The week began with an investigation at Nvidia and ended with a new CEO there. Around that story, Anthropic and Google shipped new models, agents got the ability to pay for things, and the rules around AI got tighter, both in law and inside large companies. Here are the ten stories that mattered most, counting down to the biggest.
10. llama.cpp reaches version 1.0
The project that made it practical to run language models on a laptop is now stable. llama.cpp and its ggml library both reached 1.0, with a tagged ggml release to build on.
Open-weight models kept pace. Unsloth published new Qwen3.8-27B quantizations, and a retrained Qwen3.8 now answers questions the original model was trained to avoid.
9. Data centers become an election issue
After the US midterms, analysts said the outlook for data center expansion got harder. In Michigan, a candidate who ran against data centers won a House seat and plans to push for a federal moratorium.
Reporting from Virginia suggests the backlash is mostly about electricity bills. In the same week Anthropic bought $3.4 billion of gas turbines.
8. DeepMind proves Signal’s encryption with AI agents
Google DeepMind introduced AlphaProof Code. AI agents and human experts wrote about three million lines of Lean proofs that cover Signal’s post-quantum handshake and its Triple Ratchet, Wired reports.
A machine checks every proof. DeepMind says the long-term aim is the same kind of guarantee for software in general.
7. Robots get their own foundation models
At the CoRL conference, Nvidia launched GR00T 2 and Isaac Lab 3.0, and Google released Gemini 4 Argon for robot control and simulation.
GR00T 2 predicts future video and robot actions together, so a skill learned on one machine can move to another. Runway’s world-action model Praxis-1 is now available with open weights.
6. AI safety laws start to apply
New York’s AI safety law took effect. OpenAI, Anthropic and other large developers must register, disclose their safety practices and report critical incidents.
Australia may make knowingly selling unsafe AI a crime. OpenAI also described a kill switch for its research models that it says can stop one in about 35 seconds.
5. AI tools leak secrets and customer data
Claude Code settings files committed to GitHub are exposing thousands of secrets. A breach at inference provider Fireworks AI reached Notion’s customers.
A researcher found 35,000 apps built with xAI’s Grok Build agent left open, some with confidential SpaceX data. Hugging Face warned that rising attacks could force open AI platforms to lock down.
4. Agents get a wallet, and companies set limits
OpenAI launched ChatGPT Wallet, which lets agents in ChatGPT, Codex and Dots complete purchases after the user approves. It starts with US Pro subscribers and major US cards.
Not everyone wants agents as customers. United Airlines said personal AI agents may not book its flights. Inside companies, Goldman Sachs is reportedly pulling Claude Code from bankers and traders, and Apple is said to cap engineers at $200 a day in AI spending.
3. Google ships a week of Gemini models
Gemini 4 Hydrogen replaced 3.8 Flash in a fast, broad rollout and now powers AI Mode in Search. Gemini 4 Carbon, first limited to cyber defenders, is open to everyone.
Developers also got the Gemini Decisions API for fast multimodal decisions. DeepMind released AlphaGenome 2, which predicts how DNA variants affect specific cell types.
2. Anthropic ships Fable 5.2 and Mythos 5.2
Anthropic released two new models. Fable 5.2 took first place on Artificial Analysis, and Mythos’s bug-finding is going into Cisco and IBM products. The company is reportedly seeking a $1.3 trillion valuation in its IPO.
The money around the frontier labs kept growing. OpenAI’s quarterly revenue reportedly passed $10 billion for the first time, and TSMC will invest $1 billion in Anthropic and OpenAI. CoreWeave, meanwhile, is asking customers to pay up front to fund its build-out.
1. Jensen Huang resigns as Nvidia CEO
Early in the week, Nvidia’s board opened an investigation into Huang’s relationship with a Megaspeed executive whom a US intelligence memo describes as recruited by China’s Ministry of State Security. By Wednesday the board was reportedly preparing to remove him, and on Thursday he resigned. Colette Kress is the new CEO.
The fallout reached beyond the company, which sells most of the chips the industry trains its models on. A forensic report found that China’s top models from DeepSeek, Moonshot AI, MiniMax and Ant Group were trained on export-banned Blackwell chips. Founders and VCs expect smaller checks for AI startups, and by the weekend Elon Musk and other executives had signed a petition asking Nvidia to bring him back.
Next week will show whether Nvidia’s new leadership changes its course on China and on open-source models. The daily editions are on the AI Digest page.