Subscribe
Share
Share
Embed
Anthropic drafts a court-gated release halt
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Wednesday, October twenty-first.
In today's briefing we have Anthropic pitching a federal bill that would let a judge halt a model release over mass-casualty risk, a leaked S-1 pointing to capacity strain at Anthropic, and DeepSeek open-sourcing the inference stack behind its production services.
First up - Today in the big model news;
OpenAI
Reuters, via Gizmodo, reports that OpenAI wants a valuation of one point four trillion dollars or more, plus about thirty billion dollars for infrastructure. Sam Altman said in September that it would not list in 2026. The report is unconfirmed. The infrastructure ask matches the capacity pressure reported at Anthropic: at the top of the market, compute is the line item driving the capital raises.
Anthropic
Governments keep moving toward controlling who gets frontier models and under what conditions, and now a lab is drafting the compliance plumbing itself. Anthropic has pitched a federal bill, confirmed by a spokesperson after Politico reported it. Under the bill, independent evaluators registered with the Commerce Department would have seventy-two hours to alert the department. Commerce could then ask a judge in Washington to halt a release. The trigger is a foreseeable, imminent risk of fifty or more deaths, or one billion dollars or more in damage. No agency could halt distribution on its own; only the court could. And this part is contested: TechCrunch notes it is unclear how letting labs pay the evaluators avoids influence over the results. The proposal comes from a lab, not a regulator, which complicates the picture of control imposed from outside. Evaluator access and release-gating obligations are likely to become procurement questions, so ask model vendors how they would support a halt order mid-contract.
Separately, a leaked S-1 reportedly shows Anthropic struggling to get enough capacity from AWS, Google and Microsoft, which increasingly see it as a competitor. Figures attributed to the Wall Street Journal put the first-half net loss at eighty point seven billion dollars on sixteen point three billion dollars of revenue, with two customers making up about a fifth of sales. Anthropic did not respond. All of it is press-sourced and unverified, and the public filing would confirm or kill the capacity claim. Until then, keep a second-provider fallback on critical Claude paths.
In AI Infra
DeepSeek has released DeepInfer under the Apache 2.0 license, the framework behind its production services. It is offline-only, supports just V4.1 Flash, and leaves KV cache, prefill and decode disaggregation, and load balancing to the user. DeepSeek reports one hundred fifty-five trillion tokens a day on one cluster and a gain of four to seven percent from its DSpark speculative decoding. Every figure is DeepSeek's own, and a search found no third-party comparison with vLLM or SGLang. What it adds is the serving recipe, not only the weights, so the open stack's cost story now has a published reference. Benchmark it against vLLM on your own traffic before committing.
Google has open-sourced TorchTPU, which it says runs PyTorch apps on TPUs with a configuration change, built with Meta, Hugging Face, vLLM and SGLang. Amazon released TorchNeuron for Trainium a day earlier and submitted code to PyTorch to decouple features from Nvidia chips. TechCrunch cautions that developers will not automatically leave Nvidia, and the ease-of-migration claim is Google's. Even a test run of your PyTorch inference on TPU or Trainium changes what you can negotiate with GPU providers.
In other news…
Meta paused Muse chat sharing after the security firm Claroc found more than sixty-eight hundred shared conversations through a single passive-DNS search. Conversation titles were embedded in share-URL subdomains. Exposed items included a loan application and a bank fraud alert with a partial account number, and Claroc has no evidence of misuse. Claroc reported it to Meta's bug bounty on October eighth, and Meta suspended sharing after Ars Technica asked. Claroc concluded the behavior was by design. Titles persist in DNS history even though the pages are unreachable. Hostnames and certificate logs are public, so share links should carry only an unguessable token.
That's the briefing. Have a great day, and don't forget to subscribe.