The Omniscient Bulletin · 2026-09-04
The Omniscient Bulletin — September 4, 2026
Nvidia has agreed to buy Hugging Face for 12.9 billion dollars, and committed in the filing to keep the platform open and to keep supporting rival silicon. OpenAI released GPT-6 Astra and priced it exactly where Anthropic prices Claude Fable 5.1, while holding most customers back for now. And Tesla started carrying riders in a car with no steering wheel, and within a day regulators had opened a formal investigation.
Nvidia is buying the place where open models live, and the terms it had to write down are more telling than the price. The filing commits Nvidia to keeping Hugging Face open, to letting any model maker upload and download freely, and to supporting other silicon vendors. Those promises exist because the buyer sells the accelerators that the platform's users are choosing between. The same filing then carries a risk factor complaining that other parties are lobbying Washington to restrict open-source models. Nvidia now owns the commons it says is under political attack, and it has until the first half of 2027 to persuade regulators that both things can be true at once.
Industry
Nvidia is buying Hugging Face, and had to write into the filing that its own chips will not be required
Nvidia entered the agreement on Wednesday and announced it Thursday. Jensen Huang put the price at 12,930,300,000 dollars. The 8-K splits that number in a way no headline did: about 11.9 billion dollars goes to Hugging Face shareholders, and up to 1 billion more is retention equity for employees who join. Nvidia committed in the filing to keep the platform open, to let any model maker upload and download, and to support other silicon vendors. It closes in the first half of 2027, if regulators agree.
Frontier
An Abu Dhabi institute released six models with the training data attached, and named OpenAI and Anthropic as the counterexample
IFM released six models on Thursday together with the data and methods used to build them. Its K2 Horizon release carries weights, training data, code, methodologies and intermediate checkpoints, so outsiders can retrace how they were built and reproduce the results. Founder Eric Xing says the goal is a reference point for what a truly open model release looks like, and to show regulators that openness and competitive performance are not mutually exclusive. It runs from smartwatch size to 375 billion parameters.
Frontier
OpenAI published a price for its most capable model, and it is the same price Anthropic charges for Claude Fable 5.1
OpenAI released GPT-6 Astra on Thursday and put a number on it: 10 dollars per million input tokens, 50 per million output, 12.50 for a cache write. Those are Claude Fable 5.1's figures on all three lines, and twice what Claude Opus 5 costs. OpenAI calls it the most capable model it has ever broadly deployed, and its first to reach the Critical level for cybersecurity. Access is staged, though: Trusted Access Program enterprises now, with the API and the paid ChatGPT plans coming in the coming days.
Research
One model scored 62.7 percent and 99.9 percent on the same benchmark, and the near perfect run was the cheaper one
ARC Prize published its Astra results on Thursday. On the standard harness the model scored 62.7 percent and spent 26,000 dollars. On a provider adapter harness, which lets it keep its own opaque reasoning state between requests, it scored 99.9 percent and spent 19,000. ARC says the better score costs less because at high effort the model solves each game in fewer moves. It used fewer actions than the median tested human on 96 percent of levels, and invented a shorthand of its own to track each new environment.
ARC Prize, on its own benchmark→
Policy
A Pentagon under secretary says Anthropic is still a supply chain risk, a day after the Commerce Secretary said it was resolved
Emil Michael, the under secretary of defense for research and engineering, posted on X on Thursday that Anthropic is still a supply chain risk to the defense department and the defense industry. Commerce Secretary Howard Lutnick told Axios a day earlier that Anthropic is back on the right side, and that he trusts the company. A judge ruled on August 27 that the Pentagon illegally punished Anthropic for criticizing the government, calling the decision illegal and baseless. The department did not respond to Reuters.
Reuters (via wire syndication)→
Compute
A humanoid robot maker is buying compute at the scale a frontier lab buys it, and the supplier is taking equity in return
Nscale signed a multi year deal on Thursday to supply compute for Figure's humanoid robots, committing 3.5 billion dollars and saying it plans to scale that past 6 billion. First chips are targeted for Barstow, Texas from the second half of 2027, and Nscale intends to deploy up to 100,000 Nvidia chips. It becomes Figure's preferred compute provider and a Figure shareholder in the same deal, powering the next generation of Figure's Helix models. The two will also explore scaling Nscale's supply chain with humanoids.
Policy
Tesla put a driverless Cybercab on Austin streets, and by the next day regulators had opened a formal investigation
Tesla began charging riders Thursday for trips in a two seat Cybercab built without a steering wheel, pedals, or mirrors, features federal rules typically require. By Friday, the National Highway Traffic Safety Administration had moved past the wait and see line it gave reporters and opened a formal audit query, examining whether Tesla properly self certified up to 1,000 of the vehicles and the basis for deciding certain safety standards did not apply. Unlike Amazon's Zoox, Tesla never sought the exemption federal rules allow for wheel-less robotaxis.