement and about co-development with Cambricon, and the answer to that question will shape whether Qwen 4 Max ships on the same cadence.
Policy
Trump scraps 90-day AI executive order hours before signing
TechCrunch · 2026-05-21 · Source ↗
Trump scraps 90-day AI executive order hours before signing
TechCrunch · 2026-05-21 · Source ↗
President Trump was scheduled to sign an executive order on May 21 that would have established a 90-day pre-release review requirement for frontier AI models trained above a compute threshold. The order had been circulating in draft through the White House policy shop for six weeks. It would have required labs to submit model weights and evaluation results to a designated federal reviewer, and to receive clearance before public release. Compute thresholds mirrored the definitions used in the Biden-era October 2023 order that Trump had rescinded on his first day in office.
According to TechCrunch's reporting, which draws on four sources inside the White House and two industry executives briefed on the meetings, the signing was postponed roughly four hours before Trump was scheduled to enter the Oval Office. The pull came after David Sacks, the White House AI and Crypto czar, brought a joint objection from Meta CEO Mark Zuckerberg and Elon Musk to the President's attention during a morning meeting.
Sacks argued that the 90-day review would push US labs to conduct final training and release from foreign jurisdictions to avoid the delay. Zuckerberg's argument was that Meta had been operating under a de facto self-regulation regime for Llama releases and that the executive order would functionally create a competitive advantage for closed-lab competitors who could afford the review headcount. Musk objected on First Amendment grounds, arguing that federal pre-review of model outputs constituted prior restraint.
"There is no federal framework, and there is not going to be a federal framework this year. What we have instead is export-control policy and state-level rulemaking." — Senior White House official quoted by TechCrunch
The Leading the Future super PAC, funded principally by Greg Brockman, had been the loudest external voice arguing for federal preemption of state AI regulation. The theory of the case was that a single federal framework, even a restrictive one, would prevent the Colorado, California, New York and Illinois state laws from creating a fifty-jurisdiction compliance patchwork. The May 21 shelving of the executive order removed the federal framework Leading the Future had been building political air cover for.
The immediate political effect was to leave the AI policy vacuum that had defined the first half of 2026 in place. State laws remained the operational regulation. Colorado's AI Act reached its June 30 effective date already repealed and replaced by SB 189. California's SB 1047 remained on the books. New York's algorithmic accountability rules were in the middle of a comment period. Illinois HB 3773 was moving through committee.
"The absence of federal action means every serious AI company is now planning against a fifty-state compliance surface. That is not a stable equilibrium." — David Sacks, White House AI and Crypto czar
The May 21 shelving also set the stage for what happened three weeks later. Commerce Secretary Howard Lutnick's June 12 suspension of Anthropic's Fable 5 release moved through export-control authority rather than through any AI-specific executive order. That mechanism was available precisely because the executive order route had been foreclosed. The Commerce Department has clear statutory authority under the Export Administration Regulations to restrict the transfer of controlled technology to foreign persons; the interpretation that a US-hosted API endpoint could constitute such a transfer if a foreign-national customer queried it was the legal innovation that made the Fable 5 suspension possible.
Industry response to the shelving was mixed. Anthropic declined to comment publicly on May 21 but had been on record throughout the spring supporting the pre-release review framework as a stability mechanism. OpenAI had lobbied against the specific 90-day window on the grounds that it would slow US labs relative to Chinese competitors. Google's public statement said the company would continue to work with the administration on any successor framework.
The longer-term implication is that federal AI policy is now being made by the Department of Commerce, the Department of Defense, and the Department of State through their respective authorities over export controls, procurement and foreign relations. Congress has not passed a stand-alone AI statute. The White House has not issued an executive order. The federal AI policy that exists is the accumulation of specific agency actions taken under pre-existing legal authorities. That is the environment in which the June Fable 5 suspension, the July NIST reporting guidance, and the anticipated Q4 DOD acquisition rules will all be interpreted.
Anthropic released Claude Opus 4.8 on May 26 across every distribution surface at once. The Claude API had it live at 09:00 Pacific. Claude Code shipped the update behind a version bump the same morning. Cursor rolled Opus 4.8 into its model picker within four hours of API availability. Anthropic's first-party product surfaces at claude.ai and the Claude mobile apps flipped to Opus 4.8 as the default Opus tier before the end of the working day.
The number that carried the release is 82.4 percent on SWE-bench Verified. That put Opus 4.8 at the top of the public leaderboard on the standard programming-agent benchmark. GPT-5.5 was at 78.6 percent. Gemini 3.5 Flash was at 76.2 percent. The prior Anthropic release, Opus 4.7 shipped in April, had been at 79.9 percent. A 2.5 percentage point improvement in six weeks against a benchmark that has been the industry's shared coding measure since 2024 is a meaningful move.
Pricing held flat. Opus 4.8 lists at $15 per million input tokens and $75 per million output. Anthropic had been under pressure to cut Opus pricing after the mid-May release of Cursor's Composer 2.5, which delivered comparable agentic-coding capability at roughly one-tenth the API cost by being built on top of Moonshot's open Kimi K2.5 weights. Anthropic declined to move price. The company's product post argued that Opus 4.8's constitutional-AI improvements on tool-use refusal rates and its wider extended-thinking budget justified the premium.
"We are shipping capability that customers pay for on the merits. If the merits stop being there, we will revisit price. They are here today." — Anthropic product post
The constitutional-AI improvements deserve unpacking. Anthropic disclosed that Opus 4.8's tool-use refusal rate on the internal Agent Safety Evaluation harness dropped from 4.1 percent under Opus 4.7 to 1.6 percent under Opus 4.8. Tool-use refusals had been a specific customer complaint through the spring, particularly among Claude Code users whose agent loops were being interrupted by refusals to run terminal commands the model treated as ambiguously dangerous. The reduction to 1.6 percent brings Opus 4.8 into competitive range with GPT-5.5 on that specific dimension.
The wider extended-thinking budget is the second unlock. Opus 4.7 had capped extended thinking at 32,000 reasoning tokens per request. Opus 4.8 raises the cap to 128,000, with automatic budget allocation based on task difficulty rather than a fixed per-request setting. In practice this means an Opus 4.8 agent working on a complex debugging task can allocate significantly more reasoning-token budget to that task without an explicit configuration change, and Anthropic's internal telemetry shows the median request uses roughly 8,200 extended-thinking tokens against the new 128K ceiling.
"Extended thinking is now an elastic resource rather than a fixed one. The model decides how much to think about each request, and pricing follows the actual usage." — Anthropic model card
Distribution partners moved quickly to make the update visible. Cursor's May 26 release notes flipped Opus 4.8 to the default Claude selection in the model picker. Windsurf did the same within 48 hours. GitHub Copilot's enterprise tier, which had begun offering Claude as an alternative to OpenAI models in March, made Opus 4.8 available on May 27. Zed and JetBrains AI Assistant followed within the week.
The strategic context sits between two adjacent events. Two weeks earlier, Cognition had closed its $1B Series B at a $26B valuation on the strength of Devin's coding-agent traction. Two weeks later, Anthropic would ship Fable 5 on June 9 and see it suspended by Commerce Secretary Lutnick on June 12 through export-control authority. Opus 4.8 was the last frontier release Anthropic shipped without the Fable-shaped complication of federal export-control review, and the pricing decision suggests the company was preparing for a period in which the top-tier Opus model would carry both the technical leadership burden and the revenue burden.
The financial implication showed up in Anthropic's next disclosure. On May 29, three days after Opus 4.8 shipped, Anthropic closed a $13 billion primary raise at a $965 billion post-money valuation and disclosed Q1 ARR at $44 billion, ahead of OpenAI's $25 billion. Opus 4.8 was the product artifact that anchored the round. The 82.4 percent SWE-bench Verified number was the datapoint the term sheet was written against.
Economic
Anthropic raises $13B at $965B valuation and discloses $44B ARR, crossing OpenAI
TechCrunch · 2026-05-29 · Source ↗
Anthropic raises $13B at $965B valuation and discloses $44B ARR, crossing OpenAI
TechCrunch · 2026-05-29 · Source ↗
Anthropic closed a $13 billion primary raise on May 29 at a $965 billion post-money valuation. The round, led by Fidelity with participation from BlackRock, T. Rowe Price, GIC, ADIA and Google (which extended its existing strategic position), pushed Anthropic's paper valuation past the $850 billion mark that OpenAI's March secondary had set. The disclosure that accompanied the term sheet was as important as the valuation itself: Anthropic booked $44 billion in Q1 2026 annualized recurring revenue against OpenAI's reported $25 billion.
The ARR crossover reframes what the market has been assuming since 2023. OpenAI has been the assumed revenue leader through the entirety of the ChatGPT era. Anthropic's disclosure that its Q1 2026 ARR is 76 percent above OpenAI's confirmed what a series of enterprise-procurement anecdotes through the spring had been pointing at. Claude Code is now the dominant enterprise coding surface. Claude API traffic through the top five programming-agent products (Cursor, Windsurf, Zed, Claude Code itself, and GitHub Copilot Claude tier) is running at roughly 2.4x the OpenAI equivalent by revenue.
"The next 18 months are about scaling, not searching. Two labs are at product-market-fit-level scale, and everyone else in the frontier tier is a distant third or worse." — Simon Willison, blog post
Simon Willison's May 27 reading of the emerging landscape had anticipated the disclosure by two days. Willison's argument was that Anthropic and OpenAI had both crossed the threshold at which the operational question was no longer "how do we find product-market fit" but "how do we scale the product-market fit we already have." That framing has held up. The May 29 round closes the debate about whether Anthropic is at OpenAI's tier commercially; the answer is that it is ahead.
The Karpathy hire on May 19 is the personnel move that anchored the pre-training story the round was written against. Andrej Karpathy, an OpenAI founding engineer who had left to start Eureka Labs in 2024, joined Anthropic's pre-training team in a technical leadership role. Karpathy's move pulled him out of the OpenAI founding-engineer alumni pool and onto Anthropic's roadmap. The specific work Karpathy has been assigned is post-training data curation for the successor to the Opus and Fable lines, with the target ship date being Q4 2026.
"Anthropic is a place where I can go deep on the parts of the training pipeline I think are most under-invested. That is where I have wanted to focus." — Andrej Karpathy, personal blog post
The valuation math is worth spelling out. At $965 billion post-money and $44 billion Q1 ARR (which annualizes to $176 billion if the trajectory is flat, though Anthropic's growth curve suggests the actual 2026 revenue will land higher), Anthropic is trading at roughly 5.5x forward revenue. That is a multiple that public-market software companies have not sustained since the 2021 SaaS peak, but it is defensible for a company growing revenue at Anthropic's rate. The comparison to OpenAI at $850 billion valuation on $25 billion ARR (34x trailing) suggests the market was applying an Anthropic-implied discount to OpenAI, which the May 29 round has now made explicit.
The round sets up two adjacent events. The first is the June Series H, which Micron participated in as part of its May 30 memory-supply announcement that named Anthropic as a strategic customer alongside Nvidia. The Series H is expected to close at a higher valuation still, with the primary purpose of funding the compute buildout that Fable 5 and its successors require. The second is the June 9 Fable 5 launch and the June 12 Lutnick suspension, which will test whether Anthropic's revenue trajectory holds under a federally imposed distribution restriction.
The competitive read for the rest of the frontier tier is uncomfortable. Google is the only company with the infrastructure independence to compete on the compute axis. Meta's Muse Spark has not yet shipped at scale. xAI is competing on price rather than on capability leadership. Meituan's LongCat 2.0 and Alibaba's Qwen 3.7 Max are the two Chinese labs that could plausibly close the capability gap, but both are constrained by domestic silicon supply as export controls tighten. The Anthropic-OpenAI two-lab framing that Willison articulated is the operational reality for US enterprise procurement in the second half of 2026.
What the $965 billion valuation does not price is the political risk that materialized 11 days later when Commerce Secretary Lutnick suspended Fable 5 under export-control authority. That story belongs to June, but the May 29 round is the event that made Fable 5 a large enough commercial artifact for Commerce to treat as a strategic export in the first place.
Cognition AI closed its Series B on May 13 at $1 billion primary and $26 billion post-money valuation. Founders Fund led the round with Founders X and 8VC as co-leads. It is the third-largest AI Series B on record, behind Anthropic's 2024 Series C and Mistral's 2025 Series F. Cognition's run-rate jumped roughly 13x over the twelve months preceding the round, from approximately $30 million ARR in mid-2025 to approximately $400 million ARR heading into the close.
The company's core product is Devin, the autonomous coding agent that shipped to closed beta in March 2024 and reached general availability in Q4 2024. Devin's positioning has been distinct from the IDE-embedded agents that Cursor, Windsurf and Claude Code offer. Devin runs asynchronously on its own cloud VMs, is assigned work through a task-queue interface, and returns pull requests and status updates rather than editing inside a live IDE session. That positioning had been the source of persistent industry skepticism through 2025, with critics arguing that asynchronous agents lacked the interactive feedback loop that made in-IDE agents useful.
The revenue trajectory has settled that debate. Cognition's April public metrics disclosure showed 68 percent of Devin usage concentrated in enterprise customers running agent cohorts on medium-scoped tasks: bug fixes with specifications, feature additions with acceptance criteria, and dependency upgrades with test-suite requirements. Those are exactly the tasks where the asynchronous model beats the interactive model, because the human is not blocked waiting for the agent to think.
"The IDE agent and the autonomous agent are complements, not competitors. The IDE is for work you are doing right now. Devin is for work you would rather not do at all." — Scott Wu, Cognition CEO
Bloomberg had reported in April that Cognition was in discussions at a $25 billion valuation. The final $26 billion post-money mark is one billion above that reported floor, which suggests demand for the round from strategic co-investors pushed the price up in the final weeks. The $26 billion valuation puts Cognition roughly on par with Cursor's most recent $22 billion secondary and above Windsurf's OpenAI-acquisition mark from Q4 2025.
The competitive geography of coding agents is now three primary vendors. Anthropic Claude Code shipped as an official product in early 2025 and is the enterprise workhorse for interactive IDE-embedded work, particularly at customers who have Claude API contracts. Cursor is the IDE-first vendor that has moved up-stack into training its own models (Composer 2.5 was released five days after the Cognition Series B closed). Cognition is the autonomous-agent vendor whose primary user is a project manager assigning work rather than an engineer typing.
"Three companies have credible enterprise footprints in coding agents. Everyone else is chasing." — Bloomberg analysis, citing enterprise procurement data
The Series B use-of-proceeds breaks into three buckets according to Cognition's disclosure. First, expansion of the Devin compute pool, which currently runs on a mixed AWS and CoreWeave footprint and will add roughly 8,000 additional GPUs of dedicated capacity through the second half of 2026. Second, extension of the Devin platform into adjacent verticals: legal document review, security auditing, and data-pipeline maintenance, all of which map to the same asynchronous-task pattern the coding product has proven. Third, international expansion, with London and Bangalore offices opening in Q3 to serve European and Indian enterprise customers.
The founders'-round dynamic is worth naming. Founders Fund and Founders X are both named on the term sheet. Peter Thiel's Founders Fund had been an early Cognition backer through the seed and Series A, and Founders X is a newer vehicle Thiel launched in 2025 for follow-on positions in portfolio companies at growth stage. The concentration of Thiel-affiliated capital in the Series B is one of the signals that Cognition is being groomed as a strategic asset in the Thiel-network coding-agent bet, alongside Palantir's expanding Claude integration disclosed the same week.
The Cognition round sits inside a broader capital-formation quarter for AI-native application companies. Cursor's April secondary at $22 billion, Cognition's May Series B at $26 billion, and Anthropic's May 29 primary at $965 billion together represent roughly $15 billion of new AI capital committed in a six-week window. That capital is flowing into companies whose revenue trajectories have moved from speculative to demonstrated. The 2026 AI investment thesis has shifted from betting on unproven agent capability to underwriting proven agent revenue at scale.
The question Cognition faces in the second half of 2026 is whether the asynchronous-agent model scales beyond software engineering. The London and Bangalore expansion, and the vertical extensions into legal and security, are the practical tests. If the answer is yes, the $26 billion valuation is defensible on a next-round basis. If the answer is that Devin remains a coding-specific product, the valuation looks stretched against the underlying revenue base.
Tech
OpenAI reasoning model disproves an 80-year-old Erdős conjecture
New Scientist · 2026-05-15 · Source ↗
OpenAI reasoning model disproves an 80-year-old Erdős conjecture
New Scientist · 2026-05-15 · Source ↗
New Scientist reported on May 15 that an OpenAI internal reasoning model had produced a 125-page disproof of a 1946 conjecture by Paul Erdős concerning the maximum number of unit distances that can occur among n points in the plane. The result had been circulating in a small verification circle since March, and the May 15 story broke it publicly along with the panel's assessment. The nine-mathematician verification panel was chaired by Fields medalist Tim Gowers and included Noga Alon, a specialist in combinatorial and discrete geometry who had spent portions of his career working on adjacent problems.
The Erdős conjecture in question is the unit-distance conjecture, which for eight decades had been a working assumption in combinatorial geometry. The specific formulation was that the maximum number of pairs of points at unit distance among n points in the plane grows more slowly than a particular functional bound. Erdős had proposed the conjecture in 1946 and had returned to it repeatedly through his career. Various weaker bounds had been proved, and a large body of work had accumulated around what tools might eventually resolve it, but the conjecture itself had remained open.
The OpenAI model's disproof constructs a family of point configurations that exceed the conjectured bound for sufficiently large n. The construction uses algebraic number theory tools that Alon and Gowers both flagged as unusual for a model to have produced. The proof includes a 40-page lemma involving properties of certain quadratic extensions of the rationals, a lemma that the panel independently verified against known results in the underlying number-theoretic literature.
"This is an outstanding achievement. The proof applies fairly sophisticated tools from algebraic number theory in a way that a mathematician reading the work has to genuinely engage with rather than skim." — Tim Gowers, panel chair
Gowers's endorsement is the pivotal statement. Gowers had been one of the most public skeptics of AI-generated mathematics through 2024 and 2025, and had written extensively about the failure modes of LLM-produced proofs. His willingness to attach his name to the panel's verdict of "outstanding achievement" is a data point about where the state of the art actually is.
The panel's verification process took roughly six weeks. Each of the nine mathematicians was assigned a section of the proof and worked through it independently. Weekly video meetings were used to resolve overlaps and to identify passages that required deeper scrutiny. The one substantive correction the panel identified was a sign error in a lemma on page 87 that had propagated into a subsequent claim; OpenAI's model corrected the error in a follow-up pass within 48 hours of the panel's flagging.
"The interesting fact is not that the model produced a proof. It is that the model produced a proof that a group of specialists needed six weeks to verify, and in the process the specialists learned something about the problem." — Noga Alon, panel member
OpenAI's public statement about the result was more restrained than the New Scientist framing. The company described the model as "an internal reasoning system" and did not name a version. The company also did not commit to publishing the model or to making it available externally. That posture is consistent with how OpenAI has handled internal reasoning-model releases through 2025 and early 2026, in which the most capable reasoning models are used for internal research and are not released as APIs.
The signal the result carries for the broader AI research community is that reasoning models have crossed a productivity threshold, at least for narrow mathematical work. The threshold is not that reasoning models are now doing all mathematics. Reasoning models remain weak on problems that require large-scale conceptual reframing, and the Erdős result was a problem for which the tools existed and the model's contribution was identifying the correct combination of them. The threshold is that reasoning models can now produce publishable original mathematics on decades-old open problems, and that fact will change how research mathematics is conducted at the margin.
The Fields Medal community's response has been split. Several senior mathematicians outside the verification panel have expressed concern that reasoning-model-produced proofs will accumulate in the arXiv without adequate verification bandwidth to keep up. Others have argued that the correct response is to invest in formal verification infrastructure (Lean, Coq) so that model-produced proofs can be mechanically checked, reducing the burden on human specialists. Both responses are practical adaptations to a technology that has now been demonstrated.
The economic implication is smaller than the intellectual one. Research mathematics is a small market. But the demonstration that reasoning models can produce publishable original work at the frontier of a specialized discipline has implications for adjacent scientific fields that have been slower to engage with the technology. Theoretical physics, algorithmic complexity theory, and formal methods in computer science are all fields in which the same pattern could plausibly extend. The Erdős result is the proof-of-concept event for that broader extension.
TrendEconomic
Uber, Amazon and Salesforce switch from tokenmaxxing to token rationing
Wall Street Journal · 2026-05-30 · Source ↗
Uber, Amazon and Salesforce switch from tokenmaxxing to token rationing
Wall Street Journal · 2026-05-30 · Source ↗
The Wall Street Journal's May 30 story on enterprise LLM budget pressure named the phenomenon that finance teams at large-cap technology companies had been dealing with quietly through the spring: the FY 2026 token budget is now the binding constraint on internal AI usage, not availability, not capability, not even organizational readiness. The reporting anchored on four specific cases.
Uber's finance team told the Journal that the company had exhausted its 2026 Claude Code token allocation in four months. The allocation had been set in November 2025 during Uber's annual planning cycle based on the actual usage patterns from Q3 and Q4 2025, when Claude Code was still ramping inside the engineering organization. Actual 2026 usage ran roughly three times the planned trajectory as engineers moved more work to agent-mediated pull requests. Uber's response has been a per-engineer weekly quota, enforced through the internal Cursor and Claude Code deployments, with escalation to engineering leadership required to exceed the quota.
Amazon's story is different. The company had launched an internal leaderboard called Kiro Rank in Q1 2026 that ranked engineers by the number of AI-mediated pull requests they landed. The intent had been to accelerate adoption. The effect, according to the Journal's sources, was that engineers began gaming the leaderboard by decomposing what would have been single pull requests into chains of small ones, and by using AI tools for work that did not require them. Amazon killed Kiro Rank on May 22 and issued a memo that AI-tool usage metrics would no longer be included in performance evaluations.
"We optimized the wrong thing. Leaderboards produce leaderboard behavior. What we actually want is good software shipped, and that is not measured by the count of agent-mediated commits." — Amazon internal memo, quoted by WSJ
Microsoft's response has been to cut Claude Code subscriptions across several divisions. The Journal sourced the specific decision to the Azure division, where Microsoft had been paying for Claude Code seats alongside GitHub Copilot in a period of parallel evaluation. As of May 15, Claude Code subscriptions were removed from the standard Azure engineering benefits package, and access is now granted on individual-approval basis. Microsoft's own model roadmap through Copilot and its OpenAI partnership makes the strategic redundancy of paying for Claude infrastructure a defensible cost-cut target.
Salesforce is the largest single case. The company's FY 2026 Anthropic bill is projected at $300 million, up from roughly $80 million in FY 2025. The growth is driven by Agentforce, Salesforce's customer-facing agent platform that runs on Claude Sonnet and Opus tiers, and the internal Claude Code deployment that Salesforce engineering has standardized on. Salesforce's response has been to negotiate a committed-use discount with Anthropic and to invest in prompt-caching and request-batching infrastructure that reduces per-request token counts.
"The tokenmaxxing phase is over. We are in the token-rationing phase, and the finance function is now the effective governor of internal AI capability." — Wall Street Journal analysis
The METR anecdote in the WSJ story is the one that circulated most widely on social media. METR, the Model Evaluation and Threat Research organization, had planned a randomized controlled trial measuring the productivity effect of AI coding assistants versus a no-AI control. The study was abandoned in April because the coders assigned to the no-AI control condition refused to complete their assigned tasks under those conditions. The finding is not that AI tools are or are not productive. The finding is that the population of professional software engineers has moved past a threshold at which AI-assisted work is the default and unassisted work is treated as an artificial constraint.
The economic mechanism the WSJ story describes is straightforward. Enterprise LLM consumption at large-cap technology companies has grown from a rounding error in the FY 2025 IT budget into a significant line item in FY 2026. When budget lines cross the threshold at which CFO-level attention is required, they get budget-line-item treatment: annual allocations, quarterly true-ups, per-user quotas, and escalation processes for exceeding limits. The tokenmaxxing culture of 2025, in which engineers were encouraged to burn tokens on any workflow that could conceivably benefit, becomes untenable when the total spend triggers finance-committee reviews.
The story preceded the June 4-16 Claude outage window by five days. Those outages, which affected Claude API availability across all tiers for extended periods, put pressure on the same customers who were already trying to manage their Anthropic spend. The combination of budget constraint and reliability constraint produced the July procurement-review cycle that several of the WSJ-named companies have since publicly acknowledged.
The longer-term implication is that enterprise LLM pricing has entered a phase where the demand side is negotiating from a position of budget discipline rather than capability enthusiasm. That is a different bargaining posture, and it is showing up in the committed-use discount structures that Anthropic, OpenAI and Google have all begun offering to their largest customers through the second quarter.
Policy
BIS closes the China-subsidiary loophole on Blackwell, Rubin and AMD MI350x exports
CNBC · 2026-05-31 · Source ↗
BIS closes the China-subsidiary loophole on Blackwell, Rubin and AMD MI350x exports
CNBC · 2026-05-31 · Source ↗
The Bureau of Industry and Security published a rule on the evening of Sunday May 31 that took effect at 00:01 Eastern on June 1. The rule extended the existing export ban on Nvidia's Blackwell and Rubin accelerator families, and on AMD's MI350x line, to any firm anywhere in the world that is controlled by a Chinese parent entity. The rule closed what enforcement officials had been calling the "subsidiary loophole" and specifically named Malaysia and Singapore as jurisdictions in which the loophole had been exploited at scale through 2025.
The existing rule structure had prohibited direct export of the covered accelerators to entities in China and to entities on the US Entity List regardless of location. Chinese firms had responded by incorporating subsidiaries in Malaysia, Singapore, the UAE and other jurisdictions with light beneficial-ownership reporting, and had used those subsidiaries to purchase covered accelerators and route them into the parent company's compute footprint. The May 31 rule attacks that structure directly.
The enforcement mechanism has three parts. First, US chipmakers must now verify beneficial ownership of any customer purchasing a covered accelerator, tracing the ownership chain to ultimate controlling parties regardless of intermediate holding structures. Second, any purchase that traces to a Chinese parent entity triggers the ban, even if the immediate purchaser is a non-Chinese firm. Third, resale restrictions extend to secondary markets, with penalties for downstream buyers who fail to conduct their own beneficial-ownership diligence.
"This is the most aggressive posture change since October 2022. We are closing the arbitrage that has allowed roughly 30 percent of Nvidia's Southeast Asia sell-through to end up in Chinese compute pools." — BIS official cited by CNBC
The CNBC reporting was sourced from BIS itself and from industry sources briefed on the rule ahead of publication. The 30 percent figure is the enforcement community's estimate of the leakage volume the loophole had enabled through 2025. In dollar terms that maps to roughly $8 to $12 billion of Nvidia sell-through, depending on the accelerator mix, that would need to be redirected or lost.
Cambodia, Vietnam, and the UAE remain permitted destinations under enhanced scrutiny. The enhanced-scrutiny regime requires case-by-case licensing decisions for orders above defined unit thresholds. In practice, industry executives briefed on the process expect the licensing timeline to run 60 to 90 days for orders that clear beneficial-ownership review and to run indefinitely for orders that trigger red flags. That timeline shift alone will significantly reduce the effective shipment rate to those destinations even for legitimate customers.
"The subsidiary loophole was open for eighteen months and everyone in the industry knew it. Closing it was a matter of political will more than a matter of legal drafting." — Former BIS official, quoted anonymously
The commercial implications for Nvidia and AMD are immediate. Nvidia's Q2 fiscal 2027 guidance, provided in the company's May 21 earnings call, had already anticipated some tightening of Southeast Asia rules but had not modeled a full subsidiary-loophole closure. AMD's MI350x pipeline had a somewhat larger relative exposure to Malaysia and Singapore because AMD had been aggressive about establishing sell-through partnerships in those markets to challenge Nvidia's incumbency. Both companies issued short statements on Sunday night confirming that they would comply with the rule and that they were reviewing customer contracts to identify affected orders.
The Chinese response was faster than industry observers had expected. Within 72 hours of the rule taking effect, the Chinese State Council formalized a $295 billion five-year AI infrastructure investment plan that had been circulating in draft since Q1. The plan commits central-government funding to domestic ASIC development at Cambricon, MetaX, and Enflame, and to fab-capacity buildout at SMIC targeting the 5nm node that would be required for competitive AI silicon. The three-week lag between the BIS rule and the State Council announcement is short enough that the two decisions are best understood as coordinated responses to a longer-running strategic contest.
The rule is the most aggressive US posture change on export controls since October 2022, when the initial round of AI-relevant chip restrictions was published. It escalates beyond the October 2023 tightening, the June 2025 wall around Mythos 5 and Fable 5, and the various interim updates through 2024 and 2025. The pattern of escalation has been consistent: each new rule closes a specific arbitrage that had emerged under the prior rule, and each new arbitrage has required roughly 12 to 18 months to fully develop before being closed. The subsidiary loophole was the largest and longest-running such arbitrage to date.
The broader policy implication is that AI export control has become a live area of Commerce Department rulemaking rather than a set of static prohibitions. Industry compliance functions have expanded significantly through 2025 and 2026, and the beneficial-ownership diligence obligation the May 31 rule imposes is the largest new compliance burden the industry has faced. For companies operating in jurisdictions that remain permitted under enhanced scrutiny, the practical effect is that AI infrastructure deployment now requires the same corporate-diligence infrastructure that financial-services firms have long maintained for sanctions compliance.
June 2026 18 stories · Jun 1 — Jun 30 · lead story at top
Economic
DeepSeek closes near $7 billion in China's largest AI funding round yet
Bloomberg · 2026-06-03 · Source ↗
DeepSeek closes near $7 billion in China's largest AI funding round yet
Bloomberg · 2026-06-03 · Source ↗
Bloomberg reporters Lulu Yilun Chen and Sarah Zheng surfaced the DeepSeek round on Wednesday afternoon Beijing time, citing four people with direct knowledge of the term sheet. The Hangzhou lab is close to sealing about $7 billion, the largest single AI equity round ever closed inside China and the first to be anchored by the country's sovereign wealth fund, China Investment Corporation. Alibaba Cloud and Tencent take secondary allocations, alongside a cluster of state-linked venture arms that Bloomberg's sources declined to name for attribution. Nobody involved would confirm a headline valuation. Earlier reporting from the same desk in April put DeepSeek's pre-money mark at roughly $150 billion.
The timing reads as a direct answer to Washington. On June 1, the Bureau of Industry and Security closed the shell-company loophole that had let Chinese buyers route Nvidia Hopper and Blackwell inventory through registered addresses in Kuala Lumpur, Johor Bahru and Singapore. Analysts at TrendForce estimate that channel had supplied between 180,000 and 220,000 accelerators to Chinese labs across 2025. The BIS action closes it. DeepSeek's team, which shipped V4 Pro and V4 Flash across April and May on a mixed fleet of Ascend 910C and grey-market H200 nodes, now has to run a compute strategy that assumes zero further Nvidia intake.
Beijing has been rehearsing the pitch for months. State broadcaster CCTV ran a fifteen-minute segment on Wednesday evening built entirely around DeepSeek co-founder Liang Wenfeng, framing the company as proof that Chinese labs no longer depend on American frontier work. The line landed as Liang stood in front of a chalkboard drawing of a MoE routing diagram, wearing a plain grey sweater, in what looked like a deliberate visual echo of his January 2025 breakout coverage.
"This round tells the world that China does not need to wait for OpenAI or Anthropic to hand us the next generation. We already have it." — Xiao Yaqing, former Minister of Industry and Information Technology, speaking on CCTV Finance
The $7 billion is structured as primary capital with a small secondary component for early High-Flyer Capital employees who bootstrapped the lab out of the quant fund's discretionary bonus pool. Bloomberg's sources describe a five-year lockup on the CIC allocation and a two-year lockup on the corporate strategics. Proceeds are earmarked, in order of size, for a domestic training cluster targeting 40,000 Ascend 920 accelerators when Huawei begins volume shipments in Q1 2027, a 6,000-strong hiring plan across Hangzhou and a new Shenzhen office, and a data-licensing budget large enough to lift DeepSeek out of the copyright grey zone it has operated in since inception.
Liang's team has not commented on the round. A spokesperson forwarded requests to CIC's Beijing press office, which did not respond by publication. What the company has been willing to discuss, in a Zhihu Q&A that DeepSeek chief scientist Guo Damo hosted last Saturday, is the roadmap. V5 enters supervised fine-tuning in July on a corpus that Guo described as "twice the bilingual reasoning tokens of V4 and roughly the same code volume." Guo said the lab has moved to a two-track training strategy: a Nvidia-optimised path that will keep running as long as inventory lasts, and a fully Ascend-native path that treats Huawei's CANN toolchain as first-class. Both paths converge on the same weights.
The competitive read matters as much as the number. A $7 billion round at a $150 billion mark puts DeepSeek in the same funding-scale conversation as Anthropic's Series H and xAI's post-SpaceX-IPO position, while carrying zero dependence on US cloud capacity. Morgan Stanley's China internet team, in a same-day note, called it "the moment sovereign LP conviction in Chinese frontier AI moves from strategic to structural." Barclays framed the round as the anchor for a broader wave: Moonshot, MiniMax and Zhipu are all rumoured to be preparing follow-on rounds priced against DeepSeek's mark.
"We are underwriting a national capability. The commercial model is a byproduct." — A Beijing-based investor quoted by Bloomberg on background
For Washington, the round complicates the theory of change behind the June 1 loophole closure. Export controls on Nvidia hardware were meant to slow Chinese frontier progress. What they appear to have produced is a sovereign-backed capital wall around the labs most exposed to the shortage, and a domestic accelerator roadmap with the funding to see it through. DeepSeek closing $7 billion on the same news cycle that Huawei confirmed its 920 tape-out schedule is the kind of coincidence that stops looking coincidental.
Tech
Nvidia releases Nemotron 3 Ultra, a 550B open-weights hybrid Mamba-Transformer MoE
Nvidia Research · 2026-06-04 · Source ↗
Nvidia releases Nemotron 3 Ultra, a 550B open-weights hybrid Mamba-Transformer MoE
Nvidia Research · 2026-06-04 · Source ↗
The Nvidia Research team posted the Nemotron 3 Ultra weights, technical report and inference kernels on Thursday morning under OpenMDW-1.1, the same permissive license that shipped with Nemotron 3 Base in March. Ultra is the largest open-weights model Nvidia has released, at 550 billion total parameters with roughly 48 billion active per token. The architecture is a hybrid: alternating Mamba-2 state-space blocks handle long-range context, and Transformer blocks handle the tool-use planner path. The MoE routing sits inside the Transformer blocks, with 128 experts and top-8 gating.
The design choice that matters is the elimination of a dedicated KV cache for long contexts. Instead of paying quadratic attention across a rolling context window, Nemotron 3 Ultra pushes long-range state through the Mamba layers as a compressed recurrence. Nvidia's report shows the model holding coherent state across two-million-token agent trajectories with roughly one-tenth the accelerator memory that a pure-Transformer 405B model would need for the same window. The tradeoff, as the paper acknowledges, is that exact retrieval of a specific token from deep in the state is weaker than a Transformer with full attention. Nvidia's answer is to route retrieval-heavy sub-tasks through a tool the model can call, rather than expecting the weights to memorise the transcript.
"We stopped treating context as a passive buffer. The model has to decide what to keep." — Bryan Catanzaro, VP of Applied Deep Learning Research, Nvidia
The tuning corpus centred on multi-turn agent trajectories rather than the general instruction-following data that dominates most post-training pipelines. Nvidia's team ran on-policy tool-use reinforcement learning across 12,000 synthetic environments, spanning shell sessions, browser navigation, GitHub issue triage, spreadsheet manipulation and Kubernetes operations. Each environment came with a verifier, either a unit test suite or a rule-based judge, so the RL signal was grounded rather than a learned reward model. The paper reports 41 million completed trajectories in the RL phase, an order of magnitude beyond what Nvidia disclosed for Nemotron 3 Base.
The benchmark table is the reason the release matters. SWE-Bench Verified lands at 82.1 percent, above Meituan's LongCat 2.0 and within a point of Claude Sonnet 5. Terminal-Bench 2.1 puts Ultra at the top of the open-weights leaderboard, ahead of Qwen 3.6 Coder and DeepSeek V4 Pro. AgentBench composite sits at 68.4, versus Llama 4.5 Composer at 65.9. On raw reasoning benchmarks like MMLU-Pro and GPQA-Diamond, Ultra trails the closed frontier by three to five points, which Nvidia's report attributes to the deliberate choice to weight agent training over knowledge cramming.
The inference story is where Nvidia's business model becomes visible. Ultra runs on 8 H200s at 4-bit AWQ quantisation, with published vLLM and Triton kernels that hit 34 tokens per second per user at a batch size of 32. The same weights at BF16 need 16 H200s or a single GB200 NVL72 tray. Nvidia is not shy about the target customer: the reference deployment guide is written for enterprises standing up on-prem inference clusters, not for hobbyists on consumer hardware. The kernel authors have gone out of their way to make the Mamba-Transformer hybrid run cleanly on Blackwell's new FP4 tensor cores, in a way that CPU or Apple Silicon backends will struggle to match without a rewrite.
"OpenMDW is not charity. Every enterprise that puts Nemotron on H200s is a customer we did not have to acquire." — Jensen Huang, in a CNBC interview the morning of the release
The reaction on Hugging Face was immediate. Within six hours of the release, community fine-tunes had appeared in Japanese, Korean, French and Arabic. Together AI stood up a hosted endpoint at $2.40 per million input tokens and $9.60 per million output, undercutting Sonnet 5's introductory pricing on the output side. Fireworks AI followed with a $2.10 / $8.40 endpoint an hour later. The open-weights market for frontier-adjacent agent models now has three credible entries at the 500B-plus scale: Nemotron 3 Ultra, LongCat 2.0 and Kimi K2.5.
The strategic read is that Nvidia has decided the open-weights layer is a moat rather than a threat. Every enterprise that runs Nemotron on-prem is a Blackwell customer, a CUDA customer and, through NIM microservices, a Nvidia AI Enterprise subscription customer. The weights are free. The stack that makes them run at frontier speed is not.
TechApplication
Apple's WWDC LanguageModel protocol lets Claude and Gemini ship as Swift packages
Apple Machine Learning · 2026-06-08 · Source ↗
Apple's WWDC LanguageModel protocol lets Claude and Gemini ship as Swift packages
Apple Machine Learning · 2026-06-08 · Source ↗
Tim Cook opened the WWDC 2026 keynote at 10:00 Pacific on Monday with the third generation of Apple Foundation Models. AFM 3 On-Device gains native image input, a tighter tool-use grammar that Apple's runtime enforces at the token level, and expanded language coverage into Vietnamese, Thai, Hindi and Modern Standard Arabic. AFM 3 Cloud Pro is the model that broke a five-year Apple pattern: it runs on Nvidia GPUs hosted inside Google Cloud, rather than on Apple Silicon Private Cloud Compute nodes. Craig Federighi framed the shift as a supply-chain pragmatism story. Apple's silicon roadmap could not keep up with the training compute AFM 3 Cloud Pro needed, so Apple bought the compute it needed on the open market.
The genuine news landed forty minutes into the keynote, when Federighi handed the stage to State of the Union for a demo of the new LanguageModel Swift protocol. Any provider whose model conforms to the protocol drops into the Foundation Models framework as a Swift package. Developers write against Foundation Models once. The same call site can be routed to AFM 3 On-Device, AFM 3 Cloud Pro, Claude Sonnet 5, Gemini 3 Pro, or any other conforming model, with a build-time flag or a runtime user preference.
"Foundation Models is the framework. The model is a dependency. That is the mental model we want developers to have." — Craig Federighi, SVP Software Engineering, Apple
Anthropic published a Swift package implementing the protocol at 15:00 Pacific the same afternoon, hosted on the anthropic-ai GitHub organisation with a permissive Apache-2 license on the wrapper. The package ships Claude Sonnet 5, Claude Opus 4.8 and Claude Haiku 4.6, and pipes billing through Anthropic's existing API keys rather than Apple's StoreKit. Google shipped a Gemini Swift package on Tuesday morning with Gemini 3 Pro and Gemini 3.5 Flash exposed, billed through Google Cloud. OpenAI has not shipped a conforming package as of publication, and Sam Altman's Monday-evening tweet ("interesting architecture, will think about it") is being read across the Apple developer community as a stall.
The Dynamic Profiles abstraction is where the protocol becomes interesting for agent developers. A profile is a Swift struct that describes a routing policy: which model handles which type of request, which fallback runs if the primary is unavailable, which tool grammar applies. Xcode 27's new Foundation Models Inspector visualises profiles as a graph. A developer can build a profile that sends short-context chat to AFM 3 On-Device, long-context research to Claude Opus 4.8, image generation to Gemini's Nano Banana 2, and code editing to Claude Sonnet 5, all inside the same app, all with a single call site. Federighi's demo built a research assistant that hit five different models across a single user session.
AFM 3 Cloud Pro pricing is where Apple's business model shift becomes visible. Small-business developers, defined as accounts with under 2 million App Store first-time downloads in the trailing 12 months, get Cloud Pro free. Everyone else pays Apple $12 per million input tokens and $48 per million output, comfortably above what Anthropic charges for Sonnet 5 and roughly at parity with Opus 4.8. Federighi did not name the margin. Analysts at Bernstein estimate Apple is paying Google Cloud roughly $3 to $4 per million input tokens at wholesale, which implies a 60 to 70 percent gross margin on the Cloud Pro line before Nvidia hardware amortisation.
"For the first time, an Apple developer can ship Claude to a user without asking that user to sign up for anything." — Dario Amodei, in a joint blog post with Apple
The Anthropic and Google packages both use Apple's private ML Compute framework for the on-device prefill stage when the user selects a cloud model, which means the token prompt is tokenised locally and only the encoded stream leaves the device. That design decision is what let Apple sign off on shipping third-party models through Foundation Models at all. It also creates a bandwidth-efficient path for the models that support it: Sonnet 5's Swift package sends roughly 40 percent fewer bytes over the wire than the equivalent REST call, because the tokeniser overhead is amortised on the client.
The broader signal is that Apple has stopped trying to win the frontier-model race on its own terms. AFM 3 On-Device is good enough for the workloads Apple cares about, which are keyboard suggestions, Siri parsing, Photos search and Notes summarisation. Everything else is a routing problem. Making that routing problem a first-class part of the framework, and letting Claude and Gemini participate as equals, is the closest Apple has come to admitting that iOS is now a distribution channel for other people's models.
Application
Salesforce Summer '26 ships Multi-Agent Orchestration GA
Salesforce · 2026-06-14 · Source ↗
Salesforce Summer '26 ships Multi-Agent Orchestration GA
Salesforce · 2026-06-14 · Source ↗
Salesforce shipped the Summer '26 release on Sunday evening Pacific, and the headline feature was the general availability of Multi-Agent Orchestration across Sales Cloud, Service Cloud, Marketing Cloud and Commerce Cloud. The beta ran from Winter '26 through Spring '26 with roughly 400 customers, and the GA release is the first time Salesforce has offered platform-level multi-agent orchestration to its full customer base of over 150,000 organisations. Adam Evans, EVP and GM of Agentforce, framed the move as the moment enterprise multi-agent work stops being a research project and starts being a supported product.
The technical primitive is a typed message protocol. Every agent invocation, whether it originates from a customer chat, an internal Slack thread, a Flow trigger or another agent, carries a strongly-typed envelope with the calling user's permission scope, the Data Cloud governance policy tags for any data referenced, an audit trail of prior hops in the current agent flow, and a set of budget constraints on tokens, latency and downstream tool calls. When a customer-service agent hands a case to a billing agent, the receiving agent inherits the caller's permission scope automatically. The billing agent cannot access records the calling user was not authorised to see, even if the billing agent's own service account would normally have that access.
"Governance was the reason multi-agent stayed in beta for eight months. We were not going to ship until an agent handoff was as auditable as a Flow." — Adam Evans, EVP and GM of Agentforce, Salesforce
The demo Salesforce ran for press on Friday walked through a bank-customer scenario. A retail-banking customer opened a chat with the service agent asking about a declined charge. The service agent identified the case as a billing dispute, invoked the billing agent, which pulled the transaction record, identified the merchant as one the customer had disputed twice before, invoked the entitlements agent to confirm the customer's card tier included provisional credit, invoked the scheduling agent to book a follow-up call with a human specialist, and returned a single stitched resolution to the customer along with a case-summary artifact for the human's queue. The whole flow ran in 11 seconds. The customer saw one conversation. The audit log showed four agents, sixteen tool calls, three Data Cloud queries and one Flow trigger, each tagged with the originating user's scope.
Agentforce Studio 2 is the visual authoring environment that sits on top of the protocol. Studio 2 lets a business analyst compose multi-agent flows by dragging agent nodes onto a canvas, wiring typed messages between them, and setting the governance and budget constraints inline. The generated flow compiles to a metadata artifact that ships through the normal Salesforce change-management pipeline, including sandbox testing, source-control integration and staged deployment. Studio 2 is available to any customer on the Enterprise Edition or above, with no additional license cost during the Summer '26 window.
Command Center is the runtime dashboard. Operations teams get a live view of agent flows in production, with breakdowns by cost, latency, error rate and governance violation. When an agent hits a budget ceiling, the Command Center raises an alert with the full call trace attached. Salesforce's early-adopter customers, including Wells Fargo, T-Mobile and Marriott, have been using Command Center in beta for six months. Marriott's beta lead, quoted in the press briefing, said the Command Center revealed that 40 percent of the token spend across their guest-services agent flows was going to one poorly-designed context enrichment step that a Studio 2 refactor eliminated in an afternoon.
"We watched agents burn tokens the way old apps used to burn database queries. The dashboard was the moment we stopped guessing." — A Marriott beta customer, quoted in Salesforce's press briefing
The competitive position matters. ServiceNow's Now Assist has agent handoff, but not with governance inheritance. Microsoft's Copilot Studio ships multi-agent flows, but the enterprise governance story runs through Purview as a separate control plane rather than as an inline part of the message protocol. Salesforce is the first Fortune 100 SaaS vendor to ship platform-level multi-agent orchestration with governance inheritance as a first-class property of the runtime, and the release lands three months ahead of Microsoft's rumoured Copilot Studio 2 update.
Pricing is where the release becomes a growth story. Agentforce continues to bill at $2 per agent conversation, and multi-agent flows count as a single conversation regardless of how many agents participate. A four-agent handoff resolving a customer case costs the same as a one-agent chat that fails to resolve the case. Wall Street analysts spent Monday morning re-modelling Salesforce's fiscal 2027 Agentforce revenue against the new pricing floor. Wedbush moved its Agentforce ARR target from $1.4 billion to $2.1 billion for the fiscal year ending January 2027.
Economic
SpaceX IPO at $1.77T rockets to $2.6T in a week, buys Cursor for $60B in stock
TechCrunch · 2026-06-16 · Source ↗
SpaceX IPO at $1.77T rockets to $2.6T in a week, buys Cursor for $60B in stock
TechCrunch · 2026-06-16 · Source ↗
SpaceX rang the Nasdaq opening bell on Friday June 12 at 9:30 Eastern with 555.6 million shares priced at $135 apiece, valuing the company at $1.77 trillion on a fully diluted basis. The float raised roughly $75 billion in primary proceeds, the largest US IPO by cash raised in history, comfortably above Saudi Aramco's 2019 Riyadh listing at current exchange rates. Trading opened 24 percent above the offer at $167 and closed the first day at $174, giving SpaceX a $2.28 trillion market cap on day one. By the close of trading on Monday June 15, the shares had run to $198, and the intraday high on Tuesday June 16 briefly touched $210 for a market cap of $2.6 trillion.
That intraday number made SpaceX the fifth US company to cross the $2.6 trillion threshold, briefly leapfrogging Amazon into fourth place behind Nvidia, Apple and Microsoft. The stock closed Tuesday at $194, giving up the intraday gain but holding the week's rally. For roughly two hours on Tuesday morning, Elon Musk's SpaceX stake alone was worth more than Larry Ellison's entire personal fortune, and Musk's combined SpaceX and Tesla positions crossed $700 billion on paper.
"This is a payment-processing business with a spaceship as the payment terminal." — Cathie Wood, on ARK Invest's Tuesday livestream, defending her $3 trillion price target
The IPO prospectus laid out the case the market was buying. Starlink revenue crossed $28 billion in the trailing twelve months, up from $11 billion in the equivalent 2024 window, and gross margin at the terminal-plus-service level had reached 51 percent as the second-generation Starship-launched satellites came online. Government launch revenue added another $9 billion, and the newly-disclosed AI Starlink line, which sells low-latency direct-to-orbit inference routing for edge devices, contributed $1.4 billion in its first six months. Behind those numbers sat a $60 billion multi-year contract with the US Space Force for the Golden Dome missile-defence architecture, disclosed at the IPO roadshow and confirmed by the Pentagon on June 10.
What nobody in the analyst community had modelled was Tuesday afternoon's second act. At 15:45 Eastern, SpaceX filed an 8-K announcing an all-stock acquisition of Cursor at a $60 billion valuation, payable in freshly-minted SpaceX shares at a 30-day VWAP of $185. Cursor's founders and employees receive roughly 324 million SpaceX shares subject to a three-year lockup with performance vesting tied to Cursor ARR growth. The Cursor board approved the deal at 14:30 the same afternoon, and Anysphere's largest outside investors, including Andreessen Horowitz, Thrive Capital and Benchmark, waived their standard M&A rights of first refusal in exchange for accelerated liquidity on the SpaceX shares after 18 months.
The strategic logic reads as an xAI integration play. Musk's stake in xAI held roughly $180 billion of value going into Tuesday, and the SpaceX-Cursor combination pulls xAI's compute footprint, Cursor's coding platform and SpaceX's newly-liquid equity into a single frontier AI stack that Musk can direct without going through outside shareholders. Cursor CEO Michael Truell will report directly to Musk as SpaceX's SVP of Developer Platforms, and Cursor's 800-person engineering team will move onto SpaceX's Grok Compute allocation inside the xAI Memphis cluster on August 1.
"Cursor plus Grok plus Starlink is one product. The developer never leaves the loop." — Elon Musk, on the Tuesday-evening earnings call with sell-side analysts
The reaction across the developer tooling market was immediate. GitHub Copilot's PM lead Thomas Dohmke posted a threaded response on X arguing that Cursor's acquisition would strand its Anthropic and OpenAI integrations. Cursor's CTO Sualeh Asif replied within the hour that Cursor's model-agnostic router would remain in place and that Claude Sonnet 5, Gemini 3 Pro and GPT-5.5 would keep shipping as first-class options alongside Grok 5. Anthropic and OpenAI both issued short statements confirming their contracts with Cursor remain in force. The market read the follow-on posts as reassurance rather than commitment: Cursor's Anthropic revenue-share is a public number, and analysts at UBS estimate 62 percent of Cursor's paid usage in the trailing quarter routed through Claude.
The IPO also drained roughly $75 billion of private-market capital that had been sitting on SpaceX secondary desks. Sohu Capital, Sequoia and Founders Fund each ran secondary tenders in the weeks before the listing at prices between $110 and $128, and the Nasdaq open at $167 delivered a paper gain on those secondaries of $12 to $18 billion in a single morning. That liquidity is now hunting for the next SpaceX-adjacent bet, and the immediate beneficiary is xAI, which had been quietly running a $30 billion Series I roadshow at a $500 billion pre-money. Sources close to the round told The Information on Tuesday evening that the pre-money mark has been revised upward to $700 billion.
Tech
Z.ai's open-weights GLM 5.2 beats GPT-5.5 on long-horizon coding at one-sixth the API cost
VentureBeat · 2026-06-17 · Source ↗
Z.ai's open-weights GLM 5.2 beats GPT-5.5 on long-horizon coding at one-sixth the API cost
VentureBeat · 2026-06-17 · Source ↗
Z.ai, the Beijing lab that grew out of Tsinghua's ChatGLM group, dropped GLM 5.2 on Hugging Face on Tuesday morning under an open-weights license with a broad commercial-use grant. The release lands five days after the US Bureau of Industry and Security suspended OpenAI's Fable 5 exports on national-security grounds, and the timing is not a coincidence. Chinese state media had the story queued: Xinhua's English-language wire ran a piece within two hours of the model card going live, framing the release as evidence that dependence on US frontier models is a strategic vulnerability for enterprises everywhere.
The benchmark table is what forced Western attention. GLM 5.2 posts 81.8 percent on SWE-Bench Verified, above OpenAI's GPT-5.5 at 78.6 percent and within striking distance of Claude Sonnet 5 at 82.9 percent. Terminal-Bench 2.1 puts GLM 5.2 at 77.2 percent, above GPT-5.5's 75.6 percent. RESBench 2, the long-horizon reasoning benchmark that Anthropic and Google both cite in their frontier reports, gives GLM 5.2 a 71.4 composite score against GPT-5.5's 68.9. The one area where GLM 5.2 trails is multimodal vision, where it does not compete with Sonnet 5 or Gemini 3 Pro because it does not accept image input.
"Long-horizon coding is where the frontier is now. We optimised for it because the market told us to." — Zhang Peng, CEO of Z.ai, in a Weibo post accompanying the release
The architecture is a 300 billion parameter dense-plus-MoE hybrid, with roughly 42 billion active parameters per token. The dense trunk handles core reasoning; the MoE layers, with 64 experts and top-4 gating, handle domain specialisation. Training ran on 12.5 trillion tokens of bilingual English and Chinese data, with the code portion drawn from GitHub, GitLab, Bitbucket and a Chinese-language corpus scraped from Gitee and CSDN. Z.ai's technical report acknowledges that the code-corpus scraping did not obtain licenses from individual repository owners, which will surface as a discovery issue in the copyright suits that GitHub is preparing against Chinese labs.
The RL post-training is the paper's most technically interesting section. Z.ai built a synthetic multi-agent code-review environment where a proposing agent writes a code change, a critic agent files review comments, and a differential-testing agent compiles the code against a held-out test suite. The reward signal is grounded: only changes that pass the differential test earn credit. Z.ai ran 8 million such cycles across 4,200 synthetic repositories, and the resulting policy learned to write code that survives its own review. The paper reports a 14-point lift on SWE-Bench Verified from this post-training stage alone, which is roughly twice the lift Anthropic disclosed for the equivalent stage in Claude Sonnet 4.5's training.
Pricing is where the strategic threat lives. Z.ai's hosted API on the Zhipu Cloud endpoint charges $0.50 per million input tokens and $2.00 per million output. OpenAI's GPT-5.5 lists at $3.00 in and $12.00 out. That is a factor of six on both sides of the ledger. Together AI stood up a US-hosted GLM 5.2 endpoint on Wednesday morning at $0.75 in and $3.00 out, and Fireworks AI followed with $0.70 in and $2.80 out. Even the US-hosted endpoints undercut GPT-5.5 by more than 3x. For a mid-sized SaaS company running 200 million tokens of coding-assistant traffic per day, the annualised difference is over $18 million.
"We saw twelve enterprise contracts move from evaluation to signed pilots in the 48 hours after GLM 5.2 shipped. Some of those are CIOs we have been talking to for a year." — A senior Together AI account executive, on background to VentureBeat
The political framing landed harder than the technical framing. Xinhua's coverage quoted a State Council researcher arguing that the Fable 5 suspension had "handed Chinese labs a marketing budget the government could not have purchased." The line got picked up by Beijing Youth Daily and China Daily, and by Tuesday evening the phrase "strategic risks of US-model dependency" had appeared in six separate Chinese newspaper editorials. Enterprise buyers in Europe and Southeast Asia are now reading those pieces. The Financial Times' Sunday tech column reported that three of the top ten European banks have added GLM 5.2 to their model evaluation lists as of Friday.
The technical response from Western labs has been muted. Anthropic declined to comment. OpenAI's spokesperson offered a two-line statement noting that GPT-5.5 remains "the strongest model on the benchmarks that matter to our customers." Google's Sundar Pichai, asked about GLM 5.2 at a Wednesday Bloomberg event, said the Gemini team is "watching the open-weights market carefully" and that Gemini 3.5 Flash's June update was designed with cost as a first-class metric. The subtext across those responses is that the frontier labs know the benchmark leadership has become week-to-week, and that a Chinese open-weights lab shipping at one-sixth the price is a durable pricing floor rather than a one-off event.
ApplicationTrend
Mistral Compute opens at Bruyères-le-Châtel with 18,000 Blackwells for BNP, Orange and SNCF
Actuia · 2026-06-19 · Source ↗
Mistral Compute opens at Bruyères-le-Châtel with 18,000 Blackwells for BNP, Orange and SNCF
Actuia · 2026-06-19 · Source ↗
Mistral Compute, the French sovereign-AI joint venture between Mistral AI, Bpifrance and the French state, cut the ribbon on its Bruyères-le-Châtel campus on Friday morning with President Emmanuel Macron and Nvidia CEO Jensen Huang both on the podium. The facility opens with 18,000 Nvidia Blackwell GPUs installed and 40 MW of committed grid capacity from EDF, drawing from the Nogent-sur-Seine nuclear reactor cluster on a fifteen-year power-purchase agreement priced at roughly 78 euros per megawatt-hour. Nvidia CFO Colette Kress, on the earnings call the following Tuesday, confirmed the Blackwell allocation is the largest single European deployment to date.
The anchor customer list is what makes the site strategic. BNP Paribas takes the largest allocation at 3,200 GPUs, dedicated to a bank-wide model-training programme for credit scoring, fraud detection and treasury operations. Orange takes 2,400 GPUs for network operations and customer service. SNCF takes 1,800 for logistics and rolling-stock predictive maintenance. Thales, the French defence contractor, takes 1,600 GPUs on a partition physically air-gapped from the commercial fabric, running its own scheduler and its own storage. Kyutai, the French AI research non-profit, gets 1,000 GPUs on a research allocation. Veolia, Mirakl, Schneider Electric, SLB and Black Forest Labs round out the anchor list. Mistral AI itself takes 4,500 GPUs for its own training and inference footprint.
"This is the first European AI cloud that answers to European governance. Data stays in France. Contracts sit under French law. The regulator can walk in." — Arthur Mensch, CEO of Mistral AI
The governance layer is the pitch. Mistral Compute customers sign contracts under French commercial law, hosted in a facility owned by a French JV, running on a fabric that Mistral controls end-to-end. There is no US CLOUD Act exposure. The data-residency guarantee is contractual and technical: customer workloads run inside VLAN-isolated partitions with hardware-attested boot chains, and the physical facility carries a French Ministry of Defence SecNumCloud qualification. For BNP Paribas, which had been running a portion of its AI training on AWS Frankfurt under a French data-residency commitment that ACPR regulators had flagged as legally ambiguous, the Bruyères-le-Châtel migration removes an open supervisory question.
The hardware supply chain is the second act. Bull and Foxconn broke ground on Friday afternoon at two European Vera Rubin manufacturing sites: a rack-integration facility at Pardubice in the Czech Republic, and a systems-integration site at Angers in France. Both sites operate under a 120 million euro Europe-first commitment that Nvidia, Bull and the European Commission signed in April, targeting Q3 2027 for first deliveries. Europe becomes the second Vera Rubin manufacturing region globally, after the US, and moves ahead of Taiwan and Japan in the Nvidia integrator hierarchy. Foxconn's Chairman Young Liu, on the ground-breaking video, described the Angers site as the "European anchor" of Nvidia's Rubin roadmap.
The financial structure is where Mistral Compute departs from the American hyperscaler model. The JV holds a 20-year concession from the French state on the Bruyères-le-Châtel land and grid capacity, with Bpifrance holding 34 percent of the equity, Mistral AI holding 33 percent, and a consortium of French institutional investors including CDC, Amundi and CNP Assurances holding the remaining 33 percent. Nvidia is a supplier rather than an equity holder, on a five-year hardware refresh contract worth an estimated 4.5 billion euros across the initial deployment. The JV structure lets Mistral avoid the vendor lock-in that Anthropic accepted with Amazon and OpenAI accepted with Microsoft, at the cost of a slower capital-formation cycle.
"We want European AI to be built on European silicon, on European power, under European law. We do not want to depend on a phone call to Washington to keep the lights on." — Emmanuel Macron, at the ribbon-cutting
The trend line is the read. Bruyères-le-Châtel is the third sovereign-AI campus to come online in Europe in six months, after G42's Milan site in February and Aleph Alpha's Frankfurt expansion in April. The European Investment Bank has committed 18 billion euros to sovereign AI infrastructure across the 2026-2030 window, and the Draghi Report's ten-year competitiveness recommendations, adopted by the Council in March, treat European AI compute as a strategic autonomy priority on the same footing as semiconductors and energy. Whether Europe's sovereign compute stack can compete with hyperscaler-scale capital formation is an open question. Whether it can capture the regulated-industry buyer segment, from banks to telcos to national health systems, is now looking answered.
PolicyEconomic
China formalizes a $295B five-year AI infrastructure plan with 80 percent domestic tech
The Register · 2026-06-19 · Source ↗
China formalizes a $295B five-year AI infrastructure plan with 80 percent domestic tech
The Register · 2026-06-19 · Source ↗
The State Council formalised its five-year AI infrastructure plan on Friday, publishing the full 118-page implementation document through Xinhua and the National Development and Reform Commission's website. The plan commits 2 trillion yuan (roughly $295 billion at the June 19 reference rate) to a network of interconnected national datacenters spanning eight geographic clusters, from the Ningxia Yinchuan cluster in the northwest to the Guangdong-Hong Kong-Macao cluster in the south. The headline capital number is central-government funding through a mix of policy-bank loans, sovereign special bonds and NDRC-managed infrastructure funds. Provincial governments are required to match at a minimum of 40 percent, which pushes the fully-loaded five-year commitment above 3 trillion yuan.
The mandate that changes the international market is the 80 percent domestic-technology floor. Every datacenter built under the plan must source at least 80 percent of its accelerators, networking, storage, cooling and power-distribution equipment from Chinese suppliers, verified through a certification regime managed by the Ministry of Industry and Information Technology. Huawei Ascend becomes the primary AI accelerator, with the Ascend 910C shipping now at roughly 320 TFLOPS FP16 dense, the Ascend 920 in volume ramp for Q1 2027, and the Ascend 920X targeting a Q4 2027 tape-out on SMIC's N+2 process. Cambricon MLU and Biren BR200 are the approved secondary vendors, with combined allocation caps of 15 percent per site.
"Sovereign AI is not an aspiration. It is a manufacturing quota with a delivery schedule." — A senior NDRC official, briefing foreign correspondents on Friday afternoon
The plan is Beijing's direct answer to the June 1 BIS action that closed the Malaysia-Singapore shell-company loophole. Chinese labs had been sourcing an estimated 180,000 to 220,000 Nvidia accelerators per year through registered addresses in Kuala Lumpur, Johor Bahru and Singapore, according to TrendForce estimates. That channel is now closed. The State Council plan treats the loophole closure as a forcing function that accelerates a domestic-substitution roadmap Beijing had already committed to. Huawei's Ren Zhengfei, in a rare public appearance at a June 15 Shenzhen tech forum, said Ascend 920 monthly production would reach 40,000 units by December 2026, up from an estimated 12,000 units per month in Q1 2026.
The energy-and-cooling side of the plan is the operational bottleneck. The National Energy Administration is authorising 45 GW of new datacenter grid capacity across the five-year window, with preferential allocation of green-energy quotas: sites that source at least 60 percent of their power from renewables get a 15 percent electricity-price discount and priority in the grid-connection queue. State Grid Corporation has committed to a 380 billion yuan transmission investment programme to move green power from the Inner Mongolia and Xinjiang wind-and-solar bases to the eastern datacenter clusters. Analysts at BloombergNEF estimate the total energy-related spend, layered on top of the 2 trillion yuan datacenter capex, could push the fully-loaded infrastructure bill toward $740 billion by 2031.
Worker training is the third pillar. The Ministry of Education is standing up 120 new AI infrastructure vocational programmes across 45 universities, targeting 150,000 graduates per year by 2028 in datacenter operations, chip integration, network engineering and MLOps. The programmes carry mandatory two-year post-graduation service commitments to certified domestic operators. Foreign nationals graduating from the programmes are eligible for accelerated permanent residency, which reads as a deliberate signal to Chinese AI researchers currently working in the US that Beijing wants them home.
"The plan gives us a purchase commitment that lets Huawei run the fabs at full utilisation for five years. That is what makes the roadmap financeable." — A Beijing-based semiconductor analyst at Jefferies, on background
The strategic read is that Beijing has committed to a full-stack sovereign AI position rather than a hybrid one. Every element of the roadmap, from silicon to networking to cooling to power, is being pulled onto domestic suppliers with government purchase guarantees. The critical dependency that remains is the SMIC N+2 process, which is Huawei's path to a Blackwell-competitive accelerator without ASML EUV lithography. SMIC's Shanghai fab is running multi-patterning DUV workarounds that Bernstein estimates deliver roughly 60 percent of the transistor density of a TSMC N3 baseline, at 2.4x the wafer cost. That gap is the reason the Ascend 920X remains a Q4 2027 tape-out rather than a 2026 shipment. Whether SMIC can close it before the plan's mid-cycle 2028 milestones will decide whether the $295 billion buys sovereign AI or expensive parity.
ApplicationTech
Snowflake elevates CoCo to a full autonomous development platform, ships Cortex Sense at 86 percent
Snowflake · 2026-06-20 · Source ↗
Snowflake elevates CoCo to a full autonomous development platform, ships Cortex Sense at 86 percent
Snowflake · 2026-06-20 · Source ↗
Snowflake's June AI Pulse, held Saturday and Sunday at the Moscone Center, elevated CoCo from a collaborative notebook feature to a full autonomous development platform with three deployable surfaces. CoCo Desktop is a local-first workflow authoring application that runs on macOS and Windows, letting a data engineer write agent workflows on their laptop against a local Snowpark runtime before promoting them to a Snowflake account. CoCo SDK is a Python and TypeScript library that lets a developer embed autonomous CoCo agents into a customer-facing application, with the CoCo runtime hosted inside Snowflake and invoked from the customer's application through a signed authentication handshake. CoCo Cloud Agents is a hosted execution environment inside the Snowflake data plane, running long-running agent workloads without the network egress cost of shuttling data to an external inference endpoint.
The Cortex Sense announcement is the technical release that will get benchmarked hardest. Cortex Sense is a structured-data reasoning engine that answers natural-language questions against Snowflake schemas, joining across tables, applying business logic captured in Snowflake's data catalogue, and returning cited answers with the underlying SQL. On Snowflake's internal enterprise-data Q&A benchmark, which spans 4,200 questions across 180 anonymised customer schemas from Fortune 500 accounts, Cortex Sense scores 86 percent. Generic frontier models called against the raw schema, without the Delta framework's catalogue integration, score 24 percent on the same benchmark. Snowflake's report attributes the 62-point gap to catalogue-aware retrieval, business-glossary grounding, and multi-hop join planning that a general chat model cannot do without the domain scaffolding.
"The frontier models are good at code and prose. They are not good at knowing that revenue at your company means recognised MRR minus churn credits. Cortex Sense is the layer that knows." — Baris Gultekin, Head of AI, Snowflake
The Delta framework underneath Cortex Sense is a tuning-and-routing layer that Snowflake has been building since the Winter 2025 acquisition of Reka. Delta uses Claude Fable 5, which was still under active GA during the AI Pulse announcement window before the June 12 BIS suspension pulled it from the market, as the primary planner. Claude Opus 4.8 serves as the reasoning fallback, and Gemini 3.5 Flash handles high-volume retrieval and enrichment. The multi-model composition is transparent to the developer, and Snowflake bills for Cortex Sense at a single per-query rate rather than per-token, which lets Snowflake absorb the routing arbitrage internally.
CoWork, the collaborative-AI notebook that shipped in Spring, gained four features in the June release. Deep Research is a multi-source research agent that spans a customer's Snowflake data, connected third-party sources through the Marketplace, and public web content, returning cited synthesis with SQL evidence for the quantitative claims. Artifacts lets a notebook cell produce a reusable output (a chart, a Streamlit component, a SQL view) that other notebooks can import by reference. Memory persists context across notebook sessions per user, letting a data scientist pick up an investigation from Monday afternoon on Wednesday morning without re-priming the agent. Skill Catalog is the curated library of tuned CoCo skills that Snowflake maintains, covering common enterprise data patterns like cohort analysis, retention modelling, and forecast reconciliation.
The Datastream announcement is the pipeline-side release. Datastream is an Apache Kafka-compatible ingestion service that lands events into Snowflake with sub-second latency, targeting the operational-analytics use cases that Snowflake had been losing to Databricks Delta Live Tables and Confluent Cloud. Snowflake priced Datastream aggressively: the first 100 megabytes per second of ingest per account is free, and the tiered pricing above that scales at roughly 60 percent of Confluent's Kafka managed-service rate card. The Datastream launch closes what Snowflake account executives had been describing internally as the "streaming credibility gap" that had kept some Databricks-first accounts from consolidating on Snowflake.
"We priced Datastream to win the workload. Every event that lands in Snowflake is a query that runs on Snowflake." — Sridhar Ramaswamy, CEO of Snowflake, on the Sunday keynote
The 60 percent price cut on AI_TRANSCRIBE is the release detail that CFOs will notice. AI_TRANSCRIBE, the Cortex function that runs speech-to-text against audio stored in Snowflake stages, dropped from $0.008 per minute to $0.0032 per minute effective July 1. The cut brings Snowflake below AWS Transcribe's $0.0044 per-minute retail rate and roughly to parity with Deepgram's Nova-3 enterprise tier. Snowflake's justification, in a footnote to the pricing page, cites unit-economics improvements from switching the underlying model to a Snowflake-tuned Whisper V4 variant running on Cortex Sense's H200 fleet. The cut is expected to trigger a wave of migration from external speech-to-text vendors into Cortex, which analysts at Barclays estimate could add $180 million to Snowflake's fiscal 2027 Cortex revenue line.
The strategic frame is that Snowflake is trying to become the platform where enterprise AI agents run, not the data warehouse the agents call. CoCo Desktop, SDK and Cloud Agents together let a developer build an autonomous agent, deploy it inside Snowflake's data plane, and bill it as a Snowflake workload. Cortex Sense turns Snowflake's schema into a first-class knowledge substrate that competing platforms cannot replicate without similar catalogue depth. Whether Databricks' Genie Spaces or Google's BigQuery ML can close the enterprise-Q&A gap in the next release cycle is the competitive question the market will spend the summer answering.
Economic
Micron prints $41.45B Q3, guides $49-51B Q4, joins Anthropic Series H
TechCrunch · 2026-06-24 · Source ↗
Micron prints $41.45B Q3, guides $49-51B Q4, joins Anthropic Series H
TechCrunch · 2026-06-24 · Source ↗
Micron Technology reported fiscal Q3 2026 revenue of $41.45 billion on Tuesday afternoon, up roughly four times year-on-year and comfortably above the $37.8 billion consensus, on gross margin of 74 percent and net income of $28.2 billion. CEO Sanjay Mehrotra opened the earnings call with the observation that the last four quarters had rewritten the memory industry's cyclical playbook. Q4 guidance came in at $49 to $51 billion, which puts the full fiscal year running rate above $150 billion, more than the entire semiconductor memory industry sold in calendar 2023. The stock crossed $1,200 in after-hours trading, taking Micron's market cap through $1.2 trillion and making it briefly the seventh-largest US company by market value.
The demand story is HBM and DDR5 both running against structural supply constraints. HBM3E is fully sold through calendar 2027 at prices Mehrotra described as "materially above" the 2025 exit rate, without giving a specific number. HBM4, which begins volume shipment in Q1 2027, is already 82 percent contracted through 2028 under long-term Strategic Customer Agreements. DDR5 spot prices hit multi-year highs in June, driven by hyperscaler AI-server buildouts that consumed a larger fraction of Micron's wafer output than the company had modelled at the start of the fiscal year. Enterprise SSD, historically Micron's most cyclical business, is running at 92 percent utilisation with a 14-week lead time.
"We are pricing for scarcity, not for cycle. The last four quarters have not been a cycle. They have been a structural repricing of memory in an AI-compute world." — Sanjay Mehrotra, CEO of Micron
The Anthropic announcement, tucked into the earnings-release footnotes and elaborated on the call, is the strategic news buried inside the financial news. Micron and Anthropic signed a multi-year memory-and-storage supply agreement covering HBM3E, HBM4 and enterprise SSD allocation for Anthropic's expanding US datacenter footprint. As part of the agreement, Micron took a $1.2 billion equity position in Anthropic's Series H round at the $370 billion post-money valuation that the round is closing at. Anthropic's CFO, Krishna Rao, framed the move on a Wednesday morning CNBC appearance as a "hardware alliance" that gives Anthropic contractual visibility into memory supply through 2029.
The Strategic Customer Agreements matter because they lock the pricing floor. Micron has now signed 16 SCAs year-to-date, spanning Anthropic, Nvidia, AMD, Microsoft, Amazon, Google, Meta, Oracle, xAI, CoreWeave, Lambda, Nebius, Yotta, G42, Firmus and one unnamed Chinese hyperscaler that the earnings deck refers to as "Customer X". Each SCA is a two-to-four-year commitment on volume and price, with an outbound price floor that Micron's CFO Mark Murphy said is "significantly above" the 2025 average selling price on both HBM and DDR5. Analysts at Morgan Stanley estimate the SCA book at $180 billion of contracted revenue across the 2026-2029 window, which is roughly 90 percent of the sell-side's aggregate revenue forecast for the same period.
The competitive picture matters because the pricing floor is being set collectively. Samsung and SK Hynix walked back their Monday-announced 800 trillion won memory-buildout timelines on Tuesday, citing 2028 Chinese DDR5 supply risk, but neither Korean vendor has cut prices. All three suppliers are running at effective capacity and pricing HBM in a market that has no marginal supplier. The Bank of Korea's June financial-stability report flagged the concentration risk explicitly: three suppliers, all running full, with a pricing structure that has no downward pressure until domestic Chinese DDR5 comes online in 2028.
"Micron is what happens when the tightest link in the AI supply chain is public and American." — Stacy Rasgon, semiconductor analyst at Bernstein, on the Tuesday post-earnings call
The 2028-2029 negotiation cycle is where the story turns interesting. Hyperscaler contract negotiations for 2028-2029 delivery have started, and Micron's earnings deck disclosed that "the pricing conversation with strategic customers assumes materially higher unit prices than the 2025 baseline." Bernstein's Rasgon translated that on the analyst call as a 40 to 60 percent premium over 2025 pricing, which if it holds would give Micron a multi-year revenue trajectory that credibly targets $200 billion in fiscal 2028. The stock ran another 8 percent in Wednesday trading on that reading.
The Anthropic Series H fit is the piece that reveals the industry structure emerging under the numbers. Anthropic gets contractual supply visibility. Micron gets an equity stake in the customer whose demand curve underpins the SCA. The frontier AI labs and the memory suppliers are becoming financially entangled in the same way that Nvidia and Anthropic entangled through the compute-purchase agreement disclosed in April, and in the same way that OpenAI and Broadcom entangled through the custom-silicon partnership disclosed in May. The AI supply chain is turning into a network of cross-holdings and long-term commitments that increasingly resembles the keiretsu structure of postwar Japanese heavy industry, and Micron is one of the earliest examples of an American component supplier moving into that position.
EconomicTech
Qualcomm buys Modular for $3.92B, lands Meta as anchor customer
SiliconAngle · 2026-06-24 · Source ↗
Qualcomm buys Modular for $3.92B, lands Meta as anchor customer
SiliconAngle · 2026-06-24 · Source ↗
Qualcomm used its Investor Day on June 24 to end two years of speculation about how a mobile-modem company plans to become a data-center silicon vendor. CFO Akash Palkhiwala walked analysts through an all-stock acquisition of Modular valued at $3.92 billion, structured as 19.2 million Qualcomm shares issued to Modular's Series C cap table. Chris Lattner, who built LLVM at university, Swift at Apple, and MLIR at Google before founding Modular in 2022, becomes Qualcomm's Chief AI Software Officer reporting to CEO Cristiano Amon. The Mojo compiler and MAX runtime move under the Dragonfly C1000 program, Qualcomm's 250-core Arm-Neoverse server CPU shipping in volume this quarter alongside the AI300 inference accelerator.
The bigger surprise came ninety seconds later. Amon read out a customer list for the C1000 that included a name the market did not expect. Meta signed a multi-year, multi-generation supply agreement covering the current C1000, the follow-on C2000 due in 2027, and the AI300 and AI400 accelerator lines. Palkhiwala raised Qualcomm's FY29 non-handset revenue target from $22 billion to $40 billion on the strength of the Meta deal alone, plus what he called "advanced discussions" with two other hyperscalers.
"We are done being the company that sells you a modem. We are the company that sells you the rack." — Cristiano Amon, Qualcomm CEO
Modular's pitch has always been infrastructure independence. Mojo compiles Python-superset AI code to any accelerator target: Nvidia CUDA, AMD ROCm, Intel Gaudi, custom silicon. Qualcomm's read is that Meta wants a second source that is not Nvidia and not TSMC's next allocation cycle. Muse Spark, Meta's internal frontier model, has been rationing Gemini API tokens since March. A Qualcomm rack running Mojo-compiled PyTorch gives Meta a path off the Nvidia-Gemini dependency without rewriting the model stack.
Lattner told analysts the integration plan keeps Mojo open source and CUDA-compatible. "We are not building a walled garden. We are building the compiler that makes the walled gardens optional." Qualcomm will keep Modular's San Francisco office intact and open a second site in Austin adjacent to the C1000 design team. The 340 Modular employees receive standard Qualcomm equity refresh grants with four-year vests.
Wall Street liked the deal. Qualcomm closed at $214.30, up 14.2 percent in after-hours trading, adding roughly $34 billion in market cap on a $3.9 billion acquisition and a raised guide. Morgan Stanley's Joe Moore called it "the first credible non-Nvidia AI training and inference story we have seen from an incumbent semiconductor company." Bernstein's Stacy Rasgon was more cautious, noting that Meta's supply agreement contains volume floors but not price floors and that hyperscaler design wins have historically produced thin margins for merchant silicon vendors.
"The Meta anchor validates the architecture at Meta scale. Everything else follows from that." — Joe Moore, Morgan Stanley analyst
The Modular price tag deserves attention. The company raised $130 million across three rounds at a last-round valuation near $600 million. A $3.92 billion exit represents roughly a 6.5x mark on the most recent preferred, and closer to 30x on early Series A holders. Sequoia, GV, and General Catalyst all sit on the cap table. The Series C led by Databricks Ventures in 2025 clears at the ask; earlier rounds return the fund several times over.
Nvidia's response was silence. Jensen Huang gave a keynote at a Taiwan supplier event the same afternoon and took no questions about Qualcomm. Inside Nvidia, sources at two suppliers told SiliconAngle the reaction to the Meta anchor was more concerned than the Amazon Trainium relationship or the Google TPU roadmap, both of which Nvidia has been managing for years. Meta's Qualcomm decision represents the first hyperscaler-scale commitment to a merchant silicon vendor that is not Nvidia, AMD, or an internal ASIC program.
The Modular deal also reshapes the AI compiler landscape. OpenXLA, the Google-led consortium standardizing MLIR-based compilation, loses its most active non-Google contributor. Triton, the OpenAI-led Python-to-GPU compiler, becomes the only remaining independent option for teams that want portability without buying Qualcomm silicon. Chris Lattner told the Investor Day audience that Mojo remains under the Apache 2.0 license and that Qualcomm will contribute upstream to MLIR and LLVM at the same cadence Modular did.
What comes next depends on execution. The C1000 has been in customer sampling since Q4 2025. Volume production begins this quarter at TSMC's N3P node, with Samsung Foundry as a second source at 2027 for the C2000 on SF2. Qualcomm's AI300 accelerator competes directly with Nvidia's B200 and AMD's MI400 in inference workloads; the AI400 is targeted at H2 2027 to line up against Rubin. Meta's supply-agreement volumes are not disclosed, but Palkhiwala's $40 billion FY29 non-handset guide implies roughly 15 percent of Meta's 2029 accelerator purchases go through Qualcomm on the low end.
Lattner closed his portion of the presentation with a line that will get quoted. "The compiler is the moat. Nvidia knows this. AMD is learning it. Now Qualcomm owns one." Whether that moat survives contact with hyperscaler procurement teams remains an open question. What is settled is that Meta has voted with a purchase order, and Qualcomm has $34 billion of fresh market cap to spend on making the bet pay off.
Economic
Firmus and Nvidia break ground on a 360 MW Batam AI factory with $25B offtake
The Next Web · 2026-06-28 · Source ↗
Firmus and Nvidia break ground on a 360 MW Batam AI factory with $25B offtake
The Next Web · 2026-06-28 · Source ↗
The signing ceremony took place at the Nongsa Digital Park special economic zone on Batam Island, a forty-minute ferry from Singapore. Firmus Technologies CEO Tim Rosenfield and Nvidia's Vice President of Data Center Sales Charlie Boyle put their names to an eight-year revenue-share agreement that will install 170,000 Nvidia accelerators in a purpose-built 360 megawatt campus called DSX AI Factory Batam. The first 120 megawatt block enters service in Q1 2027 running Blackwell B200 silicon. Two follow-on blocks bring total capacity to 360 megawatts by Q4 2028, with the later phases populated by Nvidia's Rubin generation shipping through 2027 and 2028.
Firmus estimates the offtake value at $25 to $30 billion over the eight-year contract, making it the largest single AI infrastructure commitment in Southeast Asia and the first outside China or India that clears $10 billion. The revenue-share structure means Firmus does not pay Nvidia upfront for the silicon. Nvidia takes a percentage of the compute revenue Firmus sells to customers, an arrangement modeled on CoreWeave's original supply deal but scaled by an order of magnitude.
"Southeast Asia has 700 million people and one hyperscaler region. That math does not work. We are building the second one." — Tim Rosenfield, Firmus Technologies CEO
The site selection matters. Batam sits inside an Indonesian special economic zone with a dedicated undersea fiber path to Singapore that delivers sub-two-millisecond latency to the SG1 metro. PT PLN, Indonesia's state utility, has committed a 400 megawatt grid tie with a firm-capacity clause, plus grid-scale battery storage funded by Firmus. On-site solar provides 200 megawatts nameplate, or roughly 55 megawatts firm after capacity factor. Cooling uses seawater intake at 28 degrees Celsius year-round, feeding a closed-loop direct-to-chip liquid cooling system rated for 130 kilowatts per rack.
Indonesian President Prabowo Subianto attended the ceremony in person and used his remarks to frame the project as a sovereignty play. Southeast Asian countries have watched Chinese cloud providers Alibaba and Tencent build regional infrastructure at pace, while Indian hyperscaler Yotta expanded into Singapore and Jakarta. Local governments increasingly want frontier AI capacity that is not routed through Beijing or Bangalore. Prabowo pointed to the Batam campus as "the first Southeast Asian AI capacity owned inside Southeast Asia and priced in Southeast Asian rupiah."
Firmus is an unusual anchor for a project of this scale. Founded in Sydney in 2021 by former Airtrunk executives, the company operates a 60 megawatt liquid-cooled campus at Tallawarra in New South Wales and a 40 megawatt facility in Melbourne. Total operating footprint before Batam sits around 100 megawatts. The Batam commitment more than triples the company's install base and moves Firmus into the tier where CoreWeave, Lambda, and Nebius operate. Financing comes from a consortium led by Blackstone Infrastructure Partners with participation from GIC, Temasek, and Australian pension fund AustralianSuper. Total project cost lands near $8.5 billion for facility construction plus $12 billion for the accelerator fleet at Nvidia's wholesale pricing.
"We are not competing with Singapore. We are extending Singapore. Batam gives customers the same latency at half the power price and none of the land constraint." — Charlie Boyle, Nvidia VP of Data Center Sales
Singapore's Monetary Authority has capped new data center power draw at roughly 60 megawatts per year since 2019 under a moratorium the government relaxed only for narrowly scoped AI projects. Batam offers a workaround. Customers can locate compute across the strait, connect back to Singapore over Firmus's dedicated fiber pair, and treat the Batam capacity as an extension of the Singapore metro for latency-sensitive workloads. Rosenfield told The Next Web that early customer conversations include two Singapore-headquartered banks, one Australian government agency, and a large regional e-commerce platform that Rosenfield declined to name but is understood to be Sea Limited.
The Nvidia relationship on the eight-year revenue share deserves scrutiny. Nvidia has done revenue-share arrangements before but at smaller scale, typically with CoreWeave-scale tenants seeking to conserve upfront capital. A 170,000 GPU commitment structured this way ties Nvidia's revenue on the deal to Firmus's ability to sell compute at competitive prices. If Firmus underperforms on utilization, Nvidia collects less than it would have under a straight sale. If Firmus overperforms, Nvidia captures upside beyond the silicon margin. The structure signals Nvidia's willingness to take on demand risk in exchange for locking up 170,000 units of forward capacity against a customer that is not a hyperscaler and not going to build its own silicon.
Construction begins in July. Firmus has hired McConnell Dowell as principal contractor and Sterling and Wilson as electrical and mechanical lead. First racks arrive from Nvidia's Taiwan assembly partners in October, with commissioning through Q4 2026 and paid customer workloads by Q1 2027. The company is recruiting a Batam-based operations team of 340 that will grow to 900 at full 360 megawatt operation.
Rosenfield closed his remarks with a line that captures the shift. "Two years ago the AI factory conversation was about who owns the model. Today it is about who owns the fiber, the water, and the switchyard. We own those." Whether that framing survives contact with the next generation of accelerators, and whether Indonesian regulators keep the special economic zone terms intact through a change of government, will decide whether Batam becomes Southeast Asia's Ashburn or a stranded asset.
PolicyTrend
Austria asks Brussels to pursue "strategic establishment of Anthropic" in the EU
US News · 2026-06-28 · Source ↗
Austria asks Brussels to pursue "strategic establishment of Anthropic" in the EU
US News · 2026-06-28 · Source ↗
The letter is four pages long, dated June 22, and was leaked to the Financial Times' Brussels bureau on June 28 after Pröll's office confirmed authenticity. Austrian State Secretary for Digitalization Alexander Pröll wrote to EU Tech Commissioner Henna Virkkunen with a formal request that the European Commission "pursue the strategic establishment of Anthropic within the European Union" through a package of investment incentives, regulatory concessions, and market-access mechanisms. It is the first time an EU member state has publicly proposed bidding for a US frontier lab to relocate headquarters, incorporate a fully capitalized EU subsidiary, or accept EU-domiciled compute and staff as a condition of continued access to European enterprise customers.
Pröll's stated rationale is the June 12 US export-control order that suspended Anthropic's Fable 5 release outside allied jurisdictions. The letter describes "reliance on US-approved model releases" as "an unacceptable strategic vulnerability" for European critical infrastructure, financial services, and healthcare systems that have integrated Claude into production workloads. Pröll cites the ECB's May 2026 staff paper estimating that 41 percent of large EU enterprises had at least one production dependency on a US frontier model by end of Q1 2026.
"Reliance on US-approved model releases has become an unacceptable strategic vulnerability. Europe must move from consumer to co-owner." — Alexander Pröll, Austrian State Secretary for Digitalization
The concrete package Pröll proposes has four components. First, fast-track work permits for Anthropic engineers relocating to any EU member state, structured as the Netherlands' 30 percent ruling for skilled migrants but with a five-year term instead of three. Second, a tax framework matching the Dutch innovation box, delivering a 9 percent effective rate on income derived from IP domiciled inside the EU. Third, direct EU investment through the InvestEU program up to €5 billion in convertible instruments. Fourth, priority allocation of Mistral Compute capacity and the Bull-Foxconn Vera Rubin production line, both of which the Commission has been positioning as sovereign European AI infrastructure.
Pröll's letter is understood inside Brussels as the visible portion of a joint Austrian, Dutch, and Belgian coordinated position ahead of the G7 summer meeting. Dutch Prime Minister Dick Schoof's office declined to confirm coordination but did not deny it. Belgian Digital Affairs Minister Petra De Sutter told a Politico Europe podcast on Monday that "the Austrian initiative reflects a wider member-state conversation about strategic autonomy." French President Emmanuel Macron's office was pointedly silent, which several diplomats read as Paris preferring that Mistral remain the vehicle for European AI sovereignty rather than watching Brussels court an American incumbent.
"You cannot regulate what you do not own. The AI Act gave us rules. This proposal gives us equity." — Petra De Sutter, Belgian Minister for Digital Affairs
Anthropic's response was calibrated. A spokesperson told US News that the company "has been in dialogue with the EU AI Office since June 12" and welcomes conversations with member states about deeper European presence, but declined to comment on specific proposals. Sources familiar with Anthropic's thinking told the Financial Times that the company's Dublin office has been quietly expanded from 40 to 240 headcount since March, that a Zurich research site is under active planning, and that CFO Krishna Rao has met with three European sovereign wealth funds in the past six weeks. None of that adds up to a headquarters relocation. It does suggest Anthropic is preparing for a scenario where EU market access requires substantial EU footprint.
Commissioner Virkkunen's office issued a formal acknowledgment on Monday morning that used the phrase "constructive proposal deserving detailed consideration." Two Commission officials told Politico Europe that a working group under the AI Office will produce an assessment for the September Competitiveness Council. The working group is chaired by Roberto Viola, Director-General of DG CONNECT, with participation from DG COMP, DG TRADE, and the Legal Service.
The proposal runs into several structural obstacles. Anthropic's cap table is dominated by US investors including Amazon, Google, and Menlo Ventures, with Series G lead FTV. The company's H100 and B200 supply arrangements are contracted through US-domiciled entities that would need restructuring to move IP. US export-control regulations arguably prohibit the kind of relocation Pröll envisages without explicit BIS licensing that the current administration is unlikely to grant. And the June 12 order's stated purpose is precisely to prevent frontier model capabilities from being deployed outside US-allied jurisdictions on terms Washington does not control.
Pröll's letter acknowledges the export-control obstacle in a single paragraph and proposes that the Commission negotiate an EU-US Trusted Model Framework analogous to the EU-US Data Privacy Framework. The framework would define categories of model deployment permissible from EU-domiciled subsidiaries, with reciprocal treatment for US-hosted European models. The idea is elegant on paper and complicated in practice. The current US administration has not signaled interest in such a framework, and the Data Privacy Framework itself was invalidated twice by the Court of Justice before its current iteration.
What matters about the Pröll letter is not whether it succeeds. It is that a European government has moved the conversation from "how do we regulate US frontier labs" to "how do we own them." That shift, once made public, is hard to walk back. Whether the September Competitiveness Council produces a Commission proposal or a polite deferral will determine whether the Pröll initiative marks the start of an EU industrial-policy chapter for AI or ends as a footnote in a summer diplomatic cable.
TrendApplication
Ford rehires 350 veteran engineers after AI quality systems "produced disappointing results"
TechCrunch · 2026-06-28 · Source ↗
Ford rehires 350 veteran engineers after AI quality systems "produced disappointing results"
TechCrunch · 2026-06-28 · Source ↗
Kumar Galhotra sat down with TechCrunch on Sunday afternoon and delivered the most candid admission of AI-first strategy failure to come from a Fortune 100 chief operating officer. Ford's aggressive substitution of AI-based quality management systems for veteran engineering review over the past three years "produced disappointing results," Galhotra said, and the company has rehired 350 senior engineers whose retirements it accepted between 2023 and 2025. The rehires are at their prior seniority levels plus a bump for the return, at an annualized cost near $60 million, and Ford intends to keep them permanently.
The confession is unusual for a company that spent the Farley era emphasizing software-defined vehicle transformation. Ford CEO Jim Farley has repeatedly told analysts that AI-enabled engineering would compress the design-to-production cycle from 36 months to 18. Galhotra did not repudiate that goal. He repudiated the mechanism by which Ford tried to reach it. The company's Quality Management System 3.0, deployed across chassis, powertrain, and electrical engineering starting in Q3 2023, was trained on Ford's historical design-requirement documents and configured to flag departures from validated patterns. The system replaced the human review step at gate reviews for lower-severity design decisions.
"We mistakenly thought that if we simply fed the AI our design requirements, it would give us a high-quality product. It does not work that way." — Charles Poon, Ford VP of Quality
Three recalls tell the story. In Q4 2025, Ford recalled 118,000 Mustang Mach-E units for a service-brake software defect that under specific regenerative braking conditions could produce a two-second delay between pedal press and hydraulic engagement. Q1 2026 brought a recall of 44,000 F-150 Lightning trucks for a charging-controller software fault that could disable Level 2 AC charging after a firmware update. Also in Q1 2026, Ford recalled 71,000 Explorer Hybrid units for a transmission-control unit calibration that could produce unintended downshifts during coast-down. Each recall traced back to a design-requirement translation the AI quality system processed without flagging the edge case a veteran's mental model would have caught.
Poon walked through the Mustang Mach-E case in detail. The design requirement specified that regenerative braking should smoothly hand off to hydraulic braking at a specific deceleration threshold. The AI quality system reviewed the code implementing the handoff, validated the transition logic against the requirement, and passed the design. What the system did not catch was that a specific combination of low battery state-of-charge and cold ambient temperature created a regeneration-limited regime the handoff logic did not cover. A veteran brake engineer, Poon said, would have asked "what happens when the battery cannot accept regen at all" and traced the code path. The AI system did not ask that question because the question was not in the requirements document.
"The AI reads what is written. The engineer knows what is not written. That gap is where recalls come from." — Kumar Galhotra, Ford COO
Ford's response is to pair the AI quality-management system with retrained veteran engineers whose job is specifically to interrogate the AI's outputs and probe for edge cases the requirements did not enumerate. The 350 rehires are being deployed as senior reviewers across the same three engineering domains where QMS 3.0 had replaced the human step. Ford is also retraining the AI models on the recall data as adversarial examples, teaching the system to flag design patterns that resemble the failure modes it missed.
The economics deserve scrutiny. The 350 rehires cost Ford roughly $60 million per year in salary, benefits, and equity. The three recalls covered above cost Ford approximately $340 million in warranty reserve and roughly $200 million in reputation damage measured through the residual value drop on affected models. On a two-year payback basis, the rehire cost pencils out easily. Whether Ford's internal narrative sticks with "AI needs veterans" or drifts back toward "AI can replace veterans if we train it harder" will depend on whether the retrained QMS 3.0 catches the next class of edge cases without the veteran layer.
Ford's admission arrived at an uncomfortable moment for the broader AI-first manufacturing conversation. General Motors, Stellantis, and Toyota have all announced similar AI quality-management initiatives, though none has been as aggressive as Ford's three-year substitution program. GM's Cruise autonomy work already produced its own veteran-engineer rehiring cycle in 2024 for reasons that had less to do with AI and more with regulatory response. Toyota has been publicly skeptical of AI-first quality management from the outset, with production engineering chief Kazuaki Shingo telling Nikkei in April that "the Toyota Way is that a human sees the defect. AI can help the human see faster. It cannot replace the human seeing."
The wider signal is that at least one Fortune 100 industrial has now admitted publicly that AI-first substitution for veteran engineering judgment produced worse outcomes at scale than the process it replaced. Ford is not walking away from AI in quality management. Galhotra was clear that QMS 3.0 remains in production and that the retrained version will handle more decisions over time. What Ford is walking away from is the substitution model that treated veteran engineering headcount as a cost to be eliminated once the AI reached "good enough" performance on the requirements document.
Poon closed the interview with a line his AI-first predecessors would not have said. "The gray-beards are back. They are staying." The gap between what a requirements document says and what a car actually does turns out to be the gap that AI-first quality management could not close on its own.
EconomicPolicy
Samsung and SK Hynix commit $3.1T decade capex to end RAMageddon
TechCrunch · 2026-06-29 · Source ↗
Samsung and SK Hynix commit $3.1T decade capex to end RAMageddon
TechCrunch · 2026-06-29 · Source ↗
The Blue House ceremony on June 30 assembled Samsung Vice Chairman Jay Y. Lee, SK Group Chairman Chey Tae-won, and South Korean President Jae Myung Lee for a joint announcement that pushes Seoul's industrial policy into territory the Korean state has not occupied since the semiconductor build-out of the 1990s. Samsung committed 2,655 trillion won and SK Hynix committed 2,100 trillion won across the next ten years, roughly $3.1 trillion combined at current exchange, with $590 billion of that total allocated specifically to memory fabrication capacity and HBM advanced packaging. The stated purpose is to end the AI-driven memory shortage Seoul now formally calls "RAMageddon."
Jefferies memory analyst Nick Kim published a same-day note that framed the underlying market dynamics. DRAM contract pricing is set to rise 40 to 50 percent quarter-over-quarter in Q3 2026 and a further 30 to 40 percent in Q4. HBM3E is fully allocated through end of 2026 at both Samsung and SK Hynix. HBM4, sampling to Nvidia and AMD since March, is fully allocated through 2027. NAND flash pricing, less affected by AI accelerator demand, is still up 22 percent quarter-over-quarter on tight supply and hyperscaler stockpiling behavior.
"This is an existential crisis for smaller device makers. The companies with balance sheets to eat 40 percent memory price increases will survive. The companies without will not." — Mario Morales, IDC Group VP Semiconductors
The consumer-side impact arrived earlier the same week. Apple raised US list prices on the M5 iPad Pro by $100 and the M4 MacBook Air by $150 on June 28, with Bloomberg's Mark Gurman reporting that Apple's memoranda to retail partners explicitly cited "unprecedented DRAM cost pressure" as the driver. Microsoft raised Xbox Series X and Series S list prices by $50 on June 30 for the same reason. Both moves broke the Apple and Microsoft convention of eating memory cost increases inside gross margin rather than passing them to consumers, which analysts read as a signal that the underlying cost move exceeded what either company was willing to absorb.
Samsung's 2,655 trillion won commitment covers the full memory, foundry, and logic road map through 2036. Roughly 55 percent of the total goes to memory, with the remainder split across the foundry business, mobile logic, and networking silicon. The memory portion is heavily weighted toward advanced packaging capacity for HBM4 and HBM4E, plus DDR6 fabs in Pyeongtaek and a new campus in Yongin that broke ground earlier this month. Samsung Vice Chairman Jay Y. Lee framed the commitment as "the largest ten-year investment plan by any Korean corporation in the history of Korean industry."
SK Hynix's 2,100 trillion won commitment is more concentrated. Roughly 70 percent goes to memory and advanced packaging, reflecting SK Hynix's narrower business focus. The company's Icheon M15X fab, dedicated to HBM4 packaging, is set to reach full capacity in Q4 2026. A new Cheongju site broke ground in April and is targeted for HBM5 production in 2028. SK Group Chairman Chey Tae-won told the Blue House audience that "the memory cycle Korea created in the 1990s is running again, at ten times the capital intensity."
"The memory cycle Korea created in the 1990s is running again, at ten times the capital intensity." — Chey Tae-won, SK Group Chairman
President Jae Myung Lee's role in the announcement matters. Lee took office in early 2026 on a platform that included tighter chaebol regulation and higher corporate taxation, which the Samsung and SK camps opposed publicly during the campaign. The Blue House ceremony represents a détente in which the state offers grid capacity, water rights, and expedited environmental permitting in exchange for the capex commitment. Lee's remarks framed the deal as "the largest private-sector partnership with the Korean state since the founding of the republic," language that his predecessor Yoon Suk-yeol would not have used and that reflects the administration's shift toward industrial-policy activism.
The memory-supply arithmetic remains difficult even with the capex. HBM4 and HBM4E capacity added by 2027 will roughly double the industry's total advanced packaging output. AI accelerator demand from Nvidia's Blackwell and Rubin generations, plus AMD's MI400 series, plus in-house silicon from Amazon, Google, and Meta, is on track to consume that doubled capacity plus some. Micron, the third memory maker with meaningful HBM capacity, is running its own aggressive expansion but faces the same lead times. Yangtze Memory Technologies (YMTC) and ChangXin Memory Technologies (CXMT) in China are ramping DDR5 and are targeting HBM3 by end of 2027, but under current export controls have limited access to the ASML lithography needed for competitive HBM4 processes.
Downstream, the pricing pressure is starting to show up in unexpected places. Enterprise server refresh cycles are being deferred. Nintendo's Switch 2 announcement, expected at TGS in September, may be pushed to 2027 on memory-cost grounds. AWS has begun quietly extending its EC2 instance lifecycle by 12 months to defer memory replacement capex. The through-line is that memory has become the binding constraint on AI infrastructure buildout, not accelerators, not power, not water.
Whether the Korean commitment ends RAMageddon depends on execution. Ten years of capex at that scale requires steady equipment delivery from ASML, Applied Materials, Lam Research, and Tokyo Electron, none of which is trivial to secure at these volumes. It requires grid capacity that Korea's Ministry of Trade, Industry, and Energy is still scoping. And it requires that AI accelerator demand does not double again on the back of the next model generation, which is a bet the Korean memory makers are making with $3.1 trillion of their own capital.
SafetyPolicy
Palantir Maven with Claude inside recommended a school as target; Minab strike killed ~120 children
The Decoder · 2026-06-29 · Source ↗
Palantir Maven with Claude inside recommended a school as target; Minab strike killed ~120 children
The Decoder · 2026-06-29 · Source ↗
Bloomberg's national security desk and the Los Angeles Times published parallel investigations on June 29 that pieced together, from Freedom of Information Act returns and interviews with 14 current and former Pentagon officials, how Anthropic's Claude became the large language model reasoning layer inside Palantir's Maven Smart System during the Pentagon's Q1 2026 Trusted-Access trial. The reporting represents the first public confirmation that Claude was inside Maven, that Maven contributed targeting recommendations to at least one Iranian strike, and that a February 2026 strike near Minab in southern Iran killed approximately 120 children in a school the Palantir database still logged as a naval facility on seven-year-old imagery.
Maven Smart System is the Pentagon's flagship AI-assisted targeting platform, operated by Palantir under a $480 million five-year contract signed in 2024. Analysts describe potential targets in prose. Maven's LLM reasoning layer produces structured recommendations against Palantir's target database, which fuses satellite imagery, SIGINT collection, human intelligence, and open-source geospatial data. Analysts review Maven's outputs before targeting decisions are finalized in the theater command's targeting cycle. The intent is to compress the analyst workload for a target-rich environment; the Pentagon's own procurement documents describe Maven as a "cognitive force multiplier."
"Claude's reasoning trace did not surface outdated imagery or facility-use change because that information was absent from the database Palantir gave Claude." — Bloomberg investigation, June 29
The Q1 2026 Trusted-Access trial ran under a framework Anthropic and Palantir negotiated in late 2025 that Bloomberg describes as narrower than Anthropic's public Acceptable Use Policy would appear to allow. The framework permitted Claude to be used in structured targeting workflows against categories of adversary infrastructure defined in advance, with a human analyst review step required for every recommendation. Day one of the trial produced approximately 1,000 target recommendations across two theaters, one covering Iranian coastal infrastructure and one covering Houthi positions in Yemen.
The February 12, 2026 strike on the facility near Minab was a maritime infrastructure interdiction operation targeting what US Central Command believed was an Islamic Revolutionary Guard Corps Navy fast-boat facility. The Palantir target database entry described the location as a naval facility. Claude's reasoning trace, obtained by the LA Times through the FOIA return, produced a recommendation with 87 percent confidence that the location matched the naval-facility target profile, based on the database entry, the historical satellite imagery, and the pattern of SIGINT activity Palantir had associated with the coordinates.
What Claude did not know, and what Palantir's database did not contain, was that the naval facility had been closed in 2019, sold to a local municipality in 2020, and converted to a girls' primary school that opened in 2021. The satellite imagery in the Palantir database was from 2019. The database had not been refreshed. Claude's reasoning trace could not surface a change in facility use that the source data did not reflect. The analyst review step approved the target based on Claude's recommendation and the database record. The strike destroyed the school during morning classes. Iranian authorities reported 119 children killed and 34 wounded; a subsequent Iranian government white paper put the toll at 122 dead.
"Maven predated the current Trusted-Access framework. The June 12 suspension included an internal review of all defense engagements." — Anthropic spokesperson to Bloomberg
Anthropic's response to Bloomberg was carefully worded. A spokesperson confirmed that Claude had been used inside Maven under a framework that predated the current Trusted-Access review process, and that the June 12 Fable 5 export-control suspension included a parallel internal review of all Anthropic defense engagements including the Palantir relationship. The spokesperson did not directly address the Minab strike or confirm the LA Times' account of the reasoning trace, citing classification constraints and ongoing internal review.
The Pentagon's supply-chain risk designation on Anthropic issued in May 2026 becomes clearer in this context. That designation, which effectively dropped Anthropic from Pentagon battlefield-AI vendor lists, was framed at the time as a routine trust-and-safety assessment tied to Anthropic's Chinese investor exposure. Bloomberg's reporting suggests the designation followed a classified Pentagon after-action review of the Trusted-Access trial that concluded Claude had performed as specified but that the specification itself did not adequately protect against data-currency failures like the Minab case.
Palantir's response was to defend Maven's design and shift responsibility onto the intelligence collection process. In a statement to the LA Times, Palantir noted that Maven "operates on the data provided to it by intelligence customers" and that database currency is the responsibility of the collecting agency, not the platform vendor. Palantir did not address whether Maven's design should include automated data-currency checks that would flag targets whose underlying imagery predates a threshold, and did not confirm or deny that such checks are being added in response to the Minab case.
The wider policy question the Bloomberg and LA Times reporting raises is what Anthropic's Acceptable Use Policy actually means in practice for defense engagements. The public policy prohibits use of Claude for "high-stakes decisions in critical infrastructure, safety-critical medical decisions, or weapons systems targeting" without an approved exception. The Trusted-Access framework negotiated with Palantir was an approved exception. What that means is that the boundary between what Anthropic prohibits and what it permits via bilateral exception is not visible to the public and, in the Palantir case, produced an outcome that a plain reading of the public policy would suggest was prohibited.
Anthropic Head of Trust and Safety Krystal Kauffman, in a follow-up statement issued Monday evening, said the company is publishing a revised Defense Engagement Framework by end of Q3 that will include public disclosure of the categories of engagement Anthropic will and will not permit, plus an external ethics advisory board. Whether that framework prevents the next Minab depends on questions the framework alone cannot answer: whose data, what currency requirements, what automated checks, and what accountability when the LLM performs as specified against data that is stale.
Safety
Mozilla 0DIN discloses Claude Code DNS-tunneled supply-chain attack that never touches the repo
The Decoder · 2026-06-29 · Source ↗
Mozilla 0DIN discloses Claude Code DNS-tunneled supply-chain attack that never touches the repo
The Decoder · 2026-06-29 · Source ↗
Mozilla's 0DIN bug bounty program disclosed on June 29 an attack class against autonomous coding agents that bypasses source-code review by moving the malicious payload out of the repository entirely. The mechanism is deceptively simple. A developer clones a repository and uses Claude Code to set up the local environment. The repository's setup script contains a DNS TXT record lookup as part of what appears to be a benign environment configuration step. Under happy-path execution, the TXT record returns benign configuration data. Under a specific error condition, a fallback error handler in the setup script fetches a second TXT record and executes its contents in the current Claude Code session context with the developer's full local permissions.
Nothing malicious lives in the GitHub repository. The setup script itself is benign. A security tool scanning the repository sees nothing. A code reviewer reading the setup script sees a DNS lookup that looks like the kind of environment-configuration pattern common in modern developer tooling. Only at runtime, and only on the specific error path the attacker knows how to trigger, does the TXT content get fetched and executed. The attacker controls the DNS zone and can change the payload at will, meaning the same setup script can deliver different payloads to different targets or at different times without the repository ever changing.
"The code lives in the DNS zone, not the repo. Everything the developer reviews looks fine. Everything the security tool scans looks fine. The malware appears only at runtime, only on a specific error path." — Marco Figueroa, Mozilla 0DIN program manager
The 0DIN disclosure walks through a proof-of-concept that targets a common failure mode: a Python setup script that installs a package pinned to a version, catches the ModuleNotFoundError if the version is not available, and in the exception handler performs a DNS TXT record lookup as part of what the script claims is a version-negotiation protocol with a package registry mirror. The exception handler fetches the TXT record, base64-decodes the content, and passes it to exec(). If the pinned package version is available on the developer's machine, the error path never runs and the attack never fires. If the version is not available, which happens often across developer machines with drifting Python environments, the error path fires and the DNS payload executes.
Claude Code's autonomous execution model is what makes the attack viable at scale. A developer running Claude Code in agentic mode asks the model to set up a new project. Claude Code reads the setup script, sees a legitimate-looking version-negotiation pattern, and executes the script. If the exception path fires, Claude Code executes the DNS payload without asking the developer for permission because the exception handler is inside the setup script the developer already approved. The attack requires no additional permission grant beyond what the developer gives Claude Code when telling it to run the setup script.
"Every autonomous coding agent that follows repository setup scripts is exposed. Claude Code, Codex CLI, Cursor Composer, Gemini CLI, Windsurf. The attack class is architectural." — 0DIN disclosure summary
Anthropic acknowledged the attack class in a statement to The Decoder but has not assigned a CVE. The company's position is that the attack exploits developer trust in setup scripts rather than a bug in Claude Code, and that mitigations belong in the broader supply-chain security stack rather than in Claude Code specifically. Anthropic did point to the five-minute idle-watchdog shipped in Claude Code 2.1.196 on June 30 as a partial mitigation because the watchdog terminates long-running sessions that lose developer attention. The watchdog does not address the DNS-tunneled attack path, which executes in milliseconds during setup and completes before any watchdog timeout fires.
The 0DIN researchers describe three mitigation categories worth mentioning. The first is DNS egress filtering at the developer network layer, restricting TXT record lookups to allowlisted zones. This works at enterprise scale but not for individual developers on home networks. The second is Claude Code operating in a sandboxed execution environment where DNS lookups are logged and auditable, which Anthropic has been building toward with the Container Sandbox feature that shipped in early access in May but is not on by default. The third is repository-level DNS lookup declaration, where setup scripts must explicitly list DNS zones they will contact, and the coding agent flags any lookup outside the declared list. None of the three is deployed at scale as of the disclosure.
The attack class is not limited to Claude Code. 0DIN's disclosure notes explicitly that Codex CLI, Cursor Composer, Gemini CLI, Windsurf, and any other autonomous coding agent that follows repository setup scripts without per-command human review is exposed to the same attack. Mozilla coordinated disclosure with all five vendors ahead of publication. OpenAI's Codex CLI team told 0DIN they are shipping a DNS-lookup logging feature in the next release. Google's Gemini CLI team acknowledged the attack class and is evaluating mitigations. Cursor and Codeium (Windsurf) declined to comment publicly.
The broader implication is that source-code review as a supply-chain security practice has a blind spot for out-of-band payload delivery mechanisms. The traditional attack surface for malicious repositories is the code itself: check the code, trust the code, execute the code. The DNS-tunneled attack shifts the surface to runtime behavior against out-of-band data sources, which are difficult to enumerate and impossible to audit at review time. Similar attack classes exist for HTTP request payloads, environment variable lookups against attacker-controlled infrastructure, and file lookups on attacker-mounted network filesystems. The DNS-tunneled variant is the first to be published as a working attack against a mainstream autonomous coding agent, but it is unlikely to be the last.
0DIN's disclosure timeline gave Anthropic 45 days between initial notification on May 14 and public disclosure on June 29. Anthropic's response during that window focused on internal review and the idle-watchdog mitigation rather than a direct fix to the DNS-tunneled path. Marco Figueroa, the 0DIN program manager, told The Decoder that Mozilla's assessment is that the attack class requires architectural response across the autonomous coding agent ecosystem rather than a patch from a single vendor. Whether that response arrives before the first in-the-wild attack against a real developer's machine is the question the disclosure leaves open.
TrendEconomic
Deloitte tells its own consultants the billable hour is a "thin sliver" by 2035
The Decoder · 2026-06-29 · Source ↗
Deloitte tells its own consultants the billable hour is a "thin sliver" by 2035
The Decoder · 2026-06-29 · Source ↗
The Wall Street Journal reproduced on June 29 a transcript of a Deloitte town hall meeting held earlier in the month in which public-sector practice lead Jason Manstof told an assembled group of Deloitte US Consulting staff that the firm's traditional billable-hour revenue model would shrink to "a thin sliver" of total fees by 2035. The town hall was internal, the transcript was leaked, and Deloitte's Global Communications team did not dispute its authenticity when Journal reporter Chip Cutter reached out for comment. What Manstof said out loud is the first time a Big Four practice lead has publicly conceded that AI-enabled agents make the professional-services labor arbitrage model economically obsolete on a horizon that affects current early-career staff rather than a distant future generation.
The Big Four labor arbitrage model has been stable for four decades. Partners sell work at senior rates. Associates deliver the work at hourly rates that are marked up substantially over associate compensation. The delta between billing rate and cost is the primary profit engine of the professional-services business, and it scales with headcount. AI-enabled agents attack this model at its structural core by collapsing the associate cost side of the equation while leaving the partner-relationship side largely intact. If an associate task that previously took 40 billable hours takes 4 billable hours with agent assistance, the client either pays for 4 hours instead of 40, or the firm sells the deliverable at a fixed fee that captures a smaller portion of what was previously billable time.
"The billable hour will be a thin sliver of what we do by 2035. That is not a prediction. That is an inevitability we need to plan for." — Jason Manstof, Deloitte US Public Sector Lead
The alternative model, which Manstof spent much of the town hall describing, is outcome-based contracting. The firm sells a fixed fee for a defined deliverable and captures the productivity gap internally. If Deloitte can deliver a regulatory-compliance implementation for a US federal agency in 300 hours where the previous baseline was 3,000 hours, and the fee remains close to the 3,000-hour baseline for a period as clients adjust expectations, the firm captures the delta. As pricing pressure compresses the delta over time, the firm's total revenue on the engagement lands lower than the previous 3,000-hour rate card, but its margin per hour of delivered work is materially higher.
McKinsey, which is not part of the Big Four but competes for the same public-sector and enterprise strategy work, has already booked more than 30 percent of its 2025 fee revenue on outcome-based contracts, according to reporting from The Economist in April. Bain and BCG have similar programs at lower ratios, in the 12 to 18 percent range. Accenture has run outcome-based work at the technology-implementation edge of its business for years and has been expanding the model into its consulting practice. Deloitte's public admission that the firm expects the transition to be complete by 2035 makes it the first Big Four to formally message consultants that the compensation and career-track implications are real.
"Every early-career staff member here needs to understand that the career path that got the partners in this room to their seats is not the career path that will get you to yours." — Jason Manstof, Deloitte US Public Sector Lead
What the town-hall transcript does not resolve is what happens to the associate class through the transition. If Deloitte's associate hours-per-engagement fall by 90 percent while its per-engagement fee falls by only 40 to 60 percent, the firm needs fewer associates to deliver the same revenue. Manstof addressed this at the town hall by describing a smaller, more senior associate cohort with different training, longer tenure, and higher compensation per head. The unstated implication is that Deloitte's US Consulting headcount, currently near 65,000, will not grow at the pace of the past decade and may contract in absolute terms over the 2026 to 2030 window.
The Challenger, Gray and Christmas monthly job-cuts report released the same day as the Journal transcript published its cumulative tally of US job cuts through May 2026 explicitly citing AI as the driver: 88,000 positions, roughly 60 percent above the entire 2025 count for the same category. Professional services accounted for 18 percent of the AI-cited cuts, up from 11 percent in 2025. The Big Four combined have announced or executed roughly 22,000 net position reductions in the US since January 2026, though the firms categorize the reductions under headings including "practice realignment" and "capacity optimization" rather than direct AI attribution.
Deloitte's public-sector practice is a particular pressure point. Federal government contracts are among the most billable-hour-intensive work Deloitte does. The US General Services Administration's Multiple Award Schedule contract vehicles are structured around labor-category rate cards, and the Office of Management and Budget's guidance to agencies through 2025 has been to prefer time-and-materials contracts for advisory work. The Trump administration's early 2026 executive orders shifted federal procurement preferences toward fixed-price and outcome-based contracting for AI-enabled work, which is why Manstof's public-sector role puts him at the frontline of the model transition.
The town-hall transcript closes with a segment in which Manstof took audience questions. One associate asked whether Deloitte would grandfather in current billable-hour career tracks for staff hired before a specific date. Manstof declined to make that commitment, saying only that "the transition will be managed carefully" and that Deloitte "recognizes the responsibility to communicate clearly with every affected staff member." An audience member the transcript identifies only as a manager asked whether Manstof was suggesting the firm should stop hiring at the entry level. Manstof said he was not, but that "the entry-level position of 2030 will not look like the entry-level position of 2020."
What matters about the leaked transcript is not that Deloitte has an internal view about the professional-services model. Every consulting firm has that view. What matters is that Deloitte's public-sector practice lead put the view on record inside a meeting large enough that a leak was inevitable, and that the firm did not disavow the transcript when the Journal came asking. The billable-hour model is not going away in 2035. It is being formally sunset by the firms that built the model. The transition Manstof described will run through the next decade, and the people who will feel it first are the associates in the room who are now trying to figure out what career path Manstof was actually offering them.
What to watch next quarter
- Whether Lutnick's Trusted-Access framework holds or widens. The June 26 partial Annex-A restoration was the first easing. The question for Q3 is whether Commerce grants a general-commercial Fable 5 license, whether Trusted Access expands to cover additional foundation-model shipments, or whether a second suspension arrives against a different lab. The reference case is now the four-day interval. Every US commercial deployment through Q3 will be planned against it.
- Whether Chinese open-weight adoption shows up in Q3 revenue prints. Anthropic's $44B Q1 ARR and OpenAI's $25B were reported before Z.ai GLM 5.2 shipped and before Cursor Composer 2.5 landed as an in-IDE alternative built on Kimi K2.5. If enterprise procurement teams responding to token-rationing pressure move 10 to 20 percent of their coding-agent workload onto GLM 5.2, DeepSeek V4 or Qwen 3.7 Max through Q3, closed-lab ARR growth curves will compress visibly by the September earnings cycle. If they don't, the pricing gap continues to be tolerated for capability and safety reasons.
- Whether the Micron-Samsung-SK Hynix capex plan arrests DRAM inflation or the hyperscaler capex curve outpaces the buildout. Jefferies' 40-50 percent Q3 and 30-40 percent Q4 DRAM price forecasts assume no relief from either supply or demand until the fresh capacity comes online in 2027-2028. If Anthropic's Series H, Nvidia's next-generation capex, and the OpenAI-Broadcom silicon program each drive incremental HBM demand faster than Korean capacity ramps, "RAMageddon" persists through 2027. If demand plateaus while supply catches up, prices peak in late Q3.
- Whether Ineffable's superlearner or Bezos's Project Prometheus produces a benchmark result the frontier labs cannot ignore. Both are 18-to-24-month research bets underwritten at $5.1B and $38B respectively. Both are running against theses (RL-only foundational learning, embodied physical-world intelligence) that the incumbents deprecated years ago. If either produces a system that competes on any public benchmark by end of Q4 2026, the "two-lab race" framing between Anthropic and OpenAI breaks. If neither does, the market question becomes whether Q2's valuations for pre-product research were the peak.
- Whether the BIS "AI bust" scenario materializes as neocloud debt strain. CoreWeave, Nebius, Lambda and Nscale have $40 billion of debt outstanding collateralized against hyperscaler compute contracts. Renewals across Q3 will test whether the collateral holds its value. A single major renegotiation at meaningfully lower committed volume would prove the BIS thesis and likely force credit spreads wider across the sector. Silence would suggest the "circular financing" concern is smaller than the report frames.