AI News — August 28, 2026: Nvidia's $13B HF Handshake, GLM-5.3-Flash Tops Charts on Chinese Silicon

Nvidia has agreed in principle to acquire Hugging Face for $13 billion, nearly double the $7 billion valuation at which Hugging Face rejected a $500 million Nvidia investment six months ago. Separately, Z.ai released GLM-5.3-Flash, a 320B MoE model with 18B active parameters that matched frontier performance on Chinese chips and topped OpenRouter's popularity charts, while a federal judge blocked the Pentagon's blacklisting of Anthropic as unconstitutional retaliation.

Good morning. The Hugging Face saga has a price tag and a signature — Nvidia is buying, for $13B — but the more interesting numbers today come from China, where GLM-5.3-Flash matched frontier performance on Chinese silicon and quietly topped OpenRouter’s popularity charts before anyone knew what it was. Also: a federal judge told the Pentagon it can’t blacklist AI companies for refusing to build autonomous weapons. Nvidia’s Hugging Face deal moves from rumor to handshake. The $13B acquisition we flagged Monday is now reportedly agreed in principle https://techcrunch.com/2026/08/26/nvidia-closes-in-on-hugging-face-acquisition/ , though no papers are signed. Business Insider https://www.businessinsider.com/nvidia-in-talks-to-buy-hugging-face-13-billion-dollars-2026-8 has the same figure, and the HN thread is fixated on the reversal: HF turned down a $500M Nvidia investment at a $7B valuation six months ago citing worries about a dominant shareholder, and now they’re selling outright to that same shareholder for nearly double. Community concern centers less on the price than on what Nvidia gets access to — hardware survey telemetry, model download patterns, and the discovery layer for essentially all open-source AI. GLM-5.3-Flash matches frontier models on Chinese chips. Z.ai released GLM-5.3-Flash https://z.ai/blog/glm-5.3-flash , a 320B MoE with 18B active parameters that reportedly matches its bigger sibling at a fraction of the cost. The reveal buried in the announcement: the model was tested anonymously as “ox-alpha” on OpenCode and OpenRouter for weeks, became the most-used model of the week, and served all that traffic from Chinese AI accelerators. One HN commenter tallied the pace: Kimi K3 caught up to Opus in mid-July, GLM 5.3 cut parameters by two-thirds four weeks later, and Flash cut them again twelve days after that. The catch is Z.ai’s terms of service, which grant a perpetual license over inputs, outputs, and even user profile pictures, plus vague prohibitions tied to “national interests.” A judge blocks the Pentagon’s Anthropic blacklist. Federal judge Rita F. Lin ruled that Defense Secretary Pete Hegseth’s designation of Anthropic as a “supply-chain risk” was “arbitrary and capricious” and unconstitutional retaliation under the First Amendment, per The Verge https://www.theverge.com/ai-artificial-intelligence/985947/anthropic-supply-chain-risk-lawsuit-judge-ruling and Wired https://www.wired.com/story/a-judge-has-blocked-the-pentagons-attempt-to-blacklist-anthropic/ . The blacklisting followed Anthropic’s refusal, during a $200M contract negotiation, to allow Claude to be used for mass surveillance or lethal autonomous weapons — terms most other labs accepted. NYT coverage https://www.nytimes.com/2026/08/27/technology/anthropic-government-blacklisting-ruling.html is out too, though HN commenters mostly wonder whether rulings against the administration mean much in practice, and note the chilling effect already worked on the labs that quietly agreed. Google ships two new Gemini models. Gemini 3.5 Transcribe https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5-transcribe/ posts a 2.6% word error rate for non-streaming speech-to-text across 85+ languages, though testers on HN say Soniox STT v5 and Deepgram still beat it on real-world benchmarks, particularly for real-time diarization beyond three speakers. Separately, Gemini Omni 1.1 Flash https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-omni-1-1-flash/ extends video generation to 40 seconds with 360p drafts and 4K upscaling, though commenters point out that non-deterministic outputs make draft-then-upscale workflows nearly useless — the same prompt at higher resolution produces different video. Jensen Huang declares AGI achieved. Again. On Nvidia’s earnings call, Huang said the company has “achieved AGI” for many tasks while simultaneously calling AGI milestones “senseless” https://www.theverge.com/ai-artificial-intelligence/985597/jensen-huang-says-nvidia-achieved-senseless-agi without a shared definition. His pivot: what matters is AI doing “productive and useful work” and “generating profitable tokens.” Convenient framing given the day’s other Nvidia news. Barret Zoph lands at Google. After being ousted from Thinking Machines the Mira Murati venture and a five-month return-to-then-departure from OpenAI, Zoph is now VP of Research at Google https://techcrunch.com/2026/08/27/barret-zoph-the-thinking-machines-co-founder-who-defected-to-openai-is-now-at-google/ , working on Gemini post-training. It’s a return to a former employer, and one more data point in the ongoing exodus of senior researchers from OpenAI ahead of its IPO. That’s the morning. If GLM-5.3-Flash keeps its OpenRouter lead now that everyone knows who made it, the Nvidia–HF deal will look even more strategically necessary — and even harder to swallow for anyone counting on open infrastructure staying neutral.