Home · Politics · Nov 16 archive
Moonshot AI Uses Banned Nvidia Chips in China
· Confirmed
In Short: Researchers have discovered that leading Chinese artificial intelligence (AI) models were likely trained on Nvidia's Blackwell chips, which are banned from export to China due to US restrictions.

Victor Davis Hanson: Punish Nvidia’s Wrongdoing—Don’t Hand China the AI Crown. The Daily Signal
According to a new investigation, models like Moonshot AI's Kimi K3 and DeepSeek's V4 show fingerprints matching Nvidia's B200 or B300 processors, indicating that these chips were likely smuggled into China.
The White House is using Moonshot AI's Kimi K3 as a test case to address both model distillation and the use of blocked Nvidia hardware.
The findings suggest that Chinese companies may access these chips through overseas data centers in countries such as Malaysia and Singapore, circumventing US export controls.
Moonshot AI's main entity, Beijing Dark Side of the Moon Technology Co., agreed to lease 32 Super Micro Computer servers fitted with Nvidia's B300 chips, and these machines are located in China.
Lingmeixin, a company involved in the leasing agreement, pledged the servers and 345.6 million yuan in earnings from a service contract with Moonshot as collateral for a 200 million yuan ($30 million) loan.
Moonshot AI disputes claims linking the company to the financing arrangement, stating that it neither authorized nor knew about the related financing agreements.
The research, published by the US-based Unemployment Labs, found distinctive hardware fingerprints in six AI models developed by four Chinese companies: DeepSeek, Moonshot AI, MiniMax, and Ant Group.
Nvidia's board has said it will investigate, and three congressional committee chairs have asked Jensen Huang to attend a closed-door briefing in Washington on November 18.
The political fallout from the Nvidia spy scandal has prompted questions over whether senior Trump administration officials turned a blind eye to Nvidia's dealings with Chinese-linked companies.
Background
Will DePue, a former OpenAI researcher leading the nonprofit Unemployment Labs, analyzed the weights of large language models released by Chinese companies and found evidence that they were trained using Nvidia's Blackwell chips, which are banned from export to China.
What's confirmed
- The leading open-weight AI models built by Chinese companies were trained on Nvidia chips that Washington barred from export to China, according to a research effort reported by TechCrunch on November 13.
- It sorts what is confirmed from what is not, walks through the access routes that could explain the findings, and sketches what happens next for Nvidia, for cloud providers and for the people who build on Chinese open-weight models.
- The research, published by the US-based Unemployment Labs, an independent research organisation led by former OpenAI researcher Will DePue, found distinctive hardware fingerprints in six AI models developed by four Chinese companies: DeepSeek, Moonshot AI, MiniMax and Ant Group.
What's still developing
- US restrictions prevent Nvidia’s advanced Blackwell processors from being exported directly to China.
- Meanwhile, DeepSeek V4 Pro, Moonshot AI’s Kimi K3 and K3.1 Pro, and Ant Group’s Ling 3.1 Flash showed patterns associated with GB200 or GB300 chips.
- CEO William Li says proprietary chips will better match NIO's algorithms and eventually lift overall profit margins NIO $NIO CEO William Li said the Chinese electric vehicle maker is developing its own automotive chips to reduce its dependence on suppliers such as Nvidia $NVDA.
- For NIO to remain competitive on a global scale over the long term, Li pointed to two pillars: the company's automotive-grade chips, developed at nanometer scale, and its whole-vehicle operating system.
- Speaking to Reuters, Li said that building chips internally would help NIO 's bottom line over time, even accounting for the steeper upfront R&D investment required — noting that Nvidia 's automotive chips command "very high gross margin."
- The Moonshot portion was smaller than the total, but still large: Anthropic said it traced more than 3.4 million exchanges to Moonshot, targeting agentic reasoning, tool use, coding, data analysis, computer-use agents, and vision.
- Moonshot AI says the Beijing-built model has 2.8 trillion parameters, a 1 million-token context window, and open weights due on July 27.
- Some of China’s leading artificial intelligence (AI) companies may have used Nvidia’s most advanced chips to train their latest models, despite US restrictions on exporting the processors to the country, a new investigation has found.
- Based on their calculations, the analysis estimated that five distinct training runs collectively required around 12,000 Blackwell GPUs, with a plausible range of 7,000–22,000 chips.
- In a September analysis, nonprofit research organisation Epoch AI found that trade data was consistent with more than $3 billion worth of advanced chips being smuggled into China through Malaysia between April 2024 and June 2025.
- The agreement suggests a possible route for accessing restricted Nvidia hardware without directly purchasing the chips.
- China-focused policymakers have argued that restrictions should cover access to controlled chips, not only their ownership or installation in particular countries.
