AI Briefing

Nvidia just bought the open-source AI world's town square for $12.9 billion, roughly 80 times Hugging Face's annual revenue, and the reason why says a lot about who Nvidia thinks its real customers will be. Meanwhile Meta quietly shelved a plan to cut teams by 60 percent after its own AI agents underdelivered and its staff revolted, and OpenAI caught a Russian operation using ChatGPT to launder fake think-tank research into five countries. Plus a legal publisher trained its own frontier-ish LLM for the price of a used car, and Alibaba shipped a preview of Qwen4's architecture.

01IndustryTHE DECODER ↗

Nvidia Buys Hugging Face for $12.9 Billion, About 80x Its Annual Revenue

Nvidia is acquiring Hugging Face, the default hosting hub for open AI models, for $12.9 billion against roughly $150 million in annual revenue. The timing is not subtle: OpenAI is building custom chips with Broadcom, signed a deal with Cerebras and secured AMD access, while Anthropic and Google build their own silicon. As the closed labs peel away from Nvidia hardware, owning the open-weight ecosystem, and potentially turning Hugging Face into an OpenRouter for open models, is how Nvidia keeps demand flowing.

02IndustryTHE DECODER ↗

Meta Planned to Cut Teams by 60% and Replace Them With Agents, Then the Agents Failed and Staff Revolted

Internal documents reported by Reuters describe 'Project OT,' a Meta plan to shrink many teams by up to 60 percent and hand the work to small human groups supervising virtual AI workers. Hours before the first round of layoffs on May 19, Zuckerberg halted the planned November second wave. The agent technology never produced the promised productivity gains, investors balked at the AI budget, and employees openly rebelled after concluding that keystroke-logging software was training their own replacements.

03PolicyTHE DECODER ↗

OpenAI Caught a Russian Operation Using ChatGPT to Build a Fake Israeli Think Tank

OpenAI banned a cluster of accounts running a covert Russian influence campaign that used ChatGPT to write social posts while explicitly instructing the model to strip out linguistic tells of Russian origin. The operation promoted the 'International Burke Institute,' a fictitious Israel-based think tank whose 'sovereignty index' ranked Russia above Western countries. Of 36 expert-attributed IBI articles, 34 were plagiarised, including a Cambridge University Press piece falsely credited to a Nottingham professor. Reach was small, but OpenAI notes the infrastructure was built to scale.

04ProductLawSites ↗

Thomson Reuters Trained Its Own Legal LLM, and the Final Run Cost About $450,000

Thomson Reuters launched 'Thomson,' its first proprietary large language model, built from an open-source base on decades of Westlaw, Practical Law, Checkpoint and Reuters content. The company spent roughly $40 million over two years on people and compute, but says efficiencies brought the final training run down to about $450,000. It is a template other data-rich incumbents will notice: you do not need frontier-lab budgets to own a domain model, you need proprietary content nobody else has.

05Open SourceTHE DECODER ↗

Alibaba Ships Qwen3.8-Flash-Next, a 125B Model That Activates Only 6B Parameters

Alibaba's Qwen team released Qwen3.8-Flash-Next, a multimodal mixture-of-experts model billed as an architecture preview of Qwen4. It carries 125 billion total parameters but activates just 6 billion per token, plus a novel 51-billion-parameter N-gram embedding layer, a 'phrase dictionary' that can live in ordinary system RAM instead of GPU memory. Context runs to 262,144 tokens natively and a million with YaRN, weights are on Hugging Face and ModelScope, and the hosted version costs $0.16 per million input tokens.