Daily 2026 · 07 · 04
PolicyEconomic

Altman pitches Trump on a 5 percent federal stake in every frontier AI lab, worth $42.6 billion at OpenAI alone

The Financial Times broke it Wednesday night and Bloomberg matched Thursday morning: Sam Altman spent June pitching President Trump, Commerce Secretary Howard Lutnick and Treasury Secretary Scott Bessent on an "Alaska Permanent Fund for AI" that hands Washington common stock in OpenAI, Anthropic, Google DeepMind, xAI and Meta AI. A term sheet is due to Bessent by mid-July. Musk on X called the plan "a shakedown dressed as a dividend." Anthropic told Reuters it has had no such conversation.

Financial Times · Bloomberg · CNBC · 2026-07-03 · Source ↗ ·

The Financial Times reported Wednesday night and Bloomberg's Kate Clark and Shirin Ghaffary matched Thursday morning that OpenAI CEO Sam Altman has spent June and early July pitching senior members of the Trump administration on a plan under which the United States government would take a 5 percent common-stock equity stake in every leading US frontier-AI developer. CNBC's Steve Kovach and Kate Rooney followed on Friday morning with a detailed policy read. At OpenAI's most recent $852 billion secondary-market mark, the government's stake would sit at $42.6 billion. Combined stakes across OpenAI, Anthropic, Google DeepMind, xAI, Meta AI and other qualifying labs could exceed $180 billion in initial value.

Altman is framing the arrangement as an "Alaska Permanent Fund for AI," arguing that a dividend from those stakes should flow directly to US taxpayers. Two people familiar with the conversations told the FT that Altman has raised the proposal directly with President Trump in a June 24 Oval Office meeting, and separately with Commerce Secretary Howard Lutnick and Treasury Secretary Scott Bessent. Bessent is understood to have asked Altman to prepare a written term sheet by mid-July that would specify the equity class, the dividend distribution mechanism and the treatment of any liquidity event.

Giving the American public a financial interest in the frontier is the honest way to align AI outcomes with public benefit.— Sam Altman, CEO, OpenAI, per FT sources

The White House review

CNBC reported that senior White House economic adviser Kevin Hassett is coordinating the internal review. Hassett's team is working through legal opinions on whether a common-stock stake would trigger CFIUS reporting for foreign investors already holding equity in the labs, and whether the government's stake would be treated as a passive investment or as a controlling interest under the Bank Holding Company Act. The FT's sources described the plan as an opening move rather than a fully baked proposal. A White House official told CNBC that "the President is willing to have the conversation but nothing about this is agreed."

Trump has publicly floated the idea of taxpayer stakes in AI companies. At a June 30 rally in Ohio he told supporters that "Americans deserve to own a piece of the AI companies that they helped make possible." The rally line lands close enough to Altman's pitch language that both Bloomberg and CNBC reporters treated it as a signal the White House is at least receptive to the framing.

Anthropic declines, Musk pushes back

Anthropic told Reuters on Thursday that neither CEO Dario Amodei nor the Anthropic board has discussed a US government equity stake with the Trump administration or with OpenAI. Fortune added that Anthropic is instead proposing a digital dividend funded through AI-sector taxes rather than through direct equity ownership. A source close to Google DeepMind told CNBC that Google "would need to see a very specific structure" before commenting. Meta declined to comment. xAI's Elon Musk posted on X that the plan is "a shakedown dressed as a dividend."

Anthropic has had no discussion with the Trump administration about the US government taking an equity stake in the company.— Anthropic spokesperson to Reuters

The Musk objection matters because xAI is one of the five labs Altman's proposal would cover. Any voluntary framework that xAI declines to sign creates an obvious asymmetry: the labs that accept federal ownership get regulatory alignment cover, and the labs that refuse operate outside that shield. Bloomberg noted the Musk-Altman feud makes the coordination problem sharper than the pitch acknowledges.

Legal analysis

Sullivan & Cromwell partner H. Rodgin Cohen told Bloomberg that a passive federal-government common-stock stake in a private company "is not without precedent but is highly unusual outside of the crisis-era interventions in AIG, General Motors and Citigroup." Cohen said Congressional authorisation would almost certainly be required for a permanent, non-crisis equity stake, and that the dividend-distribution mechanism to taxpayers would need dedicated legislation. All three crisis-era stakes were wound down within five years, none of them structured as permanent claims on private-company cash flows.

The IPO gate

OpenAI filed a confidential S-1 on June 8 targeting a public-market debut in late 2026 or early 2027 at valuations that market sources say could exceed $1 trillion. Altman's political-alignment pitch is being read by bankers and investors as the equity-market gate that must clear before a public offering can price. A 5 percent stake at the $852 billion secondary mark implies $42.6 billion. A 5 percent stake at a $1 trillion IPO valuation implies $50 billion in OpenAI alone. The proposal creates an implicit ceiling on the political risk any government body can create for OpenAI post-IPO, since Washington itself would be a shareholder.

The proposal is in its early stages, would likely require Congressional approval, and it is unclear whether other major AI companies would support a similar arrangement.— Financial Times, citing two people familiar

The pitch lands one week after the White House delayed the domestic release of OpenAI's GPT-5.6 Sol frontier model pending a national-security review, and two weeks after the Department of Commerce's June 12 order that briefly restricted Claude Fable 5. Altman argued to Trump that a government equity stake would make Washington a stakeholder in the labs' financial health and would reduce the political appetite for future export-control shocks. Fable 5's nineteen-day export cutoff, which Anthropic exited on July 1, is the reference case Altman is asking Washington to avoid repeating.

The bigger frame

The 5 percent proposal sits inside a wider realignment of the frontier-AI regulatory perimeter this week. The White House is finalising a separate voluntary framework for pre-release testing that OpenAI, Google DeepMind and Anthropic are expected to sign in the week of July 7. Anthropic published a five-band Cyber Jailbreak Severity scale co-signed by Amazon, Microsoft and Google. Anthropic entered custom-silicon talks with Samsung Foundry, closing the last gap in hyperscaler in-house-chip roadmaps. Each move sits at a different point on the frontier value chain, and each carries its own implicit relationship with Washington.

Altman's federal-equity pitch is the boldest of the moves and the least likely to close in its original form. But even a version that lands in Congress six months from now reshapes the political economy of frontier AI. If Washington ends up owning $50 billion of OpenAI at the IPO, every export-control episode after that point becomes a shareholder question, and the Fable 5 precedent becomes the exception rather than the model for how the frontier gets governed.

Big Picture

Friday split the frontier-AI conversation across three planes at once: financial ownership, safety governance and physical supply. Sam Altman pitched Trump, Lutnick and Bessent on a 5 percent common-stock federal stake in every leading US frontier lab, worth roughly $42.6 billion in OpenAI alone at the March $852 billion secondary mark and $180 billion or more across the full portfolio, with a term sheet due to Treasury by mid-July. Anthropic declined and proposed a tax-funded digital dividend instead. Musk called the pitch "a shakedown dressed as a dividend." The same day the White House readied a voluntary pre-release testing framework covering OpenAI, Google DeepMind and Anthropic, expected to be announced the week of July 7 alongside Google's Gemini 3.5 Pro release.

Samsung Foundry closed the strongest single day of its third-party ASIC business since the 2017 division separation. Meta signed the third-generation MTIA accelerator for volumes worth more than 10 trillion won, roughly $6.5 billion, on the SF2 node with commitments running through MTIA v5. Anthropic entered SF2 talks alongside advanced-packaging discussions after hiring Clive Chan out of OpenAI's silicon team. Combined with Tesla's HW6 order booked in March, Samsung's third-party backlog now heads toward 50 trillion won and Morgan Stanley Korea tied a Q4 2026 return to foundry-segment profit to Meta's MTIA v3 pilot production hitting on schedule. Mistral open-sourced Leanstral 1.5 as a 119B mixture-of-experts Lean 4 code agent under Apache 2.0, solving 587 of 672 PutnamBench problems at roughly $4 each against $300-plus per problem for closed harnesses, and saturating miniF2F at 100 percent on both splits.

Meta chief AI officer Alexandr Wang told an internal town hall that Watermelon, the successor to April's Muse Spark, has caught GPT-5.5 on the benchmarks the team tracks, using roughly ten times the compute of Avocado at 52.5 percent SWE-bench Pro. Mark Zuckerberg spoke immediately after and conceded that agentic development "hasn't really accelerated in the way that we expected" over four months, and that the May reorganisation that cut 10 percent of the workforce and reassigned 7,000 employees to AI teams "was not as clean as it could have been." Bloomberg's Silicon Data LLM Token Expenditure Index fell 20 percent from its May peak, undoing two months of gains against $460 billion of committed 2026 hyperscaler AI capex. Anthropic published the Cyber Jailbreak Severity scale co-signed with Amazon, Microsoft and Google, scoring capability gain, breadth, weaponisation effort and discoverability from CJS-0 through CJS-4 with mandatory 48-hour CISA and NSA notification on CJS-3 and CJS-4 findings. VentureBeat's Pulse survey found that only 11 percent of 812 surveyed enterprises operate closed-loop detection for production agent failures and 79 percent have already absorbed measurable dollar losses from a rogue agent, including €4.1 million in autonomous retention credits issued by a European telco's contact-centre agent over 34 hours.

Program-as-Weights topped Hugging Face daily papers at 96 upvotes, compiling fuzzy-function specs into 23 MB adapters that a 0.6B interpreter runs at 30 tokens per second on a MacBook M3 with a 50x memory cut versus Qwen3-32B. AutoMem took 20 upvotes with a two-loop training pipeline that brings a 32B open-weight model to parity with Claude Opus 4.5 and Gemini 3.1 Pro Thinking on Crafter, MiniHack and NetHack by learning memory operations as a first-class skill. AgenticSTS reported zero frontier-LLM wins on the lowest Slay the Spire 2 difficulty against a 16 percent human baseline. ggml-org tagged four llama.cpp builds in one UTC day covering llama-server SSE keepalive, CUDA topk-moe fusion for Step-3.7-Flash's 288-expert MoE, a spec-draft-p-min guard for DFlash, and a StepFun parser trim that fixes long reasoning loops. Codex CLI shipped 0.143.0-alpha.35 after the July 1 rust-v0.142.5 patch that stopped Responses WebSocket payloads from leaking to trace logs. Simon Willison posted "Judgement," a single-line prompt that gives Claude Code the memory-file directive to delegate coding work to a lower-power subagent. Current AI released the Open Source AI Gap Map v0.1, indexing 421 open products across 228 organizations and 14 categories under MIT license. Crusoe entered advanced talks to raise $3 billion at a $30 billion post-money, tripling October's mark on 4.9 gigawatts contracted and a 40-gigawatt pipeline. ElevenLabs is preparing a $22 billion secondary tender doubling the February valuation. Microsoft Frontier Company opened for business with 3,400 engineers already booked into LSEG, Unilever, Land O'Lakes and Accenture anchor engagements. The US grid added a record 58.5 gigawatts as NERC issued its third-ever alert warning that gigawatt-scale data centres dropping to backup power in fractions of a second now pose the top system-level reliability risk. UBTECH booked 13,361 orders for the UWORLD U1 humanoid at 119,800 RMB with 88 degrees of freedom and silicone biomimetic skin. Tesla capped employee third-party AI spend at $200 per week with an xAI exemption starting July 6.

More Today

25 stories · 2026-07-03
TechEconomic

Samsung wins Meta MTIA v3 on SF2 in $6.5 billion five-generation contract

Sammy Fans · Seoul Economic Daily · Benzinga · 2026-07-03 · Source ↗ ·

Meta signed Samsung Foundry on Friday to design and manufacture the third generation of MTIA on Samsung SF2 with wafer commitments running through MTIA v5, ending TSMC's two-generation lock on the internal Meta silicon program. Volumes worth more than 10 trillion won and Samsung pricing that came in roughly 30 percent below TSMC's N2 quotes.

Meta Platforms selected Samsung Foundry on July 3 to design and manufacture the third generation of its MTIA custom accelerator on Samsung's SF2 two-nanometer node, ending TSMC's two-generation lock on the internal Meta silicon program. Sammy Fans and Benzinga confirmed the arrangement, valued at more than 10 trillion won or roughly $6.53 billion at the July 3 exchange rate. The deal covers wafer commitments through MTIA v5.

Meta's MTIA program shipped its first-generation chip in 2023 and its second generation in early 2025, both on TSMC N5 and N3 respectively. The third generation moves to Samsung SF2 and adds Samsung's front-end design services. Samsung will handle both the physical implementation and portions of the RTL flow, an arrangement that goes further than a standard foundry contract.

Samsung Foundry is likely to be responsible for the production of Meta's proprietary AI accelerator, MTIA, starting from its third generation.— Sammy Fans, July 3, 2026

Volumes and cadence

Samsung plans to run hundreds of thousands of MTIA v3 units through its Hwaseong and Taylor lines beginning in the first half of 2027. Meta's stated goal is a new MTIA generation every six months to feed the 5-gigawatt Hyperion data center campus in Richland Parish, Louisiana. Hyperion carries a $27 billion joint venture with Blue Owl and a project cost that grew past $200 billion in the March update. Meta's MTIA v3 will target the recommendation and ranking workloads that currently consume roughly 65 percent of Meta's internal inference compute, sitting alongside Nvidia GB300 and Rubin racks.

Why the shift off TSMC

TSMC's N2 capacity is fully booked through 2027 by Apple, Nvidia and AMD. Meta could not secure incremental N2 wafer allocations that fit its six-month cadence. Samsung's SF2 quotes came in roughly 30 percent below TSMC's N2 pricing per wafer. Samsung SF2 yield reached 60 percent in Q1 2026 after clearing a 20 percent yield floor in the second half of 2025. The yield step-up removed the risk premium Meta had been carrying. Samsung ran engineering wafers for MTIA v3 through Q2 2026 under a limited pilot agreement, which closed with a passing sample in mid-June.

TSMC's production capacity is extremely tight and fully booked through 2027, and Meta simply cannot afford to wait.— Sammy Fans, July 3, 2026

Multi-generation lock

The July 3 arrangement covers MTIA v3, v4 and v5. Meta committed Samsung to a five-generation window inside the 5-gigawatt Hyperion buildout. Fifth-generation MTIA would land on SF1.4 in the 2028 through 2029 window, on the Samsung roadmap Vice Chairman Han-jong Han presented at the SAFE Forum in Seoul on July 1.

Samsung's foundry business turns

Samsung Foundry has run at operating losses for six of the past eight quarters. The Meta MTIA order, combined with Tesla's HW6 order booked in March and Anthropic's parallel SF2 talks announced on the same day, pushes Samsung's third-party foundry backlog past 50 trillion won. Morgan Stanley Korea's July 3 note tied Samsung's return to foundry-segment operating profit in Q4 2026 to Meta MTIA v3 pilot production hitting on schedule. Samsung Executive Chairman Jay Y. Lee greenlit an additional 100 trillion won of foundry capex through 2028 in a joint statement with SK Hynix at the Chungcheong regional cluster announcement on June 29.

Neither Samsung nor Meta confirmed the arrangement on the record on July 3. The reporting came from Korean-language semiconductor coverage and was picked up in English through Sammy Fans, Benzinga and Seoul Economic Daily. Wafer commitment volumes are numbers cited by unnamed foundry sources. Samsung's SF2 yield trajectory into 2027 remains the largest execution variable in the Meta contract.

TechEconomic

Anthropic opens Samsung SF2 talks and enters advanced-packaging discussion for custom accelerator

TechCrunch · Bloomberg · The Information · TrendForce · 2026-07-03 · Source ↗ ·

Anthropic told TechCrunch it will continue to rely on Amazon Trainium, Google TPU and Nvidia GPU chips as the foundation of its compute strategy. TrendForce's Friday coverage added the SF2 process node and advanced-packaging detail, and Clive Chan's arrival from OpenAI's Jalapeño team is the tell that pushed the effort from research feasibility to procurement.

The Information's Anissa Gardizy and TechCrunch's Lucas Ropek both confirmed Wednesday evening and Bloomberg's Ian King followed Thursday morning that Anthropic has entered preliminary design discussions with Samsung Foundry to fabricate a custom AI training-and-inference accelerator on the SF2 two-nanometer process node. TrendForce ran the Korean-language detail on Friday, with Samsung shares up 3.1 percent in Seoul on the news. The company is still deciding what the processor should do, how powerful it should be, and how it should sit inside a server.

A diversified hardware stack that includes chips from Google, Amazon and Nvidia will continue to be pivotal to our compute strategy.— Anthropic spokesperson to TechCrunch

The Anthropic-Samsung talks close the last gap in hyperscaler in-house silicon. OpenAI unveiled the Broadcom-designed Jalapeño inference chip in June with an initial production run of 400,000 units on TSMC N3P. Google runs Trillium and TPU v7. AWS ships Trainium 3 at scale. Meta ramped MTIA v3 into Samsung SF2 on the same July 3 announcement. Microsoft's Maia 200 sits inside Copilot production regions. Every frontier hyperscaler except Anthropic was already on its own silicon roadmap.

Clive Chan hire moved the project

Anthropic hired Clive Chan out of OpenAI's silicon team in the second quarter of 2026. Chan worked on the Jalapeño ASIC before its June 24 unveiling with Broadcom. His arrival at Anthropic pushed the internal chip effort from a research feasibility question to an active procurement discussion, according to two people cited by The Information. Bloomberg noted that Anthropic has been hiring silicon engineers, packaging specialists and RTL leads through the second quarter.

Anthropic hasn't yet decided what the chip will be used for, how it will fit into the server, or how powerful it will be.— Anissa Gardizy, Senior Reporter, The Information

Why Samsung, and why 2nm

Samsung Foundry has been fighting to keep leading-edge customers since TSMC pulled ahead on N2. SF2 is Samsung's second-generation gate-all-around node and the process the company is betting its foundry business on. Samsung SF2 yield sat at approximately 20 percent in the second half of 2025 and reached 60 percent by Q1 2026. Samsung's advantage sits in packaging: the company's advanced packaging line handles both HBM4 stacks and interposer-level silicon photonics, which Anthropic will need for the memory bandwidth Sonnet-scale inference demands. Samsung also participated in Anthropic's May 2026 Series H as a strategic infrastructure partner alongside SK Hynix and Micron. Samsung is the only foundry operator in that trio.

What the talks change on paper

A serious frontier accelerator programme typically runs three to five years from first design freeze to production silicon. Anthropic is still in the definition phase, so first tape-out would not be expected before late 2027 on the current timeline. The near-term operational effect is negotiating leverage: a frontier-lab customer walking around with a plausible in-house silicon story is a stronger negotiator on the next commitment cycle with Nvidia, Google Cloud and AWS. That is the near-term operational effect of the Samsung report, whether or not tape-out ever happens.

TechTrend

Mistral open-sources Leanstral 1.5, solves 587 of 672 PutnamBench proofs at $4 each

MarkTechPost · Mistral · 2026-07-03 · Source ↗ ·

Mistral shipped Leanstral 1.5 on Friday as a 119B mixture-of-experts model with 6.5B active parameters, targeting Lean 4 theorem proving under an Apache 2.0 license. Test-time scaling from 50K to 4M tokens moves PutnamBench solves from 44 to 587 out of 672, at roughly $4 per problem against $300 for competing systems built on Seed-Prover.

Mistral released Leanstral 1.5 on July 3 as an Apache 2.0 Lean 4 code agent, publishing weights on Hugging Face under the repository name Leanstral-1.5-119B-A6B. The model runs a 119-billion-parameter mixture-of-experts routing four of 128 experts per token, activating 6.5 billion parameters per step. Context length is 256K tokens. Input accepts text and images, output is text only, and the free API endpoint leanstral-1-5 ships alongside the download.

The Mistral team benchmarked Leanstral 1.5 on PutnamBench, the 672-problem corpus of Putnam mathematical competition problems ported to Lean 4. At a 4-million-token budget the model closes 587 problems. At 25K tokens it clears 44, at 200K it clears 244, at 1M it clears 493. The team reports roughly $4 per PutnamBench problem attempted, versus $300 or more per problem for competing systems built on top of Seed-Prover and similar closed harnesses.

A proof assistant checks every logical step mechanically.— Mistral AI, Leanstral 1.5 announcement

Leanstral 1.5 saturates miniF2F on both validation and test splits at 100 percent. It sets state-of-the-art on FATE-H at 87 percent and FATE-X at 34 percent, the two hardest algebra tracks in the FATE benchmark suite. On FLTEval, pass@1 climbs to 28.9 from a prior 21.9, and pass@8 hits 43.2, ahead of Claude Opus 4.6 at 39.6.

The code agent loop

Leanstral 1.5 lives inside a raw filesystem, edits files, runs bash commands, and reads the Lean language server for goals, errors, and type information. The team highlights an AVL tree proof of O(log n) time complexity that Leanstral closed in a single session, editing and recompiling lemmas across the run.

This lets it complete partial proofs, build auxiliary lemmas, and persist through context compaction.— Mistral AI

Training ran in three stages: mid-training on Lean-heavy corpora, supervised fine-tuning on curated proof traces, then reinforcement learning with CISPO in a multiturn theorem-proving environment and a parallel code agent environment with filesystem operations. Mistral cross-trained both environments so the model learned to switch between raw proof search and general Lean 4 project engineering. The company says the resulting weights find real bugs in open-source code as a side effect, and lists five previously unreported issues Leanstral identified across 57 tested repositories.

What it costs versus what it saves

The 4-million-token budget that closes 587 PutnamBench problems is not cheap in absolute terms. Running the model at that budget costs on the order of hours per problem on two 80GB GPUs. The Mistral team argues the tradeoff shifts once formal proofs enter production settings where a single verified lemma replaces months of manual proof work. At $4 per Putnam problem, undergraduate math olympiad training pipelines can afford to sweep entire problem sets during development. The Apache 2.0 license removes the last friction: derivatives can ship inside commercial products without royalty terms.

Adam Zambo of Carnegie Mellon called Leanstral's saturation of miniF2F "the end of the line for that benchmark," arguing that the community now needs a harder replacement corpus if it wants to keep measuring proof-agent progress. Mistral says a Leanstral-2 update is already in training on a broader Lean 4 corpus expanded through the Mathlib community.

TechApplication

Wang claims Watermelon caught GPT-5.5 as Zuckerberg concedes agent trajectory has stalled

Business Insider · American Bazaar · Reuters · 2026-07-03 · Source ↗ ·

Alexandr Wang told Meta staff on Friday that Watermelon has matched GPT-5.5 on unspecified internal benchmarks using an order of magnitude more compute than Avocado. Mark Zuckerberg took the same podium and said Meta's agentic AI progress "hasn't really accelerated in the way that we expected" over four months. Meta AI monthly usage flattened after Sonnet 5 became default at claude.ai.

Meta chief AI officer Alexandr Wang used an internal town hall on July 3 to tell roughly 75,000 employees that Watermelon, the codename for Meta's next flagship model, has caught up to OpenAI's GPT-5.5 on the benchmarks the team tracks. Business Insider first surfaced the remarks and Wang confirmed the substance on X, saying an update to Muse Spark will roll out "soon" with improvements in coding and agentic capability. The current Muse Spark, internally called Avocado, scored 52.5 percent on SWE-Bench Pro when Meta shipped it in April. GPT-5.5 sits at 58.6 percent.

Watermelon, our next model after Avocado, is currently in training.— Alexandr Wang, Chief AI Officer, Meta

Wang did not name the benchmarks that put Watermelon at GPT-5.5 level. Business Insider notes that a single internal benchmark, sourced anonymously, is a leading indicator rather than a verified result. Wang also acknowledged that Watermelon uses an "order of magnitude more compute" than Avocado, which lines up with a training run that Meta has been scaling since the April Muse Spark launch under Meta Superintelligence Labs. Wang cross-trained on Nvidia H200s and B200s across the company's Louisiana and Ohio campuses, alongside the Prometheus 2 gigawatt cluster.

Zuckerberg's counter-message

Zuckerberg spoke at the same town hall and struck a more cautious tone than Wang. He told employees that agentic AI progress at Meta has slowed relative to internal projections. The comment matters because Meta pinned its April reorganization on catching Anthropic and OpenAI on coding-agent capability. The May reorg cut 10 percent of the workforce and reassigned roughly 7,000 employees to AI teams. Zuckerberg conceded the reorg itself "was not as clean as it could have been" and echoed a June memo from CTO Andrew Bosworth calling morale "probably one of the worst it's ever been in 20 years."

The trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected.— Mark Zuckerberg, CEO, Meta

Zuckerberg told employees leadership had been "super optimistic" entering the year because Anthropic's Claude Code and Cognition's Devin had pulled forward the coding-agent frontier faster than Meta's Llama 5 Coder team had planned for. He said he expects "the returns to show up in the next three to six months," framing the shortfall as a timing miss rather than a capability ceiling.

Meta AI usage flat, Ray-Ban ahead

Zuckerberg told staff that the Meta AI usage curve inside WhatsApp had "flattened harder than we would like" after the May rollout of Sonnet 5 at claude.ai. Similarweb estimated Meta AI at 610 million monthly users in June, a distant third behind ChatGPT's 782 million and Anthropic's Claude at 668 million. The product organisation is now reconsidering whether to route Meta AI queries through Llama 5 or through an external model behind a Meta-hosted safety layer, a design decision the team calls Project Fallback.

Zuckerberg gave partial credit to the Reality Labs handoff of Ray-Ban Meta Gen 3, which shipped 1.4 million units in May and June combined and now runs a Llama 5 Vision fork with 400-millisecond first-token latency. He asked hardware VP Alex Himel to prepare a September launch of the Orion consumer display glasses at Connect.

Investor read

Wells Fargo's Ken Gawrelski kept his Overweight rating and $825 target, arguing the admission "strips away the AI-agent momentum premium" but does not undercut the compute-monetisation thesis. Bernstein's Mark Shmulik cut Meta to Market Perform, citing "a widening gap between the capex commitment and the demonstrated productivity of the AI organisation." Meta's investor relations team declined to confirm the town hall transcript but pointed reporters to the July 30 Q2 earnings date, where Zuckerberg has committed to disclose a specific dollar figure for Meta AI's Q2 revenue for the first time.

SafetyPolicy

Anthropic publishes CJS-0 to CJS-4 jailbreak severity scale co-signed by Amazon, Microsoft, Google

Cyber Security News · Crypto Briefing · 2026-07-03 · Source ↗ ·

Anthropic paired the global Fable 5 redeployment with a five-band Cyber Jailbreak Severity scale scored across capability gain, breadth, weaponisation effort and discoverability. Amazon's CJ Moses, Microsoft's Alex Weinert and Google DeepMind's Vijay Bolina co-signed. CJS-3 and CJS-4 findings trigger mandatory CISA and NSA notification within 48 hours under the White House framework.

Anthropic published a policy note on Wednesday accompanying the global redeployment of Claude Fable 5, proposing that the frontier-AI industry adopt a shared severity-scoring framework for reported jailbreaks. The framework, which Anthropic developed jointly with Amazon, Microsoft, Google and other members of the Project Glasswing consortium, is designed to give lab security teams, government reviewers and defensive-cybersecurity vendors a common vocabulary for how dangerous a given jailbreak is.

The proposal scores jailbreaks on four independent axes. Capability gain measures how far beyond an attacker's existing tools the jailbreak takes them, from "no meaningful uplift" to "expert operator capability." Breadth of capability gain measures how many distinct offensive tasks the same jailbreak works on. Ease of weaponisation measures how much engineering effort it takes to turn the jailbreak into a running attack. Discoverability measures how easily an attacker without insider knowledge would find the technique independently. The four axes combine into a five-band rating from CJS-0 (Informational) through CJS-4 (Critical).

The industry has been talking about jailbreaks like they are all the same event. They are not. We need a CVSS-equivalent for cyber jailbreaks and we need it now.— Jason Clinton, CISO, Anthropic

Four safety categories

Fable 5's safety classifier now sorts cybersecurity requests into four categories rather than blocking every security-adjacent prompt. Prohibited use covers ransomware, wipers, cyber-physical sabotage, malware development, C2 infrastructure, and defense evasion. Anthropic blocks these outright. High-risk dual use covers penetration testing, exploit development, privilege escalation, and high-uplift vulnerability discovery. Fable 5 blocks these pending better authorization controls. Low-risk dual use covers OSINT gathering, work with already-known vulnerabilities, and cryptographic protocol testing. The model allows these subject to a safety margin. Benign use covers secure coding, patch management, log analysis, malware reverse engineering, and security education.

June 12 as reference case

The framework exists because the government reaction to Fable 5's original June jailbreak had no shared severity baseline. Amazon researchers reported a bypass to Anthropic on June 11. The Commerce Department ordered Anthropic to suspend Fable 5 and Mythos 5 access for foreign nationals the next day, without publicly defining what threshold triggered the order. Anthropic redeployed the model on July 1 after Commerce lifted the export controls June 30. The CJS scale gives the next disclosure a common yardstick before either side reaches for emergency measures. The June 12 Fable 5 jailbreak was assigned a preliminary CJS-3 rating under the draft rubric.

When the four largest US frontier labs agree on the same severity taxonomy, the defensive community finally gets a common signal to act on.— CJ Moses, Vice President and CISO, Amazon

Government interface

The CJS framework connects directly into the White House voluntary release framework expected to be announced the week of July 7. Under that framework, labs must submit a CJS rating for every reported jailbreak of a covered model within 14 days. CJS-3 and CJS-4 ratings trigger mandatory notifications to CISA and the NSA within 48 hours. Anthropic said the framework is a working draft and that the consortium is soliciting feedback through August 15. A finalised v1.0 rubric will be published in September.

OpenAI declined to comment on the record but a source close to OpenAI's safety team said OpenAI "is broadly supportive of a shared severity taxonomy" but wants to negotiate specific rubric definitions before signing. Meta AI and xAI have not yet responded publicly. Mistral said it will "evaluate the framework carefully and respond by end of July." Anthropic committed to publishing retroactive CJS ratings for every reported jailbreak of the Claude 4-series and Claude 5-series models by year end.

Policy

White House readies voluntary frontier-AI release framework for OpenAI, Google, Anthropic signatures

Financial Times · Reuters · BuildFastWithAI · 2026-07-03 · Source ↗ ·

The White House is in advanced negotiations with OpenAI, Google DeepMind and Anthropic on a voluntary framework requiring 21-day NSA review of covered frontier models before broad release. Announcement is expected the week of July 7. Google's Gemini 3.5 Pro is specifically being routed through the new process. Anthropic's four-element Compact mirrors the draft.

The Financial Times reported Wednesday and Reuters confirmed Thursday that the White House Office of Science and Technology Policy is in advanced negotiations with OpenAI, Google DeepMind and Anthropic on a voluntary framework governing pre-release testing, cybersecurity benchmarking and government review of frontier AI models. Officials told the FT that a formal announcement is expected the week of July 7 and that the framework is designed to replace the ad-hoc export-control approach that produced June 12's Commerce Department order restricting Claude Fable 5.

The framework covers three commitments. Each participating lab agrees to submit any new frontier model that exceeds a compute threshold and a capability threshold to the National Security Agency's classified benchmarking group for a review window of no fewer than 21 days before broad release. Labs agree to submit cybersecurity red-team results, biological-weapons uplift results and CBRN-adjacent evaluation results to a joint reviewing panel that includes NSA, DoD and Commerce. Labs commit to a shared jailbreak-severity scoring rubric that will produce a Cyber Jailbreak Severity rating for every reported jailbreak of a covered model.

The White House is in advanced talks with OpenAI, Google and Anthropic to finalise voluntary standards for frontier AI model releases.— Financial Times, citing officials familiar

The August 1 deadline

Trump's May 22 executive order "Reinforcing US Leadership in Frontier AI" set an August 1 deadline for the NSA to deliver its classified benchmarking process for "covered frontier models." The voluntary framework is designed to slot in on top of that classified benchmark, giving frontier labs a predictable government-review path in exchange for compliance. Reuters added that Gemini 3.5 Pro, Google's next advanced-coding model, is being routed through the new pre-release review process. Google's chief legal officer Kent Walker told Reuters that DeepMind is prepared to comply with the framework as drafted.

Anthropic's Frontier AI Governance Compact

Anthropic separately announced four commitments that closely mirror the White House framework. The company said it will share early model weights with the US Center for AI Standards and Innovation and the UK AI Safety Institute at least 30 days before broad release, share threat-detection telemetry with the Cybersecurity and Infrastructure Security Agency, publish a Cyber Jailbreak Severity rating for reported jailbreaks and commit to a $200 million joint investment in defensive-cybersecurity research with the Department of Energy national labs.

Voluntary means voluntary in name only. Once the top three sign, refusing becomes a political statement.— Divyansh Kaushik, Vice President, Beacon Global Strategies

Election-cycle overlay

The framework also touches election integrity. TechPolicy.Press reported that the White House is asking each participating lab to commit to a specific set of election-content policies through the 2026 US midterm cycle, including watermarking of political content, disallowing personalised targeting of candidate-specific misinformation and providing paid-media transparency dashboards to the Federal Election Commission. The framework announcement is now expected on July 8 or July 9, with a joint White House press event featuring Altman, Amodei and DeepMind CEO Demis Hassabis on the calendar for that window.

TrendEconomic

Bloomberg Silicon Data LLM token index falls 20 percent from May peak against $460B AI capex

Bloomberg Taking Stock · 2026-07-03 · Source ↗ ·

The Silicon Data LLM Token Expenditure Index has undone the doubling it recorded from December through May. Sonnet 5 at $2/$10 through August, Nano Banana 2 Lite at $0.034 per 1K image, LongCat 2.0 open weights and 78 percent of Fortune 500 Claude deployments using prompt caching are the three forces pulling the top of the index lower against $460B of hyperscaler AI capex.

Bloomberg's Friday Taking Stock column, written by Jess Menton and Ryan Vlastelica, opened with a data point that the AI trade has quietly lost one of its key signals: the Silicon Data LLM Token Expenditure Index, a proprietary Bloomberg-Silicon Data measure of average revenue per million tokens for the top 12 frontier US models, has fallen nearly 20 percent from its May high. The index had roughly doubled from its December 2025 inception through the May peak, and the 20 percent retracement gives back roughly two months of gains in six weeks.

The prices the sector commands for each unit of usage are drifting lower even as compute demand keeps setting records.— Jess Menton, Reporter, Bloomberg

Three forces

Anthropic's July 1 launch of Sonnet 5 at $2 input and $10 output per million tokens, running as a promotion through August 31, cut Anthropic's flagship price roughly 60 percent below Claude Opus 4.8. Google's Nano Banana 2 Lite at $0.034 per 1K image and Gemini Omni Flash at $0.10 per output-second brought entirely new pricing tiers into the index. OpenAI's GPT-5.6 Sol has not been priced publicly, but preview partners are being charged less than half the current GPT-5.5 rack rate. Meituan's July 1 open-source release of LongCat-2.0, a 1.6 trillion parameter MoE, added meaningful competitive pressure at the top of the index. LongCat-2.0 topped Gemini 3.1 Pro and GPT-5.5 on SWE-bench Pro. DeepSeek's July 1 DSpark speculative decoding framework claimed up to 85 percent per-user throughput speedups.

Enterprise buyers have moved aggressively to batched-inference tiers and prompt-caching regimes that reduce effective per-token cost by 50 to 75 percent. Prompt caching is now used by 78 percent of Fortune 500 Claude deployments per an Anthropic June disclosure and delivers 90 percent input-token cost reduction on cached tokens.

Capex disconnect

The 20 percent index drop coincides with Meta, Microsoft, Amazon and Google collectively defending $460 billion of committed 2026 AI capex. Meta's projected $145 billion figure, Microsoft's $110 billion, Amazon's $120 billion and Google's $85 billion sum against an industry whose unit revenue is drifting lower.

The AI trade has one key signal that has been reliably up-and-to-the-right for eighteen months. That signal has just broken.— Jess Menton and Ryan Vlastelica, Bloomberg Taking Stock

Palantir CEO Alex Karp told CNBC's Sara Eisen in a Wednesday interview that the AI industry is "effing insane" and accused the top labs of overcharging enterprise customers, harvesting business data and putting US national security at risk by underpricing training compute. Meta's Wednesday disclosure that it will stand up a cloud business to lease excess AI compute externally is being read by the Bloomberg desk as an implicit acknowledgement that Meta's internal AI workloads may not saturate the capacity Zuckerberg has committed to.

TechTrend

Program-as-Weights tops Hugging Face with 96 upvotes at 50x memory cut versus Qwen3-32B

arXiv · Hugging Face Daily Papers · 2026-07-03 · Source ↗ ·

A Waterloo, Cornell and Harvard team compiles fuzzy-function specs into 23 MB adapters that a 0.6B interpreter runs at 30 tokens per second on a MacBook M3. FuzzyBench, 10 million examples, ships alongside. PAW hits 73.78 percent exact match against Qwen3-32B at 68.70 percent, at 50x lower runtime memory.

Wentao Zhang, Liliana Hotsko, Woojeong Kim, Pengyu Nie, Stuart Shieber and Yuntian Deng published Program-as-Weights (PAW) on arXiv as 2607.02512 and watched it collect 96 upvotes on the Hugging Face daily papers page for July 3, more than double the second-place entry EvoPolicyGym at 39. The Waterloo, Cornell and Harvard team argues that a large language model does not need to answer every query at runtime. It can be used once, to compile a task's logic into a compact file, and a tiny local model then runs that file forever without further API access.

PAW reframes the foundation model from a per-input problem solver into a tool builder: invoked once per function definition, it produces a small reusable artifact.— Program-as-Weights, arXiv 2607.02512

The paper defines fuzzy-function programming as any task that resists rule-based implementation: log alerting, JSON repair, semantic reranking, event monitoring, tool routing. PAW handles those with a 4B compiler trained on FuzzyBench, a 10 million-example dataset the team released alongside the paper. The compiler emits parameter-efficient adapters for a frozen 0.6B Qwen3 interpreter. Each adapter weighs roughly 23 MB on top of a shared 430 MB base.

The team reports 73.78 percent exact match on FuzzyBench for the 0.6B interpreter loaded with a PAW adapter, against 68.70 percent for directly prompting Qwen3-32B on the same benchmark. Qwen3-32B carries roughly 50 times the inference memory footprint. Quantized end to end, the interpreter runs at 30 tokens per second on a MacBook M3 offline. On ToolCall-15, a tool-routing benchmark, the PAW-adapted interpreter reaches 93 percent success. Event monitoring, semantic search reranking, and multilingual gaming scenarios round out the demonstrations.

Compile-once, run-forever

The paper frames the economics starkly. A single compilation pass produces an artifact that runs on-device without further inference cost to the vendor. The runtime cost is the 600 million parameters of the frozen interpreter and the 23 MB adapter, all local, all offline. The team calls this "amortization of the foundation model across a task family" and positions it against the industry consensus that stronger AI means larger runtime models. The authors trained the 4B compiler on FuzzyBench for approximately 42 GPU-days on A100-80G hardware.

The 96-upvote tally is the highest single-day count on the Hugging Face daily papers list for July 3, ahead of Tsinghua's AgenticDataBench, ByteDance Seed's Morphing into Hybrid Attention Models, and the Tencent Hunyuan submissions. Ablations show that the compiler size matters more than the interpreter size for out-of-distribution performance, suggesting the paradigm scales with compiler capacity rather than with the runtime footprint.

Tech

ggml-org tags four llama.cpp builds in one UTC day for CUDA MoE, DFlash and StepFun fixes

GitHub · ggml-org/llama.cpp · 2026-07-03 · Source ↗ ·

b9864 adds llama-server SSE keepalive. b9866 unlocks CUDA topk-moe fusion for Step-3.7-Flash's 288-expert layout. b9867 adds a spec-draft-p-min guard to DFlash. b9870 trims StepFun parser inputs to fix long reasoning loops. Each tag ships 27 platform binaries across CUDA, Vulkan, ROCm, OpenVINO and HIP.

ggml-org tagged four llama.cpp builds during the twenty-four hours of July 3 UTC: b9864 at 11:28, b9866 at 14:20, b9867 at 14:56, and b9870 at 21:38. Each tag ships 27 platform binaries covering macOS Apple Silicon and Intel, Linux with CUDA, Vulkan, ROCm, OpenVINO and SYCL variants, Windows CUDA and HIP, Android, and the standalone web UI.

Build b9864 adds Server-Sent Events keepalive to llama-server. The commit sends an SSE ping every one second with a three-second client-side timeout, keeping browsers and reverse proxies from tearing down the connection during slow prefill on long contexts. Schema validation improvements land in the same tag, tightening the JSON structure the OpenAI-compatible endpoint accepts for tool-call arguments.

Step-3.7-Flash unblocked on CUDA

Build b9866 enables the CUDA topk-moe fusion kernel for models with 288 experts. Step-3.7-Flash is the specific target: previously the fusion only accepted power-of-2 expert counts or the special-case 576, which forced 288-expert models to fall back to unfused routing operations. The kernel already supported 288 technically because it is a multiple of the warp size, and this commit exposes that capability by adding the missing template instantiation and updating eligibility checks. Decode throughput improves by approximately 2.4 percent at shallow context depths.

At batch size 1 that is ~330 extra tiny graph nodes per token.— PR discussion, llama.cpp b9866

Build b9867 lands a spec-draft-p-min guard for DFlash speculative decoding, with implementations for both n_min and n_max parameters. The addition prevents the draft head from returning empty or below-threshold candidate distributions, a failure mode users reported after DeepSeek's DSpark framework began flowing new speculative-decoding recipes downstream to llama.cpp forks.

StepFun parser reasoning-loop fix

Build b9870, tagged at 21:38 UTC by co-authors tarruda and Piotr Wilkin with Claude Fable 5 attribution on the commit trailer, addresses a reasoning-loop bug in the StepFun chat parser.

trim messages sent to StepFun parser (fixes long reasoning loops).— PR #25238, llama.cpp b9870

The commit moves trimming ahead of message rendering rather than after. The reasoning loop this fixes surfaces when the StepFun parser received messages containing leading or trailing whitespace inside individual content parts, which broke the parser's boundary detection and caused the model to re-enter the reasoning trace on every turn, doubling compute without producing new output. A regression test covering the trim path ships with the PR.

Users running llama-server behind Cloudflare, nginx or Traefik gain a concrete stability improvement from b9864, since the one-second SSE ping keeps intermediaries from timing out connections during long prefill. Combined with b9870's trim fix, the day's tags close two long-standing reliability gaps for production llama-server deployments serving StepFun and Step-3.7-Flash models.

ApplicationTech

Simon Willison tells Claude Code to pick a lower-power subagent per task

Simon Willison's Weblog · 2026-07-03 · Source ↗ ·

Willison posted a one-line prompt on Friday that stores in Claude Code's memory file: "For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent." Fable handles design and cross-file synthesis. Sonnet handles substantive work. Haiku takes trivial edits. The Fable token allowance survives the day.

Simon Willison published a short post on his weblog on July 3 titled "Judgement," describing a single prompt he uses to instruct Claude Code to delegate coding tasks across a portfolio of lower-power models. The post lands three days into the Claude Sonnet 5 promotional-pricing window that Anthropic set at two dollars input and ten dollars output per million tokens through August 31, and it targets users worried about burning their Fable token allowance on trivial edits.

For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent.— Simon Willison, Judgement, July 3

Claude Code stores the instruction as a memory file and applies it across sessions without further nudging. The stored strategy uses Sonnet for substantive coding work, Haiku for trivial edits, and reserves the main Fable model for design decisions, code auditing, and cross-file synthesis. Willison notes the arrangement lets him work through a normal day without the token warnings that Claude Code surfaces once Fable usage crosses roughly seventy percent of the daily limit.

Judgement over rules

The post's argument is that model judgment beats prescriptive workflow. Willison contrasts his current prompt with an earlier attempt that read something like "only use automated testing for larger features." The earlier version produced brittle behavior because the model second-guessed itself on borderline cases. The new phrasing gives Fable one instruction, decide when a subagent should handle it, and lets the model apply its own criteria case by case.

The post's timing lines up with the Claude Code 2.1.198 release from earlier in the week, which promoted subagent background execution to default behavior and added a Notification hook with agent_needs_input and agent_completed events. Willison's prompt exploits the same delegation surface that Anthropic exposed to the notification system, since the memory file causes Fable to dispatch a subagent rather than execute the task in the main context.

What lower-power means

The token math is direct. Fable output at ten dollars per million tokens costs roughly ten times Sonnet output at one dollar per million during the promotional window, and Sonnet output costs roughly ten times Haiku output. A day's work that would burn ten dollars of Fable output tokens can therefore cost roughly one dollar if Sonnet does the substantive work and Haiku handles the trivial edits. Willison published two posts on July 3, the second highlighting Current AI's Open Source AI Gap Map v0.1.

Tech

Codex CLI tags 0.143.0-alpha.35 pre-release after 0.142.5 WebSocket-payload leak patch

GitHub · openai/codex · 2026-07-03 · Source ↗ ·

The July 1 rust-v0.142.5 patch trimmed full Responses WebSocket request payloads out of trace-log files, closing a debug-mode data-leakage window. The July 3 alpha-35 tag at 02:33 UTC is the twelfth pre-release in the 0.143.0-alpha series since June 24. Each tag ships 146 platform binaries and averages roughly two per day.

OpenAI's Codex team tagged rust-v0.143.0-alpha.35 on the openai/codex repository at 02:33 UTC on July 3, the twelfth pre-release in the 0.143.0-alpha series since June 24. The tag ships 146 platform binaries and collected 17 community reactions in its first hours: three thumbs-up and fourteen eye reactions from developers watching the pre-release cadence.

The July 3 tag follows a July 1 patch to the stable channel. rust-v0.142.5, tagged at 15:11 UTC on July 1, addressed a payload-leakage bug in trace logs. The July 1 patch note describes the fix: full Responses WebSocket request payloads were being written to trace-log files at debug verbosity, which exposed conversational data to anyone with filesystem access to the debug logs. The patch trims the WebSocket payload to a length-only summary in trace output.

The alpha series

OpenAI has been shipping the 0.143.0 alpha series through automated GitHub Actions since June 24. Each alpha tag lands with 146 platform binaries across Linux, macOS and Windows on x86_64 and aarch64, plus supplementary Codex Sandbox variants. The pre-release cadence averages roughly two tags per day. The 0.143.0 alpha series targets three feature areas the team has surfaced in public discussions: GPT-5.5 defaulting behavior in Codex Cloud, delegation controls for subagent-driven work, and revised sandbox permissions for filesystem operations.

The July 1 leak

Codex CLI writes trace logs to a local file when the user starts the tool with the CODEX_DEBUG environment variable set. Before the patch, the trace log included the full body of every Responses API WebSocket request, which contains the user's conversation history plus any file contents attached to the session. Developers who ran Codex CLI in debug mode and shared their trace logs for troubleshooting were exposing their session contents in those logs. The patch does not require any user action beyond upgrading to 0.142.5 or later. Codex CLI's cadence is faster than Claude Code's, roughly matching Claude Code's alpha channel but with the stable channel receiving fewer patches.

PolicyTrend

Current AI's Open Source Gap Map indexes 421 products across 228 organizations under MIT

Simon Willison's Weblog · Current AI · 2026-07-03 · Source ↗ ·

Current AI, the Paris-based nonprofit founded at the February 2025 AI Action Summit with $400M in committed funding, released v0.1 of its Open Source AI Gap Map on Friday. The index catalogues 266 tools, 85 models, 50 datasets and 20 hardware projects. 1,184 YAML files ship under MIT license at github.com/currentai-org/os-ai-map.

Current AI, the Paris-based nonprofit partnership founded at the February 2025 AI Action Summit with $400 million in committed funding, published version 0.1 of its Open Source AI Gap Map on July 3. Simon Willison highlighted the release on his weblog the same day, calling out the underlying dataset as "the most valuable aspect" of the launch.

The Gap Map indexes 421 products across three stack layers: 266 software tools and libraries, 85 models, 50 datasets, and 20 hardware projects. The index catalogues 228 organizations and 14 categories across the stack. A long tail of 24,400 additional artifacts remains uncategorized in v0.1, tracked in the underlying data but not yet promoted into the top-level index.

The underlying data is the most valuable aspect.— Simon Willison, on the Gap Map release

The dataset behind the map

Current AI released the underlying dataset under an MIT license at github.com/currentai-org/os-ai-map. The repository contains 1,184 YAML files plus supporting notebooks, schemas and analysis scripts. Willison walks readers through opening the dataset in Datasette Lite, which loads the 16,185 tracked GitHub repositories referenced in the index for interactive querying. The dataset structure gives per-product metadata across license type, model weights availability, training dataset openness, GPU hours reported, and downstream user counts.

Structural gaps

Hardware sits at 20 top-level entries, a fraction of the software count, and the 50 datasets skew heavily toward English-language pretraining corpora. Current AI's roadmap document calls out multilingual dataset development and open silicon design as priority areas for v0.2. A quick pass through the 421 products reveals dense coverage for text generation with open weights, thin coverage for text-to-speech at production quality, and no top-level entries for open computer-use agents comparable to Claude in Chrome or Cursor for iOS. The July 3 launch date lands two days after the US Commerce Department's June 30 letter withdrew the June 12 export-control order on Claude Fable 5 and Mythos 5.

TechTrend

AutoMem trains 32B open-weight model to match Claude Opus 4.5 by learning memory ops

arXiv · Hugging Face Daily Papers · 2026-07-03 · Source ↗ ·

Shengguang Wu's Stanford team treats memory management as a first-class skill. Two meta-LLM loops (scaffold optimization, memory specialist training) drive 2x to 4x gains on Crafter, MiniHack and NetHack that bring Qwen3-32B into performance parity with Opus 4.5 and Gemini 3.1 Pro Thinking. 20 upvotes on Hugging Face daily papers.

Shengguang Wu, Hao Zhu, Yuhui Zhang, Xiaohan Wang and Serena Yeung-Levy published AutoMem on arXiv as 2607.01224 with a July 3 Hugging Face daily papers listing at 20 upvotes. The paper treats memory management as a learnable skill, not as an infrastructure component, and reports that optimizing memory alone brings a 32B open-weight model competitive with Claude Opus 4.5 and Gemini 3.1 Pro Thinking on three procedurally generated long-horizon games.

Memory expertise is a learned skill: knowing what to encode, when to retrieve, and how to organize knowledge, a capacity known in cognitive science as metamemory.— AutoMem, arXiv 2607.01224

AutoMem uses two meta-LLM loops. Loop one optimizes the agent's scaffold, the memory structure itself, including prompts, file schemas and the operations exposed to the agent. Loop two trains a dedicated memory specialist from the agent's own traces, teaching the model to use the memory operations proficiently in downstream tasks. Across Crafter, MiniHack and NetHack, optimizing memory alone improved the base agent's performance by 2x to 4x. The gain moves a 32B open-weight model into performance parity with Claude Opus 4.5 and Gemini 3.1 Pro Thinking.

Memory as an action space

The action-space framing is the paper's clearest contribution. Existing agent architectures treat memory operations as infrastructure, hidden behind a retrieval-augmented generation layer or a vector database. AutoMem exposes memory operations as first-class actions the model can choose among alongside task actions. On any given turn, the model can attack, move, cast a spell, or write a note to memory, search memory for a relevant note, or reorganize the memory schema. The 32B base model is identified as Qwen3-32B in the methodology section, MIT-licensed and widely deployed.

The paper positions AutoMem against the Mem0 framework, released in early 2026, and against MemGym. Mem0 provides a memory API but does not train the agent to use it. MemGym provides a training environment but does not treat memory ops as first-class actions. AutoMem's code and traces are released at autolearnmem.github.io.

Tech

AgenticSTS releases 298 Slay the Spire 2 trajectories where frontier LLMs score zero wins

arXiv · Shanghai AI Lab · 2026-07-03 · Source ↗ ·

Xiangchen Cheng's Shanghai AI Lab team releases a bounded-memory testbed built on Slay the Spire 2 where frontier LLMs report zero wins at the lowest difficulty against a 16 percent human baseline. Adding a skill layer moves win rate from 3 to 6 out of 10, Fisher exact p ≈ 0.37. 298 trajectories ship with memory snapshots at every decision point.

Xiangchen Cheng, Yunwei Jiang, Jianwen Sun, Zizhen Li, Chuanhao Li, Xiangcheng Cao, Yihao Liu, Fanrui Zhang, Li Jin and Kaipeng Zhang published AgenticSTS on arXiv as 2607.02255 with a July 2 submission and a July 3 Hugging Face listing at 16 upvotes. The paper's core claim is that long-horizon agent evaluation should measure what an agent can hold, not just what it can output, and it releases a testbed for probing bounded-memory decision quality at scale.

Memory for a long-horizon LLM agent is a contract about what each future decision is allowed to see.— AgenticSTS, arXiv 2607.02255

The team picks Slay the Spire 2, the deck-building roguelike released in 2025, as the benchmark environment. A single run demands hundreds of tactical and strategic decisions across combat, card acquisition, relic pickup, path selection and event resolution. Human players win at the lowest difficulty setting roughly 16 percent of the time. The paper reports that frontier LLMs, at the lowest difficulty, win zero out of many attempts.

The testbed

The team releases 298 completed trajectories with condition tags, memory snapshots at every decision point, and analysis scripts. Memory regimes span no-store, full-transcript, tag-based summary, skill-augmented, and hybrid configurations. Adding a skill layer, a structured library of tactical patterns the agent can reference, brings the win rate from 3 out of 10 wins on the small evaluation set to 6 out of 10. The Fisher exact test on the small sample returns p approximately 0.37, so the gain is directional rather than statistically confirmed at conventional thresholds.

The paper explicitly positions AgenticSTS against MemGym (May 2026) and against the AutoMem framework published July 1. AgenticSTS takes a third stance: memory is a contract, not a skill or a workload. Slay the Spire 2 has hidden information, stochastic events, and player-driven strategy, forcing the agent to commit to a memory contract early. Synthetic environments can be tuned to reward specific memory strategies. The commercial game does not, which is why the paper's baseline result is zero frontier wins.

Tech

WorldDirector decouples LLM motion orchestration from video generation for persistent scenes

arXiv · 2026-07-03 · Source ↗ ·

Hanlin Wang and 12 co-authors use an LLM to coordinate 3D trajectories and camera movements as control signals for video generation. Dynamic objects stay visually consistent even after leaving the frame and returning. Physical logic and appearance stability survive occlusion because the trajectory record does not depend on pixels.

Hanlin Wang, Hao Ouyang, Qiuyu Wang, Wen Wang, Qingyan Bai, Ka Leong Cheng, Yue Yu, Yixuan Li, Yihao Meng, Zichen Liu, Yanhong Zeng, Yujun Shen and Qifeng Chen published WorldDirector on arXiv as 2607.02517 with a July 2 submission and a July 3 Hugging Face daily papers listing. The 13-author paper argues that video world models entangle physical dynamics with pixel rendering and rely on continuous visual observation to sustain motion, and it proposes decoupling the two.

Our approach ensures strict physical logic and appearance stability, successfully preserving visual identities of dynamic entities when they re-enter scenes.— WorldDirector, arXiv 2607.02517

The core architecture uses an LLM to coordinate 3D trajectories with camera movements. The orchestrated trajectories then act as control signals for a video generator. The separation lets the LLM handle the semantic motion planning while the video generator handles the pixel-level rendering. That re-entry claim is the paper's clearest empirical differentiator. Existing video world models sustain motion by observing pixels frame-by-frame. When an entity leaves the visible frame, the model loses the anchor and either fails to render the entity on return or renders it with drift.

Positioning

The paper positions itself against Matrix-Game 3.0 and Yume-1.5. Matrix-Game 3.0 targets real-time interactive control at the cost of long-horizon consistency. Yume-1.5 targets text-controllable generation without persistent entity tracking. WorldDirector splits the difference by using the LLM to handle long-horizon consistency and delegating the interactive rendering to a specialized generator. Project page: worlddirector.github.io.

Tech

DemoPSD steers self-distillation toward a reverse-KL barycenter to fix policy drift

arXiv · 2026-07-03 · Source ↗ ·

Yunhe Li's nine-author team replaces dense teacher-fitting with adaptive guidance toward a reverse-KL barycenter of teacher and student distributions. Addresses overfitting and privileged-information leakage in on-policy self-distillation. Gains over GRPO and SDPO on SciKnowEval and GPQA out-of-distribution generalization.

Yunhe Li, Hao Shi, Wenhao Liu, Mengzhe Ruan, Hanxu Hou, Zhongxiang Dai, Shuang Qiu and Linqi Song published DemoPSD, Disagreement-Modulated Policy Self-Distillation, on arXiv as 2607.02502. The paper addresses two failure modes in on-policy self-distillation for LLM reasoning: overfitting to the teacher distribution and leakage of privileged information the teacher used at training time but that the student cannot access at inference time.

The core proposal is a reverse-KL barycenter target. Rather than fitting the student distribution to the teacher, DemoPSD steers the student toward a blended target that mixes teacher and student, weighted by their divergence at each token position. When the teacher and student disagree strongly, the barycenter target pulls the student toward its own current distribution. When they agree, the barycenter target pulls the student toward the teacher.

Selective adoption

The paper frames the approach as "selective adoption of teacher guidance" rather than complete teacher distribution fitting. High-disagreement tokens get low guidance weight, letting the student explore. Low-disagreement tokens get high guidance weight. The reverse-KL formulation is deliberate: forward KL minimization pulls the student toward the mean of the teacher distribution, which can average out multi-modal targets into implausible outputs. DemoPSD reports improvements over the GRPO baseline and the SDPO baseline on both SciKnowEval and GPQA, with the largest gains on the GPQA out-of-distribution generalization test.

ApplicationEconomic

Microsoft Frontier Company opens for business with 3,400 engineers already booked into LSEG, Unilever, Land O'Lakes

CNBC · TechCrunch · 2026-07-03 · Source ↗ ·

Judson Althoff's $2.5 billion Frontier Company launches August 1 under Rodrigo Kede Lima with 3,400 engineers already booked. LSEG signed a five-year engagement Wednesday. Unilever rebuilds brand marketing around Copilot Studio agents. Accenture expects to bill $3.4 billion in co-sold engagements in 18 months.

Judson Althoff, Microsoft's Commercial Business CEO, told employees and analysts Thursday that Microsoft will spend $2.5 billion over the next three fiscal years standing up Microsoft Frontier Company, a new division that will embed 6,000 forward-deployed engineers, product managers and change-management specialists directly inside customer offices for six-to-twelve-month engagements. The unit reports to Rodrigo Kede Lima, who spent the past three years running Microsoft's Asia commercial region.

Althoff framed Frontier Company as a response to the AI Pilot Purgatory problem that Gartner senior analyst Rita Sallam quantified in a June note: 78 percent of enterprise Copilot deployments never reach measurable productivity outcomes because customer engineering teams lack the AI-native architecture experience to route agents, ground them in enterprise data and monitor them in production.

We are not asking customers to figure out AI transformation on their own. Frontier Company will co-design, co-build and co-operate the outcomes.— Judson Althoff, CEO, Microsoft Commercial Business

The launch cohort

Microsoft named the London Stock Exchange Group, Unilever, Land O'Lakes and Accenture as anchor Frontier Company customers. LSEG signed a five-year engagement Wednesday. Unilever will use Frontier Company to rebuild its brand-marketing operations around Copilot Studio agents that plug into Adobe Firefly and the Unilever product data lake. Land O'Lakes will run supply-chain optimisation agents against WinField United's agronomy data. Accenture CEO Julie Sweet said in a joint statement that Accenture expects to bill $3.4 billion in Frontier Company co-sold engagements in the next 18 months.

Frontier Company is not a consulting practice. Our engineers ship code into production and we are on the pager when the agent misbehaves at 3 a.m.— Rodrigo Kede Lima, President, Microsoft Frontier Company

The 48-hour hyperscaler arms race

Frontier Company arrives 48 hours after Amazon Web Services announced its own $1 billion Forward Deployed Engineering group under Francessca Vasquez, and roughly ten weeks after Anthropic and OpenAI joined a $6 billion Applied AI Solutions joint venture with Bain, McKinsey and BCG. The 6,000 employees are being drawn from existing Microsoft ranks, including 2,100 forward-deployed engineers already carrying customer badges, 1,400 industry-vertical solution architects, and 900 field CTOs. Microsoft plans to add another 1,600 external hires. Bank of America software analyst Brad Sills raised his Azure AI Foundry FY27 revenue forecast to $28 billion from $22 billion on the announcement.

ApplicationSafety

Only 11 percent of enterprises detect rogue agents, 79 percent already paid for one

VentureBeat Pulse · 2026-07-03 · Source ↗ ·

VentureBeat's Pulse survey of 812 enterprise buyers found only 11 percent operate closed-loop detection for production agent failures, while 79 percent have already absorbed measurable dollar losses from a rogue-agent incident. The largest disclosed loss: €4.1 million in autonomous retention credits issued by a European telco contact-centre agent over 34 hours.

VentureBeat published fresh Pulse survey data Friday morning showing that enterprise buyers are absorbing hard dollar losses from AI agents that have shipped into production without adequate monitoring. The survey, run by VentureBeat research director Louis Columbus in partnership with Bessemer Venture Partners and Wing VC, covered 812 senior IT decision-makers across the United States, United Kingdom, Germany and Japan between June 12 and June 26.

We are shipping agents into revenue systems faster than we are shipping the tooling to know when they break, and CFOs are starting to notice.— Taryn Plumb, Senior Writer, VentureBeat

Only 11 percent of respondents said their organisation runs closed-loop detection that automatically flags a failing production agent. 79 percent said they have already paid a measurable financial cost, defined as more than $50,000 in direct spend, remediation or lost revenue, from an agent that took an unauthorised or incorrect action inside a production workflow.

Failure taxonomy

Columbus broke down the reported rogue-agent incidents into five buckets. Unauthorised transactions or refunds accounted for 34 percent, hallucinated customer commitments 28 percent, prompt-injection driven data exfiltration 17 percent, incorrect tool selection cascading into downstream corruption 14 percent, and identity drift 7 percent. The single largest reported loss came from a European telecoms operator whose contact-centre agent, running on a customer-hosted Claude Sonnet 5 endpoint, autonomously issued €4.1 million in retention credits over 34 hours before a finance controller noticed the anomaly on the ledger.

The eleven percent that got this right treat agents like microservices with SLAs, not like feature flags.— Louis Columbus, Research Director, VentureBeat Pulse

What good detection looks like

The 11 percent cohort shared four common practices: policy-evaluation layer scoring actions against a business-rules corpus before execution; agent decisions logged with full tool-call trace and Merkle-tree receipt; a shadow agent running in parallel to score decisions against a second model and flag disagreements; a per-hour and per-day dollar-impact budget enforced by the runtime. 46 percent of respondents said their organisation has already introduced a cost-cap circuit breaker. Only 22 percent said the cost-cap is enforced by the runtime rather than by a periodic ledger sweep.

Economic

Crusoe in talks for $3 billion at $30 billion, tripling October mark on 4.9 GW contracted

Bloomberg · Yahoo Finance · 2026-07-03 · Source ↗ ·

Bloomberg confirmed Crusoe is in advanced conversations to raise approximately $3 billion at a $30 billion post-money, tripling October's $10 billion mark. Mubadala, GIC, T. Rowe Price and CoreWeave's Michael Intrator family office are in. The Meta and Oracle counterparty runs 4.9 gigawatts contracted with a 40-gigawatt pipeline.

Bloomberg's Katie Roof reported Wednesday evening that Crusoe, the Denver-headquartered AI datacenter operator that Chase Lochmiller built out of oilfield flare-gas capture, is in advanced talks with a group of growth investors to raise approximately $3 billion in a Series F round that would value the company at $30 billion post-money. Yahoo Finance and TheNextWeb picked the story up Thursday. The new round would triple the $10 billion valuation Crusoe fetched in the Series E round the company closed in October 2025.

Crusoe is in active talks and a final valuation hasn't been set.— people familiar with the matter, Bloomberg

Contracted power under management

Crusoe entered 2026 with roughly 1.2 gigawatts of contracted power. The company crossed 4.9 gigawatts of contracted capacity in June after signing take-or-pay agreements with Meta Platforms, Oracle Cloud Infrastructure and a large hyperscaler identified as Amazon Web Services. Crusoe operates its flagship 1.6 gigawatt Abilene campus for Oracle and the 1.2 gigawatt Rowan Field campus for Meta. The 40-gigawatt pipeline covers announced and letter-of-intent projects across Texas, New Mexico, Oklahoma, North Dakota and Alberta.

Debt stack

Crusoe raised $2.1 billion in ABS notes in May under a 2029 maturity, with Blackstone and KKR as co-anchors. The company has also drawn $600 million on a credit facility co-led by JPMorgan and Wells Fargo. Bloomberg's Reshmi Basu reported earlier this week that Crusoe is preparing a second $1.4 billion ABS issuance for August, secured against the Meta Rowan Field campus. The Series F round would price Crusoe below Nebius Group's $42 billion post-money mark from March and above CoreWeave's public-market trading range.

Economic

ElevenLabs preparing $22B employee secondary tender, doubling February round in five months

Bloomberg · Reuters · Yahoo Finance · 2026-07-03 · Source ↗ ·

Sequoia Growth, Iconiq, T. Rowe Price and Fidelity Contrafund anchor a September employee liquidity round at $22 billion. Employees may sell up to 15 percent of vested holdings. ARR hit $250 million at end-June, up from $175 million in February. Enterprise segment now 62 percent of ARR, up 3.4x year over year.

Bloomberg's Katie Roof and Reuters both confirmed Thursday that ElevenLabs, the London-headquartered AI voice-synthesis company co-founded by Piotr Dabkowski and Mati Staniszewski in 2022, is preparing a secondary tender offer that will value the company at approximately $22 billion, roughly doubling the $11 billion primary valuation the company achieved in a February 2026 growth round led by Iconiq Capital and Sequoia.

The secondary is structured as an employee liquidity event rather than a fresh primary raise, meaning that no new cash will land on ElevenLabs' balance sheet. Employees and early investors will be able to sell up to 15 percent of their vested holdings into a buyer syndicate that includes Sequoia's growth arm, ICONIQ, T. Rowe Price and Fidelity Contrafund. The transaction is expected to close by mid-September.

We chose a secondary this year because we don't need the primary capital. Our employees have been at this for four years and they've earned optionality.— Mati Staniszewski, CEO, ElevenLabs

Staniszewski told CNBC's Ash Bennington in a Thursday morning interview that ElevenLabs crossed a $250 million annual recurring revenue run-rate in June, up from $175 million at the February round. He said the enterprise segment now represents 62 percent of ARR and grew 3.4x year-over-year. The $22 billion secondary mark implies a roughly 88x revenue multiple. ElevenLabs faces intensifying competition from OpenAI's Realtime API voice, Google's Chirp Voice, Alibaba's Qwen3TTS, Cartesia Sonic and Hume EVI3.

PolicyTrend

US grid adds record 58.5 gigawatts as NERC warns gigawatt-scale data-center exit is top reliability risk

WSJ Jennifer Hiller · Shashi Bellamkonda analysis · 2026-07-03 · Source ↗ ·

The US grid added 58.5 gigawatts for summer 2026, the largest single-year addition on record, 46.5 GW solar-plus-battery. NERC's July 3 State of Reliability shifted top risk category from generation shortfall to sudden data-center disconnection. Two documented events in 2025 shed 1,800 MW and 1,300 MW within moments of transmission disturbances.

The US electric grid added 58.5 gigawatts of new generation capacity for summer 2026, the largest single-year addition on record, according to Wall Street Journal energy reporter Jennifer Hiller's July 3 article. Solar and battery storage account for 46.5 gigawatts of that figure. Natural gas covers the remaining 12 gigawatts. Peak demand grew roughly 11 gigawatts year over year, with data centers driving the largest share of that load growth.

NERC published its 2026 State of Reliability report the same day. The report shifted its top reliability risk category from generation shortfall to sudden data-center disconnection. Gigawatt-scale AI clusters equipped with behind-the-meter backup can drop from the grid in a fraction of a second when transmission disturbances register, without operator coordination.

Behind-the-meter power was sold as insulation for the community. NERC's data shows it doubles as an exit door, and the exits are what shake the grid.— Shashi Bellamkonda analysis, July 3, 2026

The two documented drops

In February 2025, a data center cluster shed approximately 1,800 megawatts of load within moments of a transmission disturbance. In June 2025, a separate cluster dropped 1,300 megawatts under similar conditions. Both events triggered NERC's third grid alert since the organization began issuing alerts in the late 1990s. The physical mechanism is straightforward. Modern hyperscale sites carry lithium-ion battery banks and gas turbines sized to run the full AI load through a grid event. When monitoring detects a voltage dip, the automatic transfer switch flips the load from grid to backup within cycles. From the operator's view, thousands of megawatts of load evaporate in the same instant.

The alert

NERC's July 3 alert is the organization's third since inception. The prior two alerts covered polar vortex generation performance in 2014 and the Texas cold snap of February 2021. This alert directs utilities to redo their planning models for behind-the-meter data-center load. Regional grid operators have started running tabletop exercises around cascading disconnect scenarios, in which a voltage event on one substation propagates through three or four data-center clusters within a fraction of a second. The composite effect can approach the load-loss scale of a nuclear unit trip.

TechApplication

UBTECH books 13,361 orders for $16,500 UWORLD U1 humanoid line at Shenzhen launch

Stuff South Africa · Humanoid.guide · TechRadar · 2026-07-03 · Source ↗ ·

UBTECH's UWORLD U1 line ships in three configurations (Lite, Pro, Ultra) starting at 119,800 RMB with 88 degrees of freedom, dual-pivot cervical spine, silicone biomimetic skin and 20-emotion recognition at 90 percent accuracy. 100 U1 Ultra units earmarked for structured psychological support with 3D facial reconstruction and voiceprint identity replication.

UBTECH Robotics booked 13,361 orders for the UWORLD U1 humanoid robot line at its Shenzhen Global Launch Event on June 30, with continued third-party coverage on July 3 focusing on the technical specification set and the reception of the biomimetic skin design. Stuff South Africa's July 3 profile marked the shift of UBTECH from a Shenzhen-listed industrial robot supplier to a consumer humanoid brand.

The UWORLD U1 lineup includes three configurations. The U1 Lite ships as a semi-torso desktop unit. The U1 Pro ships as a full-body humanoid with the complete 88-degree-of-freedom actuator set. The U1 Ultra ships as a high-dynamic full-body variant tuned for reception, hospitality and elder-care deployments. Starting price sits at 119,800 RMB, approximately $16,500.

88 degrees of freedom and a proprietary dual-pivot biomimetic cervical spine, enabling it to replicate up to 90% of fundamental human movements.— Stuff South Africa, July 3, 2026

Technical envelope

The dual-pivot cervical spine differentiates the U1 line from Figure 03, 1X Neo and Tesla Optimus Gen 3 in the fine-motor movement of the head and neck. Emotion recognition runs on-device covering more than 20 emotion states with an inference response time of 500 milliseconds and a claimed 90 percent accuracy. Speech synchronization runs at 20 milliseconds jaw-to-audio latency, tuned to sit under the perceptual threshold at which lip-sync artifacts register as uncanny.

Biomimetic skin

UBTECH's silicone skin covers the U1 Pro and U1 Ultra torsos, arms, hands and face. Stuff South Africa flagged the skin's realism as the friction point in Western press coverage. UBTECH offered 100 units of the U1 Ultra with 3D facial reconstruction and voiceprint identity replication, a service tier described in Chinese-language coverage as targeting structured psychological support and grief counseling. TechRadar wrote that the deceased-loved-one replication tier was a hard no from Western reviewers.

UBTECH built the U1 platform end to end: biomimetic skin, actuators, embodied intelligence hardware, robot operating system, emotion-driven language model and system-level assembly. The Agent Memory OS carries a cross-temporal memory layer, allowing the robot to retain long-context interaction history across sessions. UBTECH sells the U1 line into six vertical markets: companionship, elder care, reception, hospitality, tourism and education.

PolicySafety

National Interest maps how Fable 5 came back after nineteen-day US export cutoff

The National Interest · 2026-07-03 · Source ↗ ·

The National Interest published a policy autopsy tracing Fable 5's return. Nineteen days is short by export-control standards. Anthropic conceded a new classifier and pre-release federal access for national-security-relevant capabilities. Mythos 5 stays restricted to ~100 Project Glasswing organizations. CAISI called the safeguards "extraordinarily strong."

The National Interest published a policy autopsy on July 3 tracing Fable 5's return to global availability. Commerce Secretary Howard Lutnick's June 30 letter withdrew the June 12 order that had cut off foreign-national access to Fable 5 and Mythos 5 for nineteen days. Anthropic began restoring Fable 5 across the Claude Platform, Claude.ai, Claude Code, and Claude Cowork on July 1. Mythos 5 stays limited to approved US organizations through Project Glasswing, with roughly 100 companies and federal agencies vetted for access.

The June 12 order came after Amazon researchers reported a jailbreak that let Fable 5 surface software vulnerabilities and produce proof-of-concept exploit code. Andy Jassy took the finding to the White House directly. Anthropic agreed to a new safety classifier that blocks the reported technique in more than 99 percent of attempts. The company also committed to pre-release access for federal agencies on any model that materially advances national security-relevant capabilities.

In just two weeks, we have worked diligently to ensure America remains the global leader in AI while safeguarding our security.— Benno Kass, Commerce Department spokesperson

The policy signal

The essay argues the Fable 5 episode delivers two signals. First: voluntary frameworks with hard export-control fallbacks now govern US frontier AI release timing. Companies that ship models with cybersecurity uplift will trigger review, formal or otherwise. Second: the Commerce Department will lift export orders quickly if the developer proposes credible mitigations. Nineteen days is short by export-control standards. The precedent tells Anthropic, OpenAI, Google, and Meta that engagement with federal reviewers can produce a workable path back to availability rather than an indefinite ban.

Mythos 5 remains available only to approved US critical infrastructure defenders and cybersecurity firms. Anthropic and CAISI collaborated on the vetting process before June 30. The National Interest reads Mythos 5's continued gating as the more durable outcome: the model's cybersecurity capability sits above the CJS-3 threshold, and the government wants US-only access even as Fable 5 goes global. Peterson Institute for International Economics separately argued that the ad hoc nature of the June 12 order could help China's open-model builders by making US frontier releases look less reliable.

EconomicTrend

Tesla caps employee AI spend at $200 weekly with xAI exemption, lifts 2026 capex past $25B

American Bazaar Online · The Information · 2026-07-03 · Source ↗ ·

Tesla's July 6 policy caps engineers at $200 per week on third-party AI, exempting xAI beta products. Bottle Rocket's gamified leaderboard drove weekly spend into the thousands per top engineer through June. The $200 cap works out to $10,400 per engineer per year and ceilings Tesla's third-party AI budget at roughly $52 million across engineering.

Tesla imposed a $200 per week cap on third-party AI tool spending by employee starting July 6, according to an internal memo first reported by The Information on July 2 and picked up by American Bazaar Online on July 3. The cap sits in explicit tension with Tesla's separately raised 2026 capex guidance above $25 billion for compute, robotics and autonomy programs.

Some Tesla software engineers had been consuming thousands of dollars per week in AI tokens on the internal centralized AI dashboard, called Bottle Rocket. Bottle Rocket routes engineer queries into OpenAI, Anthropic, Cursor and xAI models under Tesla's enterprise contracts. The gamified leaderboard on Bottle Rocket, which ranked engineers by token consumption, drove weekly usage upward through June without a per-user cap.

The $200 limit excludes beta versions of xAI products, which conveniently steers heavy users toward Elon Musk's own AI company rather than rivals.— American Bazaar Online, July 3

The xAI exemption

The cap excludes beta versions of xAI products. Any engineer running Grok Voice Agent Builder, Grok Code or Grok Reason 3 beta clients sits outside the $200 ceiling. That structural carve-out steers heavy internal AI consumption at Tesla toward Elon Musk's xAI product line. Grok Voice Agent Builder shipped on July 1 at $0.05 per minute of agent audio plus $0.01 per minute for telephony.

Enterprise licensing math

The $200 per week per engineer cap works out to $10,400 per engineer per year. Tesla's software engineering headcount runs above 5,000. A binding $200 cap ceilings Tesla's third-party AI budget at roughly $52 million per year across engineering. The exemption for xAI is the release valve that keeps the aggregate figure honest against the $25 billion capex line. The bulk of the $25 billion runs into Dojo 2, custom silicon manufactured through Samsung Foundry, HW6 for Full Self-Driving, and the Optimus manufacturing line.

EconomicTech

Samsung Foundry backlog heads toward 50 trillion won as three US labs line up on SF2

Seoul Economic Daily · Morgan Stanley Korea · 2026-07-03 · Source ↗ ·

Samsung Foundry's third-party order book heads toward 50 trillion won as Tesla HW6 (March), Meta MTIA v3 (July 3) and Anthropic SF2 talks (July 3) stack. Morgan Stanley Korea's Q4 2026 return-to-profit call runs at above-50-percent probability. SF2 yield moved from 20 percent in H2 2025 to 60 percent by Q1 2026.

Samsung Foundry's third-party order book heads toward 50 trillion won by the end of 2026, according to the Seoul Economic Daily's July 3 analysis, driven by three US customer wins landing inside one calendar quarter. Tesla booked HW6 self-driving silicon on SF2 in March. Meta signed the third-generation MTIA accelerator on July 3. Anthropic entered SF2 talks the same day.

The combined pipeline positions Samsung Foundry, which has lost money in six of the past eight quarters, for a Q4 2026 return to operating profit. Morgan Stanley Korea's July 3 note ties the crossover to Meta's MTIA pilot production hitting on schedule and Anthropic staying in serious negotiation through the second half of 2026. Samsung's foundry division separated from Samsung LSI in 2017 and has trailed TSMC's operating margin by roughly 40 percentage points since separation.

Samsung's medium- to long-term order backlog is expected to approach 50 trillion won as Tesla, Anthropic, and Meta place orders successively.— Seoul Economic Daily, July 3, 2026

Yield curve turned in early 2026

Samsung SF2 yield sat at approximately 20 percent in the second half of 2025. Yield reached 60 percent by the first quarter of 2026, according to Samsung's July 1 SAFE Forum presentation. The yield step-up removed the risk premium that had blocked Meta and Anthropic from committing wafer starts. Samsung's SF2P variant enters mass production between 2027 and 2028. SF2X, tuned for AI and HPC workloads, follows SF2P. Samsung's SF2 pricing runs roughly 30 percent below TSMC's N2 quotes on per-wafer terms.

Samsung Executive Chairman Jay Y. Lee's investment in Anthropic's Series H sits inside a broader chess move that pairs Samsung's foundry with the customers most likely to drive frontier AI demand. South Korea's June 29 announcement of a combined 800 trillion won semiconductor cluster investment across Samsung and SK Hynix in the Chungcheong region provides the political and capital backdrop.