Up North AIUp North
Back to news

Washington Tightens the Chip Spigot, Again

GPT-6 Astra Co-Authors a Candidate Lattice Definition of the Standard Model. Yann LeCun Bets on Federated, Sovereign Open Models via Project Tapestry.

Share

Washington Tightens the Chip Spigot, Again

The US Commerce Department issued an interim final rule today closing a loophole that let Chinese firms access Nvidia-class compute remotely through third countries like Singapore and Malaysia [1][2]. The rule requires licenses for exports, reexports, or transfers to any data center worldwide that might route compute back to China or other D:5 embargoed countries, and it targets clusters over 12 million TPP specifically. It also repeals parts of the Biden-era AI Diffusion Rule and tightens foundry due diligence [1].

This is Washington admitting the first version of chip controls didn't work. Companies found the seams — a data center in Kuala Lumpur, a reseller in Singapore — and routed around them. Today's rule is an attempt to plug those seams directly, with a 60-day comment period before it locks in.

Separately, a senator is publicly accusing Commerce of failing to actually stem the flow of chips to China [3] — which tells you this isn't a victory lap, it's damage control. For any company building on GPU infrastructure with cross-border dependencies, the message is clear: assume compute geography gets more scrutinized, not less, and build your sourcing accordingly.

GPT-6 Astra Co-Authors a Candidate Lattice Definition of the Standard Model

OpenAI launched a science blog, and its first post is a genuine physics contribution. UC Santa Barbara physicist Nathaniel Craig, working as a visiting scientist at OpenAI, used GPT-6 Astra to help develop a candidate nonperturbative lattice definition of the Standard Model using overlap fermions and gauge-invariant measures for chiral gauge theories [1][2]. The work builds on an October preprint and was reviewed by established lattice theorists — Golterman, Kronfeld, Shamir — before publication [3].

What's notable isn't just the physics (which the team is careful to frame as a candidate step, not a solved problem for practical simulation) — it's the workflow. Astra reportedly took Craig's strategy and turned it into an actual construction, leading detailed proof work with Lean formalization baked in [2]. That's not autocomplete. That's an AI functioning as a junior collaborator capable of carrying a proof to completion under human direction.

This is the clearest signal yet that "AI-assisted research" is graduating from literature summarization to actual theorem construction in hard sciences. If this pattern holds, the bottleneck in frontier physics and math stops being "can we prove it" and starts being "do we have someone with the judgment to point the model at the right problem."

Yann LeCun Bets on Federated, Sovereign Open Models via Project Tapestry

While much of the industry chases scale, LeCun is pushing a different axis entirely. As Chief Science Advisor to the AI Alliance's Project Tapestry, he's backing a globally federated training approach where only model weights move between countries — never raw data [1][2]. Consortiums in India (BharatGen) and Australia (Monash) are already building sovereign derivatives aligned to local laws and values, with first milestone results shared publicly this year [2][3].

Yann LeCun weaving a tapestry at a wooden loom in a bright studio

The pitch is explicitly anti-centralization: no single lab or government controls the foundation, data sovereignty stays intact, and safety comes from distributed guardrails rather than one company's alignment team. Common Crawl joining the project this month adds real data infrastructure behind the idea, not just a press release [3].

It's a direct counter-narrative to the "scale begets safety via one trusted lab" framing that dominates US discourse — and it dovetails with what Europe is doing with EUROPA (see below). Expect this federated-sovereign model to become the default posture for any country or bloc that doesn't want to be permanently dependent on a US or Chinese frontier lab.

Domyn's EUROPA Hits Nearly a Trillion Parameters on EU Public Infrastructure

The EUROPA consortium, led by Domyn with Fraunhofer institutes, published training results from Germany's JUPITER supercomputer showing a test configuration at 951.9 billion total parameters — a mixture-of-experts model with 384 experts, 8 active per token, yielding 32.8B active parameters [1][2]. That's well past the European Commission's 400B minimum bar and closing in on the project's ~1 trillion parameter target [2][3].

The engineering detail that matters: throughput jumped from 90 to 1,300 tokens per second per GPU through expert-network optimizations, all on public EuroHPC infrastructure rather than hyperscaler clusters [1]. EUROPA is targeting all 24 EU official languages, open weights, and EU AI Act compliance from day one — a very different starting posture than most frontier labs, which bolt on compliance after the fact.

Combined with LeCun's Tapestry push, this is the clearest evidence yet that "sovereign AI" has moved from policy slide decks into actual GPU-hours. Europe building a trillion-parameter model on shared public infrastructure, openly, is either a historic flex or a cautionary tale about consortium-speed execution — probably both, and we'll know which within a year.

What This Means For Your Business

Today's stories all point at the same shift: the valuable skill is no longer writing the model or even writing the code — it's deciding what to point powerful, cheap AI at, and knowing when to trust the output. Liquid's sub-cent decision models make high-volume automation commodity-priced; GPT-6 Astra shows AI can carry serious proof work when a domain expert sets the direction; LeCun and EUROPA show entire governments betting that owning judgment over your own model matters more than renting someone else's frontier intelligence.

For companies building AI products right now, the practical read is: stop treating model choice as a one-time decision. You'll likely be orchestrating several models — a cheap decision model for the 95% of routine calls, a frontier model for the hard 5%, maybe a sovereign or regional model for compliance reasons depending on where your data lives. The chip export rule is a reminder that compute geography is now a genuine business risk, not a footnote — know where your inference actually runs.

The thread tying it all together: code and even raw model capability are becoming commodities. What isn't commoditized is the judgment to architect the pipeline, pick the right tool for each sub-task, and know which 5% of decisions actually need a human or a frontier model in the loop. That's the work now.

Key takeaway: The models are getting cheaper and more capable by the week — the competitive edge has already moved to who orchestrates them best, not who builds them.

See what we're exploring →

Sources

  1. https://www.liquid.ai/blog/d1-2
  2. https://www.marktechpost.com/2026/10/15/liquid-ai-d1-ultra-decision-model/
  3. https://www.liquid.ai/news/models
  4. https://techcrunch.com/2026/11/16/us-tightens-ai-chip-export-controls-to-deter-chinese-companies-from-accessing-compute-remotely/
  5. https://www.forbes.com/sites/zacharyfolk/2026/11/16/us-closes-loophole-that-allowed-china-to-remotely-access-nvidia-chips/
  6. https://www.reuters.com/legal/litigation/senator-accuses-us-commerce-department-failing-stem-flow-ai-chips-china-2026-11-16/
  7. https://i10x.ai/news/gpt-6-astra-standard-model-lattice-nathaniel-craig
  8. https://clauding.de/en/posts/openai-wissenschaftsblog-standardmodell-astra/
  9. https://agipulse.org/e/1128
  10. https://thealliance.ai/projects/tapestry
  11. https://www.prnewswire.com/news-releases/ai-alliance-launches-project-tapestry-to-build-a-collaborative-foundation-for-open-and-sovereign-ai-302735918.html
  12. https://commoncrawl.org/blog/common-crawl-joins-project-tapestry
  13. https://www.domyn.com/blog/training-1t-moe-eu-public-infrastructure-pytorch
  14. https://maccus.fr/en/eu-open-ai-model-europa/
  15. https://www.banandre.com/blog/eu-open-ai-model-europa-trillion-parameters

Stay ahead of AI

No spam. Unsubscribe anytime.

Want to go deeper?

Reading the news is one thing. Exploring the frontier is another. See what we're building.