OpenAI Publishes a Navier–Stokes Proof From a Model It Will Not Sell
OpenAI posted a machine-produced resolution of the Navier–Stokes Millennium Prize problem on 8 September, and the notable part is not the mathematics. Every figure is OpenAI's own and unaudited: the work came from an internal model it describes as “significantly more capable than GPT-6 Astra”, in training since 28 August; roughly 10,000 concurrent agents ran in the winning group; the solution landed on 5 September, about 88 hours after launch, with a Lean formalisation adding another 17 hours through GPT-6 Astra. Across all problems attempted, OpenAI counts 4.9 million agent messages and around 300 billion output tokens — 2.7 million and roughly 130 billion for Navier–Stokes alone. The company does not claim the prize, and credits Alpöge and Buckmaster with priority on the forced Euler result.
The Lean proof is genuinely public, on GitHub, which is the one part a sceptic can check without trusting OpenAI. The model is not. For anyone buying inference, that is the point: frontier capability is now being demonstrated on systems that are not on any price list, at a compute scale no customer will reproduce. A vendor's most impressive result increasingly says nothing about what you can procure, and benchmarking the models actually available to you is the only evidence that survives a procurement review.
A 124B Open-Weight Multimodal Model Under MIT — and the Hardware Bill That Comes With It
Ant Group's inclusionAI has Ling-3.0-flash-VL live on Hugging Face — a 124B mixture-of-experts model with 5.5B active parameters, 1M context and native vision, under a plain MIT licence. The repository was created 4 September and finalised on the 8th, with an FP8 variant the same day. The weights are actually there: 64 safetensors shards, about 249.7 GB. Architecture is a 42-layer hybrid attention stack at a 5:1 ratio with VideoRoPE and a ViT projector. The performance claims are Ant's own, measured against an older revision of Artificial Analysis's index than the one published this week.
The honest caveat is in the serving recipe rather than the benchmark table. The reference SGLang launch configuration wants four 141GB-class accelerators (H200 or H20-3e), or a four-GPU Blackwell node (B300 or GB300) — so “downloadable under MIT” and “runnable inside your own perimeter” are separated by a hardware order with a long lead time. For an EU buyer, that gap is the whole decision: a permissive licence removes the legal obstacle to keeping a multimodal agentic model inside European jurisdiction, and does nothing about the four GPUs.
Europe's One Real Chokepoint Just Set the Clock to the 2030s
ASML and TSMC announced on 8 September an industry initiative to transition to 12-inch photomasks for High NA EUV lithography. No other participants are named: the two companies say only that other High NA EUV adopters and major suppliers have expressed their interest in joining. The dates they give: a pilot line by 2031, full system readiness for advanced-node production by 2033, with TSMC's High NA high-volume manufacturing starting from 2030. TSMC's own release carries the same timeline.
Set that against the compute assumptions in most sovereignty business cases. ASML is the only EU-headquartered company that is a genuine bottleneck in the AI supply chain, and its roadmap for the next lithography step now runs into the next decade. Anyone modelling a five- to ten-year fall in the price of European inference capacity is implicitly betting on process nodes that the equipment makers are scheduling for 2030 and beyond.
Quick Hits
- Artificial Analysis ships Intelligence Index v4.3, its third revision in five weeks — Claude Fable 5.1 and GPT-6 Astra tie at 53, Opus 5 at 51, and GLM-5.3 Flash — the third-strongest open-weights model, behind GLM-5.3 and Kimi K3 — at 42. The line worth keeping: GLM-5.3-Flash matches GPT-5.6 Terra's score of 42 at $0.25 per index task against $1.40. Private test sets now carry 45% of index weight, up from 40%. (artificialanalysis.ai)
- CADA is still a proposal, not law. The Council's Working Party on Telecommunications met on the Cloud and AI Development Act on 8 September, working from the Commission's 4 June 2026 proposal (doc 10104/26), with further sessions through the month. Widely repeated claims that CADA was adopted in June, entered into force on 4 August, or has a sovereignty tier applying in February 2028 appear in no Council or EUR-Lex record. (consilium.europa.eu)
- Anthropic reportedly walked away from a roughly $6bn acquisition of Decart — an Israeli inference- and training-efficiency startup — after due diligence. Bloomberg attributes this to people familiar with the matter on 8 September; neither company has confirmed it, and no reason has been established. (bloomberg.com)
- China's MIIT published its five-year ICT plan on 7 September — targeting 9,800 EFLOPS of “intelligent computing capacity” by 2030 against 2,185 EFLOPS at end-June 2026, with ¥3.8trn of information-infrastructure investment across 2026–2030. The metric is a Chinese-defined aggregate and is not normalised for precision, so it does not compare directly with Western EFLOPS figures. (english.scio.gov.cn)
- inclusionAI also published LLaDA2.2-mini on 5 September — an Apache-2.0 diffusion language model with DELETE and INSERT control tokens that let decoding edit sequence structure mid-generation. Its own figures show function calling on BFCL v4 rising from 28.44 to 47.68 and LongBench v2 from 12.13 to 34.99 — alongside regressions on AIME 2026 (40.37 to 35.05) and IFBench (31.60 to 24.93). (huggingface.co)