A quiet week for models, a loud one for who controls them

Nothing at the frontier moved this week. No lab shipped a new flagship model, no context window got longer, no price list changed. The numbers that got the most attention came out of an earnings release. What moved instead was ownership: who builds the silicon, who may end up owning the place open models are distributed from, and who is allowed to buy access to a frontier lab's models at all.

The silicon week ended with NVIDIA's own numbers

Hot Chips ran Sunday to Tuesday and produced two concrete things. On 24 August NVIDIA said Groq 3 LPX, its interactive inference accelerator, is in full production as an extension of the Vera Rubin NVL72 platform; it cites a record 3,400 output tokens per second on the open Gemma 4 31B model at a 100,000-token context in Artificial Analysis benchmarking, with Nebius the first cloud to adopt it. On 25 August OpenAI published the first results for Jalapeño, its custom inference chip. Those figures come from OpenAI's own announcement; we found no independent evaluation of them.

Then on 26 August NVIDIA reported its fiscal second quarter: revenue of $96.2 billion, up 18% sequentially and 106% from a year ago, with data center revenue of $89.0 billion, up 117%, at a 75.0% gross margin. The same evening it announced with AWS a plan to deploy 2 million additional GPUs across AWS infrastructure in 2027-2028, bring Vera CPUs to AWS, extend NVLink Fusion with custom high-bandwidth memory, and build AI factories for the U.S. government including 100,000 GPUs on secure AWS infrastructure.

The interesting part is that these are the same companies. The labs and clouds designing their own accelerators to escape NVIDIA's leverage — OpenAI, Google, Amazon, Anthropic — are also the ones signing its largest deployments. Vertical integration and dependence are growing at the same time, not in sequence.

The open-model commons may get an owner

On 26 and 27 August, The Information reported that NVIDIA has agreed to buy Hugging Face for about $12.9 billion, after Business Insider first reported over the weekend that the company was fielding takeover interest, then reported the same night that talks valuing it above $13 billion had not yet produced a signed agreement. Neither company responded to TechCrunch's request for comment, and Hugging Face's own blog carried nothing about it this week. Treat it as a credible report, not a fact.

If it closes, the default distribution point for open weights would sit inside the company that sells the hardware nearly all of those weights run on. NVIDIA has spent the year backing open models — it ships its own Nemotron family — and the logic is not subtle: a healthy open ecosystem gives buyers an alternative to the closed labs, and those buyers rent NVIDIA GPUs. A coherent reason to want the asset, and a reason for anyone who depends on the Hub to watch who governs it.

Access became conditional

On 28 August OpenAI announced it is winding down the contract that puts its models inside Cursor, following Cursor's acquisition by SpaceX. SpaceX closed that acquisition in mid-August, exercising a $60 billion option from an April agreement. We could not open OpenAI's announcement page — openai.com returns 403 to automated clients — so this rests on OpenAI's own news feed, which carries the post, its date and its summary: a decision "to wind down our contract providing OpenAI models to Cursor following its acquisition by SpaceX." Widely reported details, including a specific shutoff date, could not be confirmed against a source we were able to open, so they are left out here.

For anyone building on a coding agent, the practical lesson is narrow and real: which models are available inside a tool is now partly a function of who owns the tool. That is a new failure mode. If a workflow, an agent config or a prompt library assumes a particular model in a particular IDE, that assumption now carries a corporate-ownership dependency it did not have a month ago. Keeping agent instructions portable across providers is no longer a hygiene preference.

Two courtrooms, one lab

Anthropic had the week's sharpest swing. On 27 August Judge Rita Lin, in the U.S. District Court for the Northern District of California, ruled that the administration's blacklisting of the company was unlawful First Amendment retaliation and vacated the directives that had barred federal agencies and defense contractors from using its products. "The empty invocation of national security is not a blank check to punish and retaliate against government critics," she wrote.

Late on 28 August, in the same district, Sony Music Publishing, Warner Chappell and other publishers sued Anthropic and named co-founders Dario Amodei and Benjamin Mann as individual defendants, alleging a "brazen campaign of illegally torrenting, scraping, and downloading copyrighted works" and seeking up to $150,000 per work plus up to $25,000 for each instance of stripped copyright management information. Anthropic told TechCrunch it disagrees with the claims and intends to defend itself. Nothing has been decided; the point for now is that the live legal question about frontier labs is provenance of training data, not model behavior.

What actually shipped

• Gemini Omni 1.1 Flash (27 August): Google's generative video model went production-ready for developers, with scene extension in 10-second increments up to 40 seconds cumulative, 10 seconds of prior context instead of one, start- and end-frame control, 360p previews for cheap iteration and 4K upscaling. Available through the Gemini API in AI Studio.
• Double-blind evaluations (27 August): Google DeepMind piloted what it calls the first double-blind evaluation of a proprietary frontier-class model, with the Singapore AI Safety Institute, OpenMined and AVERI, running a Gemini Flash Lite model inside Confidential Space so the evaluator cannot see the weights and Google cannot see the test prompts. A structural answer to benchmark contamination is more interesting than most benchmark results.
• Gemini Notebook limits (28 August): daily prompt caps give way to compute-weighted limits that refresh every five hours and factor in prompt complexity, chat length, source count and features used, for consumer accounts on web and mobile from 2 September. Google published no numbers.

A quiet week for capability, then. The models you were using on Monday are the models you are using today, at the same prices. What changed is the set of assumptions underneath them — about who supplies the compute, who curates the open alternatives, and whether the vendor relationship you built on survives its next acquisition.