AI News: September 9, 2026

1. Mistral Raised €3 Billion at a €21 Billion Valuation in Europe’s Largest Tech Round

Mistral AI. The Series D is the biggest equity raise ever completed by a European technology company, three years after the lab’s founding. Samsung Electronics led, co-led by the EQT-managed Scaleup Europe Fund and existing investor PSG Equity, with a16z, Nvidia, Salesforce Ventures, Advent, BlackRock, and Luxembourg also participating. Mistral says the money goes to frontier research, training compute, and international expansion; it currently operates in 20 countries and lists more than 125 enterprise customers including Airbus, ASML, and HSBC. Samsung leading a French sovereign-AI champion is the detail worth sitting with, given that the pitch is independence from foreign infrastructure. Source

2. Cognition Raised $2 Billion at $48 Billion as Its Run-Rate Nearly Doubled in Four Months

Cognition. Andreessen Horowitz, Accel, Founders Fund, General Catalyst, and Avenir backed the round. Annualized run-rate revenue went from $492 million in May 2026 to $900 million now, with the company projecting $4 billion to $5 billion by the end of 2026. For comparison, Cursor was discussing a $50 billion valuation in April 2026 on more than $2 billion of annualized revenue before SpaceX acquired it for $60 billion, which means Cognition is currently carrying a higher revenue multiple than Cursor did. Cursor’s sale was driven partly by compute constraints, and whether Cognition hits the same wall is the open question behind the number. Source

3. Infostealer Malware Is Draining Claude Token Allowances Through Hijacked Sessions

Anthropic. Attackers are using commodity infostealer malware to lift saved Claude login sessions off users’ machines, then minting Claude Code OAuth tokens from those credentials for persistent access. Reported cases include an account burning its full daily token allowance for three days while the owner was inactive, usage jumping from 0% to 49% in twelve minutes, and unauthorized plan upgrades with the charges attached. Anthropic emailed affected users confirming the malware came from outside Claude, and declined to say what protections users should apply. The compounding problem is that itemized usage tracking does not exist, so victims cannot see which requests were not theirs and theft runs undetected for days. If you use Claude Code, rotating credentials and checking usage is cheap insurance. Source

4. An NYU Mathematician Says OpenAI Pressured Him to Drop His Anthropic Co-Author

Research integrity. Tristan Buckmaster of NYU says that after rumors of his Navier-Stokes progress with co-author Levent Alpöge reached OpenAI, researcher Sébastien Bubeck pressed him to remove Alpöge from the paper because Alpöge works at Anthropic, and asked “Why would you ruin your career?” when he refused. OpenAI then published its own Navier-Stokes result using what Buckmaster describes as the same solution approach. The provenance question is sharp because Buckmaster had shared drafts through Codex; OpenAI says it did not see the work through any means before public release, acknowledges training on data postdating the mathematicians’ breakthrough, and says the two proofs differ significantly. Nothing here is settled, but it is the first high-profile case where “did the assistant leak my draft” is a serious question rather than a hypothetical. Source

5. China Set a 2030 Target of 9,800 Exaflops of AI Compute Built on Domestic Chips

Policy. The Ministry of Industry and Information Technology published a 2026 to 2030 plan to more than quadruple national intelligent computing capacity, from 2,185 exaflops in June, itself up 177% year over year, to 9,800 exaflops by 2030. It commits 3.8 trillion yuan, about $532 billion, in cumulative information infrastructure investment and calls for orderly deployment of clusters in the 10,000-card and 100,000-card ranges plus inference facilities tuned to specific applications. China has already built 52 intelligent computing facilities with more than 10,000 accelerator cards each. The plan is explicitly premised on domestic silicon as US export controls tighten. Source

6. ASML, TSMC, Samsung, and Intel Committed to 12-Inch Photomasks for High NA EUV

Semiconductors. ASML and TSMC announced an industry transition from the 6-inch photomask format to 12-inch for High NA EUV lithography, with Intel Foundry and Samsung signing on. The roadmap targets a 12-inch mask pilot line by 2031 and production readiness at advanced nodes by 2033. The reason this is AI news: at current mask sizes, large AI accelerators must be split across multiple exposures and stitched together, which caps die size and hurts yield. Samsung also plans High NA EUV DRAM by 2028 and TSMC will deploy it in advanced nodes from 2030. The dates are the point, since they set a supply-side ceiling on accelerator die area for the next several years. Source

7. Alibaba Cloud and Cambricon Took PyTorch Foundation Board Seats

Open source. The PyTorch Foundation announced at PyTorch Conference China in Shanghai that Alibaba Cloud and Cambricon joined as Platinum members, each taking a Governing Board seat and a Technical Advisory Council seat, with Ant Group joining at Gold. Cambricon’s keynote covers hardening PyTorch’s device-agnostic foundation so any backend can offer native capabilities, which is the concrete stake: Chinese accelerator vendors are buying influence over the upstream abstraction layer rather than maintaining forks. Huawei, a member since 2023, helped start the Accelerator Integration Working Group and got Ascend NPU accepted as the first additional platform in PyTorch. More than 250 organizations in China now contribute to Foundation projects including vLLM, DeepSpeed, Ray, and Safetensors. Source

8. Argentine Patagonia Is Being Scoped for Gigawatt-Scale AI Data Centers

Infrastructure. Pampa Energía is seeking investors for a 500 MW facility in Neuquén province with initial contracts expected by the end of 2026, and Poland’s Green Capital is starting at 300 MW in Chubut with a 3,000 MW long-term target. OpenAI announced a project with Sur Energy in October that has not been signed. The draw is cool climate, hydropower, wind, and shale gas from Vaca Muerta, plus roughly $900 million in required grid buildout. The distinguishing feature versus US sites is the absence of organized local opposition so far, which the reporting attributes to there being almost nothing built yet rather than to settled consent; Mapuche communities near the Loma de la Lata plant have confronted energy companies before, and grid connections, undersea cables, and the 2027 presidential election all remain open variables. Source