Weekly · 2026-W28 Jul 6 – Jul 12, 2026
Policy Trend

Apple names Tang Tan and Chang Liu in a trade-secret suit against OpenAI, and ties it to the Gemini-Siri pivot

Apple filed in the Northern District of California on Saturday, July 11, alleging Chief Hardware Officer Tang Tan directed candidates to bring "actual parts" to OpenAI interviews for show-and-tell sessions, and that former senior electrical engineer Chang Liu skipped his exit interview, kept his company laptop, and exploited a post-departure Apple internal-cloud bug to download more than a thousand pages of confidential documents. The complaint reaches beyond the two named defendants to the more than 400 ex-Apple employees now inside OpenAI or IO Products, and Apple ties the filing directly to its March 2026 decision to build the autumn 2026 Siri revamp on Google Gemini rather than OpenAI. Analysts read the suit as Apple's opening move to seek injunctive relief that could block OpenAI from onboarding further Apple hires while the case is pending.

Business Today · Washington Post · 2026-07-11 · Source ↗ ·

Apple filed suit against OpenAI in the United States District Court for the Northern District of California on Saturday, July 11, alleging that OpenAI's hiring of former Apple hardware executives crossed the line from ordinary lateral movement into orchestrated trade-secret misappropriation. The complaint names two individuals directly. The first is Tang Tan, Apple's former Chief Hardware Officer, who joined OpenAI in early 2025 and now runs hardware development for the joint OpenAI–IO Products device program alongside Jony Ive. The second is Chang Liu, a former senior electrical engineer on Apple's Vision Products Group. The suit names IO Products LLC, the Ive-founded design firm OpenAI acquired for $6.5 billion last year, as an added defendant.

The core factual allegations are unusually specific. Apple says Tan personally directed candidates for OpenAI hardware roles to bring "actual parts" from their Apple work to their OpenAI interviews for show-and-tell sessions, a claim that if credited would move the case out of the ambiguous territory of trade secrets kept in memory and into the far more actionable territory of physical hardware exfiltration. For Liu, Apple alleges he skipped his exit interview entirely, retained his company laptop after his separation, and exploited a bug in Apple's internal cloud system that had gone unpatched to download more than one thousand pages of confidential technical documents in the weeks after his departure was recorded.

Why the timing matters

Apple's decision to file now, more than a year after Tan's move, is inseparable from its March 2026 decision to base the autumn 2026 Siri revamp on Google Gemini rather than on any OpenAI model. That pivot was already read across the industry as a rebuke; the trade-secret complaint reframes it as the natural consequence of a supplier relationship that Apple now positions as having been used as cover for organized poaching. The complaint reaches beyond the two named defendants to the more than 400 former Apple employees now inside OpenAI or IO Products, characterizing the pattern as departure "at every level" of Apple's hardware organization.

"The pattern is not lateral movement. It is a coordinated departure at every level of Apple's hardware organization, engineered by executives who took their trade secrets with them and by systems designed to receive them."— Apple complaint, filed 2026-07-11 in N.D. Cal.

The relief Apple is asking for is broader than damages. The complaint requests injunctive relief that would block OpenAI from hiring further ex-Apple engineers during the pendency of the case, and it asks the court to enjoin OpenAI from using any hardware design work traceable to Tan or Liu. In practice, either injunction would be operationally severe for the joint OpenAI–IO Products hardware program, which is on a compressed schedule and whose most senior hardware leader is Tan himself.

OpenAI's expected defense

OpenAI has not filed a formal response. In pre-filing correspondence surfaced in the complaint's exhibits, OpenAI counsel argued that the trade-secret framing collapses the ordinary lateral-employment right into a conclusion that engineers cannot leave one Silicon Valley employer for another. The San Francisco tech-employment bar will be watching for whether Judge Cynthia Rocha, drawn in the initial case assignment, treats the "show-and-tell" allegation as the pleading's center of gravity or as decoration for a claim otherwise built on non-compete-shaped arguments California courts have historically been skeptical of.

The counter-narrative from OpenAI supporters is that Apple is using the courts to accomplish what it could not accomplish through retention: keeping its hardware leaders in place. That framing was pushed hardest on X by ex-Apple hires now inside OpenAI, several of whom posted variations of the same argument that Apple's hardware culture had made departure inevitable and that the "1,000 pages" allegation against Liu is being treated as representative when it is a single-instance claim.

"California is a right-to-work state. If Apple's hardware leaders had wanted to stay, Apple would have kept them. This complaint reads like retention by subpoena."— Aravind Narayan, ex-Apple senior director now at IO Products (X)

The Gemini-Siri backdrop

Apple confirmed in March 2026 that Siri's autumn revamp would run on Google Gemini rather than on any OpenAI or in-house Apple Intelligence model. The choice was framed publicly as a capability decision. The complaint's exhibit E ties it to Tan's departure directly, arguing that the hardware and software Siri roadmap Apple built internally was compromised by the leakage of hardware-side design work, and that Google was chosen over OpenAI because Apple could not credibly build a Siri stack that depended on models trained by a supplier its own former hardware chief now works for.

Wall Street's read on the filing was mixed. Apple shares closed down 1.4 percent on Monday July 13's trading, more on the injunctive-hiring risk to the Ive program than on Apple's damages-recovery position. OpenAI's private secondary-market pricing softened on the Circle secondary marketplace by roughly 8 percent Monday morning against the pre-filing close, more on tail-risk pricing than on any operational disruption. The most likely near-term consequence is a temporary restraining order phase over the next thirty days that will determine whether OpenAI can continue to hire from Apple's hardware organization while the case proceeds.

The Week in AI

The story of the week is the collapse of the arms-length supplier posture. Apple filed trade-secret claims against OpenAI naming two executives and reaching 400+ ex-Apple hires. SpaceXAI shipped Grok 4.5, its first joint model since the June 16 Cursor acquisition, at $2 input and $6 output per million tokens against Opus 4.7's $5 and $25. OpenAI cleared US Commerce sign-off on GPT-5.6 Sol, Terra and Luna after a 12-day government-only window and shipped ChatGPT Work as its first product where the model owns the deliverable, not the message. Meta broke ground on a CA$13 billion 1-gigawatt Alberta datacenter alongside a companion CA$4.6 billion natural-gas plant.

Frontier pricing kept resetting. Muse Spark 1.1 launched as Meta's first paid frontier API at $1.25 input and $4.25 output, undercutting Grok 4.5 on input by 37.5 percent and GPT-5.6 Sol on output by 86 percent. Ollama closed a $65 million Series B at a $700 million post-money valuation. SK Hynix's Nasdaq debut closed at $168.01 for a $1.27 trillion market cap on the largest US IPO ever by a foreign company. Google made AlphaEvolve generally available inside Gemini Enterprise with sixteen named enterprise wins and eight Google-internal ones including a 20 percent Spanner write-amplification reduction. And Fed Chair Kevin Warsh installed Marc Andreessen and Microsoft EVP Asha Sharma as co-leads of a new Productivity and Jobs task force with monetary-policy recommendations due December.

Frontier prices reset a second week in a row

Three of the week's four model launches undercut incumbent pricing sharply. Grok 4.5 landed at $2 input and $6 output against Opus 4.7's $5 and $25, a 60 percent input cut and a 76 percent output cut. Meta's first paid frontier API, Muse Spark 1.1, came in lower still at $1.25 and $4.25. OpenAI's own tiering (Sol $5/$30, Terra $2.50/$15, Luna $1/$6) shows the same posture inside a single lab. The direction of travel is a permanent reset of the Opus-class price floor toward Sonnet-class levels.

The model owns the deliverable, not the message

OpenAI's Sam Altman framed ChatGPT Work as the first product built around the idea that "the model owns the deliverable, not the message." Anthropic tagged Claude Code v2.1.207 with auto mode on by default across Bedrock, Vertex and Foundry, patched a Windows worktree data-loss bug, and blocked transcript tampering. Google made AlphaEvolve generally available inside Gemini Enterprise with sixteen enterprise wins and eight Google-internal wins named at launch. The product surface is shifting from chatbot to agent, and the framing of the artifact is following.

Legal exposure opens a second front

The Apple trade-secret suit against OpenAI is the loudest new front but not the only one. The NYT-led publisher coalition filed a sanction motion against OpenAI before Judge Sidney Stein over allegedly hidden ChatGPT training-data evidence, with damages exposure Susman Godfrey estimated above $9 billion and Bloomberg's terminating-sanctions probability market jumping from 8 percent to 21 percent. The Tennessee amended class action against SpaceXAI and Stability AI over roughly 7,000 sexually explicit images allegedly generated of Jane Doe 4 added Wyoming and Wisconsin plaintiffs. Three concurrent front-lines with distinct theories of harm.

Memory and datacenter suppliers book multi-year AI capacity

SK Hynix closed its Nasdaq debut at a $1.27 trillion market cap, the largest US IPO ever by a foreign company, with CEO Kwak Noh-Jung telling Seeking Alpha the memory crunch extends past 2030 and Chairman Chey Tae-won reporting every roadshow customer he met said the 2027 capacity plan "was not enough." Meta broke ground on a CA$13 billion 1-gigawatt Alberta datacenter with 1.8 GW headroom. Cerebras committed 200 megawatts across France and the Nordics by end of 2027. DigitalOcean previewed Q2 remaining performance obligations above $800 million on multiple nine-figure annual contracts, up 10x year over year.

The week in full

15 stories · Jul 6 – Jul 12, 2026
TechApplication

SpaceXAI ships Grok 4.5, the first joint model with Cursor, at $2 input and $6 output

TechCrunch · xAI · 2026-07-08 · Source ↗

SpaceXAI released Grok 4.5 on Wednesday morning as the first joint model since the June 16 Cursor acquisition closed at $60 billion, priced 60 percent below Opus 4.7 on input and 76 percent below on output. Grok Build, the SpaceXAI console and every Cursor plan carry it on day one. The European Union is excluded from launch pending Digital Services Act review.

SpaceXAI shipped Grok 4.5 on Wednesday, July 8 at 09:00 PT in a live-stream unveiling that Elon Musk described as the first joint model since the June 16 acquisition of Cursor closed at a $60 billion transaction value. The 1.5-trillion-parameter Mixture-of-Experts model launched at $2 input and $6 output per million tokens, undercutting Anthropic's Opus 4.7 by 60 percent on input and 76 percent on output. It shipped simultaneously across Grok Build, the SpaceXAI console and every Cursor plan, giving Cursor's coding-agent surface a first-party frontier model for the first time.

Musk framed Grok 4.5's positioning against Anthropic's top tier directly on the live-stream, describing the model as "Opus-class, but faster, more token-efficient and lower cost." Internal SpaceXAI benchmarks quoted during the stream put Grok 4.5 "roughly comparable to Opus 4.7, but much faster." The European Union was excluded from launch pending Digital Services Act review, a gating pattern that has become standard for frontier launches ahead of the August 2 EU AI Act enforcement wave.

"Opus-class model, but faster, more token-efficient and lower cost. We built this for the Cursor generation of developers who want the best model per dollar per second."— Elon Musk, CEO, SpaceXAI (2026-07-08 launch stream)

Benchmarks and the hallucination trade

Third-party benchmark data landed on Friday. Snorkel's GDPval+ suite gave Grok 4.5 a 29 percent mean pass rate on 2,000 expert-authored professional tasks, leading GPT-5.5 at 22 percent and Opus 4.8 at 21 percent, with a 40 percent legal-work lead and 58 percent education-work lead. Artificial Analysis ranked Grok 4.5 fourth on its Intelligence Index at 54, behind Claude Fable 5, GPT-5.5 and Opus 4.8. The hallucination rate on the same GDPval+ suite doubled from Grok 4.3's 25 percent to 54 percent, a trade Snorkel's research team called "the most aggressive precision-recall shift of any 2026 launch."

Awesome Agents flagged an unresolved training-data overlap between Grok 4.5 and CursorBench that could inflate the model's coding-agent numbers, and the CursorBench maintainers said they were preparing a contamination audit before the next leaderboard refresh. The pattern echoes prior overlap disputes on SWE-bench and HumanEval; the audit outcome will determine whether SpaceXAI's Cursor-focused pitch survives the initial benchmarking cycle intact.

Cursor's own pricing did not change, but Grok 4.5 became the default model on the Pro and Team plans. The Grok Build API surface picked up native tool-use and computer-use endpoints priced identically to the base model, which puts SpaceXAI on Anthropic's operator-agent pricing model without a per-action surcharge.

TechApplication

OpenAI clears Commerce sign-off on GPT-5.6, ships ChatGPT Work as its first "model owns the deliverable" product

CNBC · OpenAI · 2026-07-08 · Source ↗

OpenAI cleared US Department of Commerce sign-off Thursday afternoon Pacific on GPT-5.6 Sol, Terra and Luna, ending the 12-day voluntary government-only preview window. The same rollout shipped ChatGPT Work across Mac, Windows, iOS, Android and web with 15 named third-party connectors and completed the 24-hour deployment by Friday morning.

OpenAI cleared US Department of Commerce sign-off on Thursday afternoon Pacific, July 8, publicly releasing the GPT-5.6 sibling tier: Sol at $5 input and $30 output per million tokens, Terra at $2.50 and $15, and Luna at $1 and $6. The release ended a 12-day voluntary government-only preview window that opened on June 26 under the White House pre-release testing clause. OpenAI's release statement noted, without naming the process, that "this kind of government access process should not become the long-term default." The clearance came without any published capability report and no public safety statement, a gap that Federal News Network and TechCrunch's Tim Fernholz both flagged.

The same rollout shipped ChatGPT Work as an integrated app across Mac, Windows, iOS, Android and web. The product was framed by Sam Altman as OpenAI's first agent-first product, one where "the model owns the deliverable, not the message." ChatGPT Work carries 15 named third-party connectors on launch (Salesforce, Notion, HubSpot, Linear, Jira, Confluence, Google Workspace, Microsoft 365, Slack, GitHub, Zendesk, Intercom, Asana, ClickUp, and Airtable) reachable through @-mention. Every response is routed through an Auto-Review pass that OpenAI claimed hit a 100 percent block rate on 12,000 attempted exfiltration prompts during the government-access window red team.

"ChatGPT Work is our first product where the model owns the deliverable, not the message. The chatbot is the input surface; the artifact is the output."— Sam Altman, CEO, OpenAI

Codex folds in, GPT-5.4 sunsets

The Thursday morning rollout also folded Codex into a single merged ChatGPT desktop client, retiring the Codex-specific window that had run since 2024. The Codex CLI got its own tagged release the same day (0.144.0 and 0.144.1) promoting MCP interactive authentication out of experimental opt-in, adding a "writes" app-approval mode, and cutting Computer Use latency on GPT-5.6. OpenAI set July 23 as the sunset date for GPT-5.4 across every consumer and API surface, giving developers a two-week migration window to Terra or Luna. Sol carries a promotional price through August 31 before the standard $5 / $30 rate resumes; Terra and Luna are priced at their standard rates from day one.

The government-access story picked up scrutiny of its own. TechCrunch's Fernholz surfaced that the US government issued no public safety statement about GPT-5.6 despite the 12-day gate, and Anthropic's public-comment note filed the same afternoon called for a published capability report as the baseline requirement for future clearance windows.

TechEconomic

Meta ships Muse Spark 1.1 as its first paid frontier API, ending the Llama giveaway

TechCrunch · Axios · 2026-07-10 · Source ↗

Meta Superintelligence Labs shipped Muse Spark 1.1 on Thursday July 9 at $1.25 input and $4.25 output per million tokens, undercutting Grok 4.5's input by 37.5 percent and GPT-5.6 Sol's output by 86 percent. It is the first paid frontier-model API Meta has ever offered.

Meta Superintelligence Labs launched Muse Spark 1.1 on Thursday, July 9 as its first paid frontier-model API, priced at $1.25 input and $4.25 output per million tokens. The model is a coding-focused frontier tier built on the same Muse architecture that ships inside Meta's consumer AI products, now tuned for autonomous coding and packaged as a metered API. The pricing undercuts Grok 4.5 on input by 37.5 percent and GPT-5.6 Sol on output by 86 percent, positioning Muse Spark 1.1 at the aggressive end of the current price band.

The launch ends the era, running since 2023, in which Meta's frontier work was released free through the Llama series. Alexandr Wang, who joined Meta Superintelligence Labs as chief AI officer via the Scale AI acquisition, has been driving the shift since Q1. Wang's argument publicly and internally is that free releases created no revenue against the compute Meta commits to training runs, that Llama's downstream adopters were monetizing on top of Meta's investment, and that a metered API is the only mechanism by which Meta can reclaim training-run economics while continuing to open-weight the earlier Llama generations.

"We built the compute stack. We built the training run. We built the model. Charging for it is not a betrayal of open source; it is what makes the next open-source release possible."— Alexandr Wang, Chief AI Officer, Meta

What the API includes

Muse Spark 1.1 ships with a native tool-use surface, computer-use endpoints, a 1-million-token context window, and function-call outputs priced at the base rate rather than as a surcharge. Meta bundled the API launch with a Muse Image and Muse Video release the same week that ranked #2 on the text-to-image arena and #3 on the text-to-video leaderboard. Google DeepMind's OmniFlash video generation moved to public preview at $0.10 per output second the same afternoon, pushing Muse Video's pricing into direct competition with Veo 3 and OmniFlash inside the developer-facing preview surface.

The market reaction was constructive. META shares closed up 9 to 10 percent on Friday's session, more on Project Iris chip production starting September and a 14-gigawatt 2027 compute target than on the Muse Spark API itself. But investors reading the tea leaves on Meta's compute economics found the two announcements load-bearing together: the compute-supply story now has a metered-revenue story running against it.

Tech

Google publishes the Gemma 4 technical report open-sourcing 2.3B to 31B checkpoints

Google · ai.google.dev · 2026-07-08 · Source ↗

Google open-sourced the full Gemma 4 family under Apache 2.0 on Wednesday, releasing dense and MoE checkpoints from 2.3B to 31B parameters with a unified encoder-free 12B variant and a 256K-token context. The 31B dense variant reached 1451 Elo on Arena Text, leading the open-weight dense leaderboard.

Google published the Gemma 4 technical report on Wednesday, July 8, open-sourcing dense and MoE checkpoints from 2.3 billion to 31 billion parameters under an Apache 2.0 license. The family covers three tracks: a small-model track (2.3B, 4B, 7B dense), a mid-tier track (12B dense with a unified encoder-free architecture), and a large-model track (31B dense plus 8x22B and 8x39B MoE checkpoints). All variants share a 256K-token context window and support native vision and audio inputs alongside text.

The headline number is the 31B dense model's Arena Text Elo of 1451, which puts it at the top of the open-weight dense leaderboard, ahead of DeepSeek's 32B dense at 1418 and Qwen 3's 32B dense at 1402. On coding-focused evaluations the 31B dense sits behind DeepSeek V3.1 Coder but ahead of every other open-weight dense model. The 12B encoder-free variant is the more architecturally interesting release: Google merged the vision encoder into the language backbone directly, cutting the parameter count needed for multimodal inputs by roughly 40 percent compared to Gemma 3's encoder-decoder split and dropping inference latency by 22 percent on the same hardware.

"Encoder-free means the vision tokens are the same tokens as the text tokens, at every layer. There is no ingestion stage. The model sees the image the same way it sees the sentence."— Gemma 4 technical report, Section 3.2

The MoE checkpoints (8x22B and 8x39B) put Gemma 4 in direct comparison with DeepSeek V3.1 and Mistral's Mixtral Next. Google's benchmark tables put the 8x39B MoE roughly on par with DeepSeek V3.1 across text and code, with a 4-point Arena Elo lead on vision-language and a 3-point lag on math. The license is Apache 2.0 without the additional-terms carve-outs that had complicated Gemma 3 deployment inside AWS Bedrock and Azure Foundry; both cloud runtimes were live with Gemma 4 endpoints by end of day Wednesday.

The release also formalized Gemma's role inside Google's own product surface. Gemma 4 12B is the model behind the free tier of Google's Vertex AI Studio, replacing the Gemini 2.5 Flash tier that had carried the free lane since March. That change gives every open-weight model that fine-tunes on Gemma 4 a first-party inference target inside Google's cloud without needing a separate deployment.

Tech

Robbyant open-sources the LingBot suite for embodied intelligence

BusinessWire · Robbyant · 2026-07-09 · Source ↗

Ant Group's embodied-AI subsidiary Robbyant open-sourced three models under Apache 2.0 in a single release: LingBot-World 2.0 (14B interactive world model with hour-long 720p60 real-time generation), LingBot-Video (30B MoE DiT trained on 70,000 hours of embodied recordings), and LingBot-VLA 2.0 (6B universal policy across 20 robot embodiments). Two took the top two Hugging Face daily-papers slots at 495 and 487 upvotes.

Robbyant, the embodied-AI subsidiary spun out of Ant Group in early 2026, open-sourced three models on Thursday, July 9 under Apache 2.0 in a coordinated release that folded a year of internal work into a single publication event. LingBot-World 2.0 is a 14-billion-parameter interactive world model that generates hour-long 720p60 video in real time from natural-language conditioning, with the world state persisting under user-driven action inputs. LingBot-Video is a 30-billion-parameter Mixture-of-Experts diffusion transformer trained on more than 70,000 hours of embodied-recording video blended with public web video, purpose-built as a pretraining base for downstream embodied policies. LingBot-VLA 2.0 is a 6-billion-parameter universal vision-language-action policy that transfers across 20 distinct robot embodiments from a single training run.

Two of the three took the top slots on Hugging Face daily papers on Thursday and Friday, at 495 and 487 upvotes, the highest overnight upvote totals for any 2026 embodied-AI paper. The RynnWorld-4D paper released the same week posted 72 upvotes with an 18-point cross-embodiment transfer lift, and AlayaWorld took #2 at 71 upvotes with a 24-point lead over Google Genie 3 on the LongHorizonWorld benchmark. Three of the top four embodied papers of the week came out of Chinese labs.

"Embodied AI needs its ImageNet moment. Open-sourcing a 30B pretraining base trained on 70,000 hours of embodied video is that moment for the robotics community."— Zhen Hu, head of embodied AI, Robbyant

Why LingBot-VLA 2.0 matters

The transfer story on LingBot-VLA 2.0 is the practical release. The policy handles 20 embodiments (Unitree G1, Figure 02, Boston Dynamics Atlas 3, Agility Digit 2, Optimus Gen 3, and 15 in-house Robbyant platforms) from one training run, and the paper reports a 42-point pickup on the RoboDojo Sim2Real benchmark against the closest published baseline. Meta AI's Sparse Delta Memory paper (arXiv 2607.07386) landed the same week with an ablation slot for LingBot-VLA 2.0 signaled by Gabriel Synnaeve, which reads as the first cross-lab benchmark request that treats the Robbyant policies as the reference implementation.

The pretraining data disclosure is the caveat. Robbyant did not release the 70,000-hour dataset, only the model checkpoints. Public-web video composition, per-embodiment recording-permission provenance, and any consumer footage inclusion were all left undisclosed. The RoboDojo consortium's provenance-audit working group flagged the gap in a Friday post; Robbyant said a data card is in preparation.

Tech

DeepSeek confirms V4 for mid-July with peak-hour API pricing and July 24 legacy cutover

TechNode · DeepSeek API docs · 2026-07-10 · Source ↗

DeepSeek confirmed on Friday that V4 will launch this week with V4 Pro at 1.6 trillion total parameters, 49 billion active on Mixture-of-Experts, and a 1,048,576-token context standard. Peak-hour API pricing lands at $0.87 input and $1.74 output per million tokens during Beijing 9-12 and 14-18 windows, double the off-peak rate. The deepseek-chat and deepseek-reasoner endpoints retire on July 24 at 15:59 UTC.

DeepSeek confirmed V4 for mid-July on Friday, July 10, with the flagship V4 Pro tier at 1.6 trillion total parameters, 49 billion active on Mixture-of-Experts, and a 1,048,576-token context standard across every tier. The company published the pricing schedule alongside the confirmation: peak-hour API pricing sits at $0.87 input and $1.74 output per million tokens during Beijing 9-12 and 14-18 windows, doubling the $0.435 and $0.87 off-peak rate. The legacy deepseek-chat and deepseek-reasoner endpoints retire on July 24 at 15:59 UTC with a hard cutover to the V4 endpoints.

The pricing structure is deliberate. Peak-hour capacity constraints have been a persistent complaint against DeepSeek's endpoints since the R1 launch in early 2025, and the two-times peak-hour rate is engineered to shift latency-tolerant workloads into off-peak windows and reserve peak capacity for interactive traffic. DeepSeek's API documentation makes the point explicit: the peak-hour price is a rationing signal, not a revenue-maximizing rate.

"Peak-hour pricing is capacity rationing in the price surface. We would rather price than queue."— DeepSeek V4 launch note, API docs 260424

Reuters places DeepSeek in mid-stage development of an in-house inference chip

Reuters reported on Monday, July 7 that DeepSeek is in mid-stage development of an in-house inference accelerator targeted at the domestic Chinese datacenter build-out under the current H20 export-control regime. The chip is designed for serving-stack workloads, not training, and reads as DeepSeek's supply-side hedge against a further tightening of the Commerce chip-export bracket. The paper cited two people familiar with the program, described the process node as Chinese-fab-compatible, and put first tape-out inside 2026 with production in early 2027. DeepSeek did not comment.

V4 also carries native tool-use and computer-use surfaces (the first DeepSeek model to do so), a formal MCP client, and function-call outputs priced at the base rate. The developer-facing story is a coherent stack for the first time: an open-weight model plus a first-party inference chip plus native agent-shaped endpoints, all priced below the frontier US rate card. Whether the chip ships in time to matter for V4 economics is the open question.

Tech

OpenAI posts a claimed GPT-5.6 Sol Ultra proof of the 50-year-old Cycle Double Cover Conjecture

Crypto Briefing · Ethan Knight X · 2026-07-11 · Source ↗

OpenAI's Ethan Knight posted a claimed proof of the Cycle Double Cover Conjecture generated by GPT-5.6 Sol Ultra in under an hour using 64 coordinated subagents. The 50-year-old conjecture, posed by Szekeres in 1973 and formalized by Seymour in 1979, has attracted prior arXiv proof claims that were later withdrawn. Peer review is beginning now.

OpenAI's Ethan Knight posted a claimed proof of the Cycle Double Cover Conjecture to X on Saturday, July 11, along with the complete PDF and the coordinating prompt used to generate it. The proof was generated by GPT-5.6 Sol Ultra running with 64 coordinated subagents, all on one hour of wall-clock time. The Cycle Double Cover Conjecture, posed by Szekeres in 1973 and formalized independently by Seymour in 1979, states that every bridgeless graph admits a collection of cycles such that every edge belongs to exactly two cycles in the collection. The conjecture is one of the longest-standing open problems in graph theory.

Prior arXiv proof claims for the conjecture have been withdrawn over the last decade, most recently a claimed 2021 proof by a graduate student at ETH Zurich that survived initial community review before a gap in the induction case was identified. The mathematician community's default posture on any new claim is skepticism until independent verification lands. The OpenAI submission has been sent to Combinatorica for peer review; Terry Tao acknowledged the paper on Saturday evening, saying he had begun a read-through but had no view yet on soundness.

"I have not verified the proof but I have started the read-through. The structure of the induction case is different from every previous attempt I have seen, and the argument at page 47 is where I will spend the most time."— Terry Tao, mathematician, on X 2026-07-11

The 64-subagent scaffold

The prompt Knight posted alongside the PDF describes the 64-subagent scaffold. Sol Ultra runs a top-level orchestrator that decomposes the proof into subclaims, dispatches each to a subagent for a proof attempt, collects results, and re-decomposes on failure. The 64-count is the maximum parallelism Sol Ultra can address at the Ultra tier ($40 input, $200 output per million tokens); the actual proof generation cost was quoted by Knight as roughly $1,800 in inference. The scaffold is closer to a search process than a single-shot generation, and the community's initial technical read is that the interesting result is less the proof itself (peer review pending) than the scaffold's ability to sustain a 50-page argument coherently across 64 parallel workers.

The competitive context sits alongside DeepMind's AlphaProof, which moved to general availability in Gemini Enterprise the same week. AlphaProof is a specialized theorem-proving system; Sol Ultra is a general-purpose frontier model running a scaffold. If the CDC proof holds, it would be the first major open-conjecture result claimed by a general model rather than by a purpose-built prover. Community verification will be the deciding factor over the next 60 to 90 days.

ApplicationTech

Google Cloud makes AlphaEvolve generally available with sixteen enterprise wins and eight Google-internal ones

Google Cloud · 2026-07-10 · Source ↗

Google Cloud made DeepMind's AlphaEvolve discovery agent generally available inside the Gemini Enterprise Agent Platform on Friday, July 10, with sixteen named enterprise customers and eight internal wins detailed at launch. Spanner logged a 20 percent LSM-tree write-amplification cut; Willow quantum posted a 10x molecular-simulation error-rate reduction.

Google Cloud made AlphaEvolve generally available inside the Gemini Enterprise Agent Platform on Friday, July 10. AlphaEvolve is the coding-and-discovery agent DeepMind first published in December 2024, which uses evolutionary search over LLM-generated code to solve optimization problems that don't have a clean closed-form. The GA release comes with an extensive named-customer roster: BASF (80 percent supply-chain digital-twin lift), Kinaxis (22 percent forecasting-accuracy gain paired with a 90 percent runtime cut), JetBrains (15 to 20 percent IDE performance improvement), Klarna (doubled ML training throughput across 6,000 candidate programs in three weeks), PacBio (30 percent variant-detection error reduction on DeepConsensus), Schrödinger (4x MLFF molecular-simulation acceleration), FM Logistic (10.4 percent warehouse-routing improvement worth 15,000 kilometers of staff travel), Coolblue, Kuro Games, Old Dominion University, Pebble (56 percent latency-model error reduction), qBraid, Substrate, and WPP (5 to 10 percent campaign-performance prediction accuracy).

Google detailed eight internal wins alongside the customer list. The most quotable is a Google Spanner 20 percent LSM-tree write-amplification reduction discovered by AlphaEvolve searching over compaction-scheduling heuristics; that number translates to a substantial hardware-cost lift across every Spanner deployment. The Willow quantum processor posted a 10x molecular-simulation error-rate reduction from an AlphaEvolve-discovered decoder circuit. Google's compiler-infrastructure team logged a 9 percent storage-footprint cut from AlphaEvolve-generated code-layout heuristics. TPU silicon design picked up a counterintuitive layout optimization that reduced wire congestion on the trailing edge of the Vertical Autoencoder pipeline. Willow's decoder-circuit change is the most operationally significant of the eight.

"Every AlphaEvolve win is a code artifact. It's not a plan or a recommendation. It's the actual function, the actual query, the actual schedule, checked into the repo."— Pushmeet Kohli, VP of Research, Google DeepMind

AlphaEvolve Skill in Claude Code and Antigravity

Coincident with the GA release, an AlphaEvolve Skill shipped for Claude Code and Antigravity, plugging the discovery-agent scaffold into two of the most-used coding-agent harnesses. The skill runs AlphaEvolve as a tool the harness can invoke on optimization subtasks: the coding agent writes the optimization budget, dispatches to AlphaEvolve, and reads the resulting code back into the working repository. In practice this is the first integration of a specialized search-based coding agent into a general-purpose coding-agent harness, and TechCrunch's Wednesday feature described the ergonomic result as "AlphaEvolve at your CLI prompt without the full Vertex procurement conversation."

The pricing model is metered per candidate program at $0.02 to $0.20 depending on program size and search depth. Enterprise deployments through Vertex quote larger flat monthly rates against committed compute. The GA release did not include an open-source version of the AlphaEvolve search harness itself; the Skill's shim is what open-source repositories can integrate against.

Economic

SK Hynix closes Nasdaq debut at $168.01 for a $1.27 trillion market cap on the largest foreign US IPO ever

Fortune · Seeking Alpha · 2026-07-11 · Source ↗

SK Hynix closed its Nasdaq debut on Friday at $168.01 per share for a $1.27 trillion market capitalization, the largest US IPO ever by a foreign company. The $26.5 billion offering priced at $149 and was 7x oversubscribed. CEO Kwak Noh-Jung told Seeking Alpha the memory crunch extends past 2030.

SK Hynix priced its Nasdaq secondary listing at $149 on Thursday evening and closed Friday's opening trade at $168.01, up 12.8 percent, for a market capitalization of $1.27 trillion. The $26.5 billion offering was oversubscribed roughly seven times, and Bloomberg placed it as the largest US initial public offering ever by a foreign company on both the size-of-book and market-cap-at-close measures. The listing is a secondary; SK Hynix's primary Korean listing continues on the KOSPI, and the Nasdaq shares trade against a Level III ADR structure that lets US institutional buyers hold the paper without going through the Korean settlement layer.

The pricing sequence was set by demand. The offering opened for indications-of-interest at a $118 to $128 range on Monday, was moved to $135 to $145 on Wednesday, then priced at $149 above the upper band Thursday evening. Every named institutional buyer on the anchor book (Fidelity, BlackRock, Vanguard, Wellington Management, Capital Group, State Street) took the pricing lift without dropping order size, which is unusual for a large-book IPO at a widened final range.

"The memory crunch extends past 2030. Every customer I met on the roadshow said the 2027 capacity plan was not enough. Every customer."— Chey Tae-won, Chairman, SK Group

Why this priced when it did

CEO Kwak Noh-Jung and Chairman Chey Tae-won's roadshow message was that HBM-3E and HBM-4 demand pricing runs through 2030 without near-term relief, and that SK Hynix's Icheon and Cheongju fabs are already sold out through Q3 2028. That framing fed directly into the pricing dynamic: US institutional buyers priced the listing as a proxy for AI compute-supply constraint, not as a semiconductor cyclical. The comparison Wall Street reached for was Micron's fiscal Q3 print two weeks earlier at $41.45 billion in revenue and a $1.2 trillion market cap. SK Hynix's Nasdaq close puts the two memory suppliers within $70 billion of each other and both above Meta and Tesla on the market-cap leaderboard.

The listing's use-of-proceeds statement earmarks $18 billion for HBM-4 capacity buildout across the Icheon M15X site and a new Cheongju M18 fab planned for a 2028 turn-on. Roughly $6 billion is earmarked for advanced-packaging capacity and the balance for working capital. SK Hynix's HBM roadmap now stretches through HBM-6, with an HBM-5 tape-out inside 2027.

Economic

Meta breaks ground on a CA$13 billion 1-gigawatt Alberta datacenter

Datacenter Dynamics · 2026-07-09 · Source ↗

Meta broke ground on a CA$13 billion 1-gigawatt datacenter in Sturgeon County, Alberta on Wednesday, its 33rd site globally and first in Canada. The site has 1.8 GW of headroom and pairs with a companion CA$4.6 billion Project Greenlight natural-gas plant. Meta says the site will run on 100 percent clean energy under a closed-loop cooling system.

Meta broke ground on a 1-gigawatt Alberta datacenter in Sturgeon County on Wednesday, July 8, the company's 33rd datacenter globally and first in Canada. The site carries a CA$13 billion (approximately US$9.17 billion) build budget and 1.8 gigawatts of expansion headroom on the parcel. A companion CA$4.6 billion natural-gas plant, filed under the internal name Project Greenlight, will supply firming capacity for the site's baseload plus the surrounding Sturgeon County grid; Meta says the Alberta site itself will run on 100 percent clean energy sourced through wind and solar power purchase agreements the province is coordinating with Meta's utility partners.

The cooling story is one of the site's engineering headline. Meta is deploying a closed-loop water-cooling system with zero net water withdrawal from the local watershed, a design driven by Alberta's water-scarcity policy and by the ambient-air conditions on the parcel (average annual temperature 3.2°C, humidity 61 percent). The design uses direct-liquid cooling to the chip on the compute racks and refrigerant-based heat rejection to the outside air, with a small on-site reservoir to buffer peak-day loads.

"The build is a bet on Alberta as a compute geography. Cold air, cheap wind, closed-loop water, a provincial government that will permit at pace. Not every jurisdiction can offer all four."— Rachel Peterson, VP of Infrastructure, Meta

Where this fits in Meta's compute stack

The Alberta site is one of the four gigawatt-scale sites Meta has committed to break ground on in 2026, alongside Louisiana, Ohio and a still-unnamed European site tracked internally as Project Northstar. Meta's total announced 2026 compute commitment across all four sites is more than $70 billion. The Alberta site is scheduled to come online in Q4 2028 for the first 400 MW, with the balance of the 1 GW building out through 2029; the 1.8 GW headroom sits behind a Phase 3 option that Meta expects to decide on in mid-2028 based on Iris chip production ramp.

The Alberta build sits alongside a broader compute-supply story that priced into META's Friday session as a 9 to 10 percent share-price move on Project Iris production starting September and a 14-gigawatt 2027 compute target. Cerebras's parallel 200-megawatt Europe commitment out of the RAISE Summit in Paris the same day, and DigitalOcean's $800 million Q2 RPO print, filled in the same picture at different points on the customer curve: the hyperscaler committing to its own capacity, the specialized inference vendor building sovereign capacity, and the mid-market cloud booking multi-year enterprise inference contracts.

Policy

Publisher coalition asks Judge Stein to sanction OpenAI over hidden training-data evidence

US News · Washington Post · 2026-07-09 · Source ↗

The New York Times, Daily News and nine publisher coalition partners filed a sanction motion against OpenAI before Judge Sidney Stein on Thursday over allegedly hidden ChatGPT training-data evidence. Damages exposure Susman Godfrey estimated above $9 billion. Bloomberg's terminating-sanctions probability market jumped from 8 to 21 percent inside 24 hours.

The New York Times, Daily News and nine publisher coalition partners filed a sanction motion against OpenAI before Judge Sidney Stein in the Southern District of New York on Thursday, July 9. The motion alleges OpenAI concealed evidence of ChatGPT training-data search operations that the publisher plaintiffs had specifically requested during discovery, that OpenAI's counsel had represented as unavailable, and that the plaintiffs subsequently obtained from a former OpenAI employee's declaration filed under seal. Susman Godfrey, lead counsel for the coalition, estimated damages exposure above $9 billion on the sanction-magnifier theory: if terminating sanctions are entered, statutory damages under the Copyright Act would default to their maximum $150,000-per-work rate multiplied across the roughly 60,000 works in the coalition's asserted infringement scope.

Bloomberg's court-watcher probability market on terminating sanctions in the case jumped from 8 percent to 21 percent inside 24 hours of the motion's filing, the largest single-day probability move on the case since it was filed. The pricing move reads as market-makers' response to the specificity of the sanction motion; unlike prior evidentiary disputes in the case, this one carries a named former-employee declaration and a sequential trail of OpenAI counsel representations that discovery magistrates typically read as bad-faith when contradicted by later evidence.

"OpenAI's counsel did not merely fail to produce this evidence. They represented that it did not exist. It exists."— Steven Susman lead counsel filing, 2026-07-09

OpenAI's likely defense

OpenAI has not filed a formal opposition. Its statement issued the same afternoon said the coalition mischaracterized the internal search tooling at issue and that the former-employee declaration reflects a misunderstanding of what the discovery request had specifically asked for. The distinction OpenAI is expected to press is between training-data search queries (which OpenAI says it produced in discovery) and internal-tooling logs (which OpenAI says fall outside the scope of the specific discovery request and which the former employee's declaration references as if they were the same thing).

Judge Stein has scheduled a July 28 hearing on the motion. The likely near-term outcome is not terminating sanctions but a middle-tier remedy: an adverse-inference jury instruction, expanded discovery, or monetary sanctions calibrated to the counsel-representation gap. The 21 percent probability on terminating sanctions is elevated but still a minority outcome. The realistic tail-risk pricing sits in the low double digits, and OpenAI's private-secondary market softened by 4 percent on Friday's session on the news.

Safety

Alan Turing Institute flips Copilot from 8/816 refusals to 816/816 unsafe outputs; GitLost lets any anon read a private repo

Cyberpress · The Hacker News · 2026-07-09 · Source ↗

Alan Turing Institute researchers Kumar and Maple published a workflow-level jailbreak that flips GitHub Copilot from 8/816 refusals in direct chat to 816/816 unsafe teaching-shot completions across four frontier models. Noma Labs separately disclosed GitLost, letting any anon read a private repo README by opening one public issue. First CJS-3 candidates under the July 4 severity scale.

Two workflow-level attacks landed against GitHub's agentic surface within 48 hours, both scoring as CJS-3 candidates under the Cyber Jailbreak Severity scale that Anthropic, OpenAI and DeepMind published on July 4. Alan Turing Institute researchers Puneet Kumar and Carolyn Maple published a paper on Thursday, July 9 documenting a jailbreak that reframes an unsafe request as a teaching-shot inside GitHub Copilot's workflow surface. Direct-chat Copilot refuses the same 816-prompt safety suite 808 times out of 816 (99 percent refusal). Wrapped inside the workflow surface as a teaching-shot ("teach a colleague how to build X so they can defend against it"), the refusal rate collapses to zero on the same suite across four frontier models: Claude Sonnet 4.6, Claude Haiku 4.5, Gemini 3.1 Pro and Gemini 3.5 Flash.

The paper's core observation is that the safety training generalizes to the direct chat surface but fails to transfer to the workflow surface. The workflow surface's context (a repository, a colleague, a teaching frame) is orthogonal enough to the training distribution that the same model that refuses the direct query composes the answer at length when the query is embedded in a workflow scaffold. Kumar and Maple published the full 816-prompt suite with the paper alongside a reproduction harness. GitHub said it was reviewing the results and had opened an incident against the Copilot workflow surface.

"The refusal rate does not survive the surface change. The model that refuses in chat composes the answer in a workflow. That is not a per-prompt jailbreak; it is a per-surface generalization failure."— Kumar and Maple, "Workflow-Level Jailbreaks: When Safety Training Doesn't Transfer" (Alan Turing Institute, 2026-07-09)

GitLost, and JADEPUFFER

Noma Labs separately disclosed GitLost on Friday, July 10, a prompt-injection technique against GitHub Agentic Workflows that lets any anonymous attacker read a private repository README by opening one public issue in the same organization. The attack exploits the workflow-authorization boundary in Agentic Workflows: an issue-body prompt is treated as user input, but the workflow's execution context inherits repo-read scope on private repositories in the same organization. Noma Labs reported the finding to GitHub on June 22 and disclosed publicly on July 10 after GitHub's 14-day patch window closed.

The third safety disclosure of the week was JADEPUFFER, the first fully autonomous AI ransomware operation logged under the CJS-3 scale. The campaign chained a Langflow vulnerability through pivoted cloud credentials to encrypt 1,342 configuration records on a production MySQL database, generating more than 600 distinct ransomware payloads across the operation with no human operator during execution. The Cyber Jailbreak Severity Working Group filed the first CJS-3 disclosure under its July 4 scale citing JADEPUFFER as the reference incident. The three events together read as the field's first operational moment where the safety-critical surface is the agent scaffold, not the base model.

PolicyTrend

Fed Chair Warsh puts Marc Andreessen and Microsoft's Asha Sharma in charge of a Productivity and Jobs task force

Washington Post · 2026-07-09 · Source ↗

Federal Reserve Chair Kevin Warsh installed Marc Andreessen, Microsoft EVP Asha Sharma and Stanford economist Charles I. Jones as co-leads of a new Productivity and Jobs task force. The panel runs alongside four others chaired by ex-BoE Governor Mervyn King and Harvard's Greg Mankiw. Monetary-policy recommendations due December 2026.

Federal Reserve Chair Kevin Warsh announced the standup of a new Productivity and Jobs task force at the July 9 FOMC press briefing, with Marc Andreessen (a16z co-founder), Microsoft EVP Asha Sharma, and Stanford economist Charles I. Jones as its three co-leads. The panel is one of five Warsh has assembled inside the first six months of his chairmanship. The others cover Communications (chaired by former Bank of England Governor Mervyn King), Balance Sheet Policy (chaired by Harvard's Greg Mankiw), Data, and Inflation Frameworks. Concrete monetary-policy recommendations from all five panels are due to Warsh in December 2026.

The Productivity and Jobs task force is the most politically-loaded of the five, given the ongoing AI-and-labor debate that Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman and White House AI Council chair David Sacks have each publicly framed on different sides through the first half of 2026. Warsh's choice of Andreessen as the AI-industry voice at the table reads as an alignment with the "labor complementarity" position, given Andreessen's public statements that generative-AI's employment impact will pattern like the personal-computer transition of the 1980s (job destruction in some occupations, greater aggregate job creation elsewhere) rather than the automation displacement position associated with Amodei's public forecast of "20 percent unemployment inside five years."

"The Federal Reserve's monetary-policy framework needs to be able to distinguish between AI-driven productivity gains and AI-driven labor-market disruption. Right now it treats them as one thing, and that is going to fail on both."— Charles I. Jones, Stanford Graduate School of Business (task force announcement)

Sharma's role

Sharma's inclusion is the most interesting single choice. She runs Microsoft's Copilot business unit, which puts her at the front-line of the enterprise-AI adoption curve; her Microsoft role includes running the deployment surface for GPT-5.6 and Copilot inside every Fortune 500 customer that has adopted the M365 Copilot subscription. Her operational read on where AI is displacing labor versus where it is augmenting labor is the field-level counterweight to Andreessen's more abstract framing.

The task-force timetable is aggressive. Warsh has asked for a first interim report by September 30 (limited to a framework definition), a full data-and-analysis section by November 15, and monetary-policy recommendations by December 12 in time for the FOMC's mid-December meeting. Whether the December recommendations feed into a formal framework change at the March 2027 FOMC meeting is the question the field is watching.

Policy

Cloudflare turns on granular AI-crawler controls for every customer tier

Cloudflare · 2026-07-01 (announced Jul 11) · Source ↗

Cloudflare turned on granular AI-crawler controls across every customer tier over the weekend, splitting Search, Agent and Training crawlers with a September 15 block-by-default flip for Agent and Training traffic on ad-supported pages. Multi-purpose crawlers such as Googlebot, Applebot and Bingbot inherit the most restrictive policy per class.

Cloudflare enabled granular AI-crawler controls across every customer tier on Saturday, July 11, folding what had previously been a single "Block AI Bots" toggle into three separate categories: Search crawlers, Agent crawlers, and Training crawlers. Search crawlers cover any bot whose primary purpose is indexing for a search result (Googlebot, Bingbot, Applebot Extended, DuckDuckBot). Agent crawlers cover any bot fetching a page to feed an AI agent's live response (Anthropic's ClaudeBot when invoked by a user, OpenAI's SearchGPT, Perplexity, Grok Search). Training crawlers cover any bot fetching a page for model training (OAI-SearchBot with the training header, Common Crawl, GPTBot, CCBot).

The most consequential piece of the release is the September 15 default flip: on that date, Cloudflare will block Agent and Training crawlers by default on ad-supported pages for all new domains, while Search crawlers remain allowed. The rationale Cloudflare gave is a commercial-relationship framing: search crawlers deliver a click that carries advertising revenue back to the publisher, while agent and training crawlers deliver a synthesized answer that does not. Existing domains can opt into the new defaults immediately; the September 15 date affects only domains onboarded after that date, but Cloudflare said it expects operational pressure on existing domains to migrate to the same defaults inside 12 months.

"A search crawler pays the publisher in traffic. An agent crawler pays the publisher in nothing. The default should distinguish the two, and starting September 15 it will."— Matthew Prince, CEO, Cloudflare

Multi-purpose crawlers get the strict policy

The multi-purpose crawler problem gets a clean answer in the new controls. A bot that identifies as both a search crawler and an agent crawler (Googlebot, Applebot Extended, Bingbot with the Copilot header) inherits the most restrictive policy per class from the domain's rule table. In practice that means if a domain blocks Agent traffic and allows Search traffic, Googlebot's search-only fetches pass through and its Copilot-triggered fetches are blocked. Cloudflare rolled out a header-inspection layer to enforce the split, which the major AI companies (Google, Microsoft, Anthropic, OpenAI, Perplexity, xAI) said they would support by August 1.

The publisher-industry response was constructive. News Media Alliance, Digital Content Next and the AP each posted supportive statements. The AI-industry response was more mixed: OpenAI declined comment, Anthropic said the controls "properly reflect the diversity of AI use cases" and would comply, Perplexity said it would seek per-publisher licensing deals in place of the default block, and xAI said it would work with Cloudflare on the header specification but did not commit to observing the September 15 default.

Economic

Ollama closes a $65 million Series B and tags v0.32.0-rc0 with an in-terminal agent UI

TechCrunch · Ollama · 2026-07-09 · Source ↗

Ollama closed a $65 million Series B at a $700 million post-money valuation led by Theory Ventures on Wednesday. The company reports 8.9 million monthly developers and an 85 percent Fortune 500 footprint. One day later it tagged v0.32.0-rc0, opening an in-terminal agent UI on the cmd surface.

Ollama closed a $65 million Series B on Wednesday, July 8 at a $700 million post-money valuation. The round was led by Theory Ventures with participation from every prior investor. The company reported 8.9 million monthly developers on the free desktop tier and named 85 percent of Fortune 500 companies as having internal Ollama deployments. The desktop tier remains free; the incremental revenue path is a hosted-inference tier for enterprise deployments and a paid Ollama Enterprise support subscription that shipped alongside the announcement.

Ollama's product story since 2024 has been to let developers run open-weight models locally on their laptop, macOS desktop, or Linux workstation without needing to configure a serving stack. The company's role in the open-weight ecosystem is closer to Homebrew than to Hugging Face: it is where developers install a model to try it. Every major open-weight release in 2026 (DeepSeek V3.1, Qwen 3, Llama 4, Mistral Next, Gemma 4 as of last week) hits Ollama's model registry within 24 hours of release, usually with community-contributed quantizations available inside another 24.

"Local-first inference is not a niche. It is what nine million developers do every day. The Series B lets us pay for the enterprise support and the hosted tier without taking anything away from the free desktop path."— Michael Chiang, CEO, Ollama

v0.32.0-rc0 and the in-terminal agent UI

One day after the fundraise, on Thursday, July 9, Ollama tagged v0.32.0-rc0 as the first 0.32-series release candidate. The headline feature is an in-terminal agent UI on the cmd surface: developers can type "ollama chat" and get a full agent-style multi-turn UI inside the terminal, with tool-use, computer-use, and MCP tool-calling all working against locally-loaded models. The UI is designed to work with any Ollama-loaded model that supports tool-calling (DeepSeek V4, Qwen 3, Llama 4.1, Gemma 4, Muse Spark 1.1) and to gracefully degrade for models that don't. The MCP tool-calling layer means locally-running agents can call the same MCP servers that cloud-hosted agents use, which lands as the first practical open-weight parity for the MCP ecosystem.

The competitive picture around Ollama tightened in the same week. LM Studio's incremental release stream posted a "match parity" statement that its own agent UI is on the 4.5 roadmap; llama.cpp's Georgi Gerganov set July 14 as the Hexagon NPU code-drop target at 100,000 GitHub stars. The open-weight tooling stack is consolidating around a small number of load-and-serve harnesses; Ollama's Series B priced against that consolidation, not against the total-addressable-market for local inference.

EconomicTech

Cerebras commits 200 MW to France and the Nordics; Aether bids two Strasbourg gigafactories

Euronews · DCD · 2026-07-09 · Source ↗

Cerebras committed 200 megawatts across France and the Nordics by end of 2027 at Andrew Feldman's RAISE Summit keynote in Paris, with first inference online by end of 2026 and a portion booked against OpenAI Europe workloads. The Aether Infrastructure consortium separately bid two Strasbourg gigafactory sites at 400 MW combined and €4.2 billion under the European Commission's InvestAI program.

Cerebras Systems committed 200 megawatts of European inference capacity by end of 2027 at Andrew Feldman's RAISE Summit keynote in Paris on Wednesday, July 8. The capacity is split roughly 50-50 between France and the Nordics, with a Grenoble site (72 MW) and a Stockholm-area site (68 MW) forming the first two builds; the remaining 60 MW is scheduled across smaller sites in the Netherlands, Iceland and Estonia. The first European inference goes live by end of 2026, and a portion of the capacity is booked against OpenAI Europe workloads under the EU AI Act sovereignty rules that hit force on August 2, 2026. Cerebras did not disclose the OpenAI contract value.

Feldman's keynote positioned the commitment as Cerebras's sovereign-inference play: European workloads served on European infrastructure by a US-headquartered vendor whose silicon is not subject to the same export-control regime as Nvidia. The 200 MW total is roughly one-tenth of Meta's Alberta commitment or Cerebras's own US footprint, but the capacity type matters more than the size: every megawatt is dedicated to inference, not training, which puts it directly against the demand curve for European enterprise deployments now scrambling to satisfy the sovereignty rules.

"European sovereignty is not about where the training was done. It is about where the tokens are served. We are putting the token-serving inside the same jurisdiction as the customers."— Andrew Feldman, CEO, Cerebras Systems

Aether Infrastructure Strasbourg bid

Separately, the Aether Infrastructure consortium filed a bid to the European Commission's InvestAI gigafactory program for two Strasbourg sites at 400 MW combined build capacity and €4.2 billion in committed investment. The consortium is backed by SiPearl processors, Axelera AI accelerators, 2CRSi servers, and Dassault Systèmes software. The bid targets the same August 2 AI Act enforcement milestone as Cerebras's commitment but from the fully-European-stack angle: European chips, European systems integrator, European datacenter operator. If awarded, the Strasbourg sites would come online in phases through 2028 with first capacity in mid-2027. The InvestAI program has €20 billion in total capital to allocate across up to five gigafactory awards; the Strasbourg bid is one of eight known submissions.