Good morning. The Jensen Huang fallout keeps compounding — today it’s export-control violations, circular financing allegations, and an AI startup funding gap. In lighter news, Claude cracked Gödel’s notebooks and llama.cpp finally hit 1.0.
China’s best models are running on banned Blackwells. A nonprofit called Unemployment Labs analyzed outputs from DeepSeek, Moonshot, MiniMax, and Ant Group and found computational fingerprints consistent with training on Nvidia’s export-restricted Blackwell GPUs. The group estimates around two million advanced chips reached China through smuggling routes and offshore workarounds, including Malaysian data centers. TechCrunch’s writeup ties the findings to the Megaspeed investigation and raises the obvious corporate liability question now that Huang is gone.
The scarcity was partly manufactured. A separate report, surfaced on TechPowerUp and discussed on HN, alleges Nvidia funded AI startups that then bought its GPUs, helping offload TSMC oversupply while preserving the appearance of runaway demand. CFO Colette Kress reportedly pushed back internally on the circular financing. Commenters were split — some noted The Guardian had the circular-finance angle a year ago, while others argued investing in your own customers is standard playbook, pointing to Henry Singleton as precedent.
Nvidia’s startup checkbook is closing. Huang’s exit has blown a hole in AI venture funding, per TechCrunch. Nvidia participated in 105 rounds in 2026 and was often the lead check for neoclouds, inference providers, and open-weight model shops — CoreWeave, Crusoe, and Mistral among them. Kress is expected to tighten the spigot considerably, and founders aren’t optimistic the pace resumes soon.
OpenAI launches ChatGPT Wallet. Pro subscribers in the US can now let OpenAI’s Dots agents make purchases on their behalf, with Face ID or fingerprint required per transaction and tokenized card details handled like Apple Pay. The Verge has the details. Full delegation within preset spending limits is on the roadmap, though retailer blocking — Amazon already shut out Meta’s Muse agents — remains an open question.
OpenAI also disbanded its robotics team. Researchers are folding into the main research and applied orgs, with the company betting that general-purpose models will eventually run robots directly. Fortune reports OpenAI is still targeting a humanoid demo by 2027. HN commenters were skeptical: one noted general models are currently bad at low-level control and that stacked specialist models remain more efficient, while another read it as pre-IPO cost cutting.
LangChain hits a $3B valuation. CapitalG led a $300M Series C, per TechCrunch, more than doubling the company’s valuation from a year ago. LangChain now claims 7,000 customers for LangSmith, including Nvidia, Toyota, and Salesforce.
Claude helped decode Gödel’s notebooks. Researchers at Helsinki University used Claude to read Kurt Gödel’s 1,009 pages of notes in Gabelsberger shorthand, surfacing hundreds of previously unknown theorems within days — work that had been crawling along at roughly one page per week. The Guardian reports lead researcher Jan von Plato found results anticipating 1970s–80s discoveries plus 14 unexplored logical themes. HN skeptics pushed back a bit, noting Gödel was a famous perfectionist who rarely published anything that wasn’t bulletproof — so “theorem” may be doing some lifting here.
llama.cpp ships 1.0.0. After 3.5 years, the project adopted semantic versioning for both libllama and libggml, acknowledging its role as downstream infrastructure for a huge chunk of the local AI ecosystem. The release post also notes expanded distribution via Docker, Debian, Homebrew, conda-forge, and Winget.
One more for the researchers. A detailed writeup on 600 experiments with JEPA world models found that the learned predictor performed nearly as well as perfect physics simulation (78% vs 77%) on long-horizon tasks, suggesting prediction quality isn’t the real bottleneck. The author instead developed four new planning approaches, including on-the-fly subgoal generation via memory graphs. HN commenters noted hierarchical planning and subgoals are well-established ideas, and that MCTS already solves Push-T without training.
That’s the briefing. Huang testifies before Congress by Thanksgiving, so expect the Nvidia story to get louder before it gets quieter.