DailyAI Briefing: Chinese AI tied to banned Nvidia chips, plus 4 more

Welcome to the DailyAI Briefing, the day’s five biggest AI stories. A former OpenAI researcher says six Chinese AI models carry the fingerprint of Nvidia chips that need a US licence before they can be sold to China. Banks handling the Anthropic share sale are turning away buyers from mainland China and Hong Kong, startup founders are bracing for less money from Nvidia after the exit of Jensen Huang, Cursor explained how a leaked key let a fake coding extension reach its users, and the popular open-source tool llama.cpp reached its first stable version.

Let’s dig in.

Six Chinese AI models carry the fingerprint of banned Nvidia chips

Six models from four Chinese labs, DeepSeek, Moonshot AI, MiniMax and Ant Group, show signs of being created on the Blackwell chips from Nvidia, according to Will DePue, a former OpenAI researcher who now runs a small nonprofit lab. Selling those chips to China requires a US export licence, and none has been reported. The full report was published on 12 November.

Before training starts, the weights of a model are filled with random numbers, and the way the software lays them out depends on how many processor cores the chip has. Traces of that layout survive training. DePue searched more than six trillion combinations of chip layout and starting number. He then reproduced the starting weights of two of the models, DeepSeek-V4-Flash and Kimi K3, on real Blackwell hardware. Earlier models from the same four labs match the previous generation of Nvidia chips.

I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.

— Will DePue (@willdepue) November 12, 2026

The report is clear about its limits. Chinese companies can rent computing power in data centres outside China, so the fingerprint says nothing about where the chips were located or whether any law was broken. DePue estimates that Kimi K3 and DeepSeek-V4-Pro each used 2,000 to 3,000 of the chips. He also warns that labs can now erase many of these traces, so this kind of check could soon stop working. Whatever the route, rules meant to keep the best chips away from Chinese labs are clearly not working.

Banks refuse Chinese orders for Anthropic shares

22 banks are working on the Anthropic stock market listing, led by Goldman Sachs and Morgan Stanley, and several of them are refusing orders from investors in mainland China and Hong Kong. The restrictions were reported on 13 November by a financial news service, citing people familiar with the matter. Some of those people pointed to the US rules on exports of defence goods. Anthropic declined to comment.

The shares are due to be priced on 18 November and to trade on Nasdaq under the ticker ANT. Our earlier report covers the price range and the size of the sale. The company already blocks firms under Chinese control from using its AI services, so this fits its record. Still, when weapons export rules help decide who may buy shares in an AI company, governments are treating AI as a national security matter.

Startups brace for less money from Nvidia

105 funding rounds. That is how many startup deals Nvidia has joined this year, often as the largest investor, according to PitchBook data. Founders and investors told a tech news site that much of that money depended on Huang himself. One investor described Colette Kress, the interim chief executive, as “much more conservative”.

Want AI to save you hours a week? Our free playbook shows you how.Get the free guide →

Nvidia backed CoreWeave before it went public, holds stakes in Mistral AI, Lambda and Black Forest Labs, and agreed this year to buy Hugging Face for $13 billion, a deal expected to close in 2027. Investors expect smaller cheques and lower valuations for young companies that build data centres or release open AI models.

The deals also tied those startups to Nvidia chips. If Nvidia steps back, rival chipmakers and cloud companies get a rare opening, though few are likely to put hundreds of millions into startups with no revenue yet.

A leaked key let a fake coding extension reach Cursor users

5,000. That is the most users who may have received a malicious copy of the Claude Code extension from the Cursor extension marketplace on 29 October. It stayed up for about 11 hours. The full account of the breach was published on 13 November.

The trail started at Fireworks AI, a company that hosts AI models for other businesses. An attacker got hold of an access key that Cursor had issued to Fireworks and used it to copy old coding sessions from shared storage. A few of those sessions still held login details for other services, which led the attacker through internal tools to the marketplace. Cursor says the original extension had no flaw, users in Privacy Mode were not affected and the last attacker activity was on 1 November.

We’ve published an update on our recent security incident. See a detailed overview of what happened and our remediation steps here: https://t.co/2lf6lRJPlL

— Cursor November 14, 2026

Cursor has moved all its AI traffic away from Fireworks. Our report on the second breach at the provider has more detail. Before a company shares data with an AI vendor for training, it should delete every password and key in it, because one leak can open many doors.

The llama.cpp project reaches version 1.0

3.5 years after it started, the open-source project llama.cpp has reached its first stable release. It is free software that runs AI models on ordinary computers. The team announced version 1.0.0 in a blog post on 13 November. From now on, any change that breaks the public interface for developers will come with a new major version number, so apps built on top of it know when an update could break them.

llama.cpp v1.0.0 is now available https://t.co/twQwJkH8kE

— Georgi Gerganov (@ggerganov) November 13, 2026

The software supports more than 150 model architecture variants and can be installed through Docker, Homebrew, Winget and other channels. A version number is not exciting news, but a promise of stability matters to the many apps that run AI on your own device rather than in the cloud. It also links back to our third story, because the team behind llama.cpp works at Hugging Face, which Nvidia has agreed to buy in a deal that has not closed yet.

That’s all for today. See you tomorrow.