# AI News — August 29, 2026: Judge Blocks Anthropic Blacklist, Nvidia–Hugging Face Closes at $13B > Source: > Published: 2026-08-29 06:00:08+00:00 Good morning. The Nvidia–Hugging Face deal we’ve been tracking all week is now official at $13B, but the more consequential rulings today came from a federal courtroom, where a judge tore apart the Pentagon’s attempt to blacklist Anthropic. Meanwhile, Anthropic researchers quietly demonstrated automated alignment work that outperforms humans at a fraction of the cost, and the EPA is trying to make sure you never find out a data center is being built next door. **Nvidia–Hugging Face is signed.** After days of rumors, [Business Insider](https://www.businessinsider.com/nvidia-in-talks-to-buy-hugging-face-13-billion-dollars-2026-8) reports the $13B acquisition is agreed. TechCrunch places it in a [broader run on open-weight platforms](https://techcrunch.com/2026/08/28/open-weight-ai-companies-are-the-valleys-hottest-acquisition-targets/) — Nvidia’s $6B Poolside buy, Stripe’s $7B for OpenRouter — even as open-weight adoption sits at just 6% of companies and 2% of engineers. Nvidia’s motive is straightforward: hyperscalers and frontier labs are building their own chips, so owning the distribution layer is a hedge. **A federal judge rules the Anthropic blacklisting unconstitutional.** Judge Lin found the Trump administration’s designation of Anthropic as a supply-chain risk was “unlawful retaliation” under the First and Fifth Amendments, per [TechCrunch](https://techcrunch.com/2026/08/28/anthropic-gets-its-first-court-win-over-the-pentagons-supply-chain-risk-label/) and the [New York Times](https://www.nytimes.com/2026/08/27/technology/anthropic-government-blacklisting-ruling.html). The designation, ordered by Defense Secretary Hegseth after Anthropic refused to remove guardrails against autonomous weapons and domestic surveillance, was contradicted by the government’s own continued contracting with the company — including collaboration on the Mythos cybersecurity model. The administrative record justifying the ban was a four-page memo, most of it dated after the actions it was meant to justify. HN commenters were less celebratory than you’d expect: one noted “you can beat the rap but not the ride,” pointing to the smaller companies that folded rather than fight. **Anthropic’s automated alignment researcher beats humans.** A paper out of the Anthropic fellows program shows an AI system that runs the full alignment research loop — literature review, method proposal, iterative training — and fixes all 10 tested misalignment benchmarks without regressing capabilities. [TechCrunch](https://techcrunch.com/2026/08/28/an-anthropic-researcher-just-gave-us-a-peek-at-self-improving-ai/) reports it beats experienced human researchers in six hours at roughly $4/hour versus $150/hour. The authors are careful to note the ceiling is set by benchmark quality and the underlying literature, which is where the interesting failure modes will live. **A multi-agent system produces novel math.** A paper describes “the Station,” a decentralized multi-agent setup that produced new results on kissing configurations, Kakeya sets, and an improved bound on Erdős’s minimum-overlap problem — with human-readable proofs, not just numbers. From the [arXiv paper](https://arxiv.org/abs/2608.23691) and the [HN discussion](https://news.ycombinator.com/item?id=49481455), the design choices are charming: agents work without central coordination and receive scheduled “holidays” with random prompts to encourage open-ended thought. One commenter dryly observed that the researchers have reinvented the Cambridge Senior Common Room, only with GPUs. **GLM-5.3 arrives on Hugging Face.** Z-AI followed up Flash with the full [GLM-5.3](https://huggingface.co/zai-org/GLM-5.3) open weights, and early users are comparing it favorably to Claude Opus for reasoning and coding. The catch, flagged repeatedly on HN, is that 5.3 lacks vision, which pushes several users back toward the Flash variant for UI work. DeepInfra was first to host it on OpenRouter. **Lambda borrows another billion to buy GPUs.** Neocloud Lambda has raised [$1B in short-dated debt from JPMorgan](https://techcrunch.com/2026/08/28/neocloud-lambda-secures-1b-in-debt-to-buy-more-chips/) to acquire Nvidia chips it will lease to Microsoft, on top of a $1B credit facility in May and a separate $926M loan this week. A $3B pre-IPO round is reportedly in the works. Total AI-related debt raised globally in 2026 has crossed $400B, which is the sort of number that only looks reasonable if inference demand keeps compounding. **The EPA wants data centers to skip public notice.** The Trump EPA is [proposing to gut the New Source Review rule](https://www.theverge.com/ai-artificial-intelligence/986176/data-center-pollution-epa-rule-change-air-permit), which requires public comment periods before industrial sites get air permits. Critics say the timing and language are aimed squarely at data centers and their associated gas plants, which have been running into community opposition. Under the new rule, residents would find out when the bulldozers show up. **A JAMA paper argues doctors should get out of the loop.** Ezekiel Emanuel and Vinod Khosla [co-authored a piece in JAMA](https://www.wired.com/story/ai-has-human-doctors-asking-whats-left-for-us/) arguing AI alone will soon outperform physicians on core tasks and that “humans in the loop degrade AI performance.” Emanuel, previously an AI skeptic, says the recent literature converted him. The disclosed conflicts are non-trivial: co-author Neal Khosla runs an AI healthcare company, and Vinod Khosla holds related investments. That’s a lot of court, cash, and code for one Friday. Have a good weekend — and if you hear bulldozers, check the permit filings.