Other SBS Sports · 6d ago

US Lab Finds Evidence Chinese AI Models Trained on Banned Nvidia Blackwell Chips

Key Points
  • US non-profit Unemployment Labs reported forensic evidence that six recent Chinese AI models—including DeepSeek's V4 series and Moonshot AI's Kimi K3—were trained on Nvidia's export-restricted Blackwe
  • The findings are drawing intense attention in South Korea as they coincide with the dismissal of former Nvidia CEO Jensen Huang over an alleged affair and chip-smuggling scandal.
  • Researchers estimate the training clusters required roughly 12,000 Blackwell chips worth $500 million, though they noted the study does not directly prove export-control violations.
Summary A new forensic analysis from a US research institute has reignited debate over the effectiveness of Washington's semiconductor export controls, suggesting that leading Chinese artificial intelligence companies have been training their latest models on Nvidia's restricted Blackwell chips. On November 14, Unemployment Labs—a US non-profit led by former OpenAI researcher Will Depue—released a report detailing digital forensic tests on six publicly available Chinese AI models released since April. The models examined include DeepSeek's V4-Flash and V4-Pro, Moonshot AI's Kimi K3 and K3.1-Pro, MiniMax's M3, and Ant Group's Ling-3.1-Flash. According to the researchers, the models' outputs revealed numerical patterns and initialization fingerprints unique to Nvidia's Blackwell architecture. Specifically, DeepSeek-V4-Flash and MiniMax-M3 matched the compute characteristics of B200 and B300 chips, while four models, including Kimi K3 and DeepSeek-V4-Pro, aligned with GB200 and GB300 systems paired with Nvidia's Grace CPU. The Blackwell series, launched in February 2025, served as Nvidia's flagship AI accelerator until its successor, Rubin, hit the market in July 2026. While the Trump administration permitted limited exports of Nvidia's older Hopper-based H200 chips to China, it has never approved the sale of Blackwell or Rubin processors. Unemployment Labs estimated that the Chinese AI projects relied on approximately 12,000 Blackwell GPUs, with chip procurement and cluster setup costs reaching $500 million (roughly 670 billion KRW) for final training runs alone. In South Korea, where semiconductor supply chains and US-China tech rivalry are closely watched, the report has garnered widespread attention—especially given its timing. It arrives just days after former Nvidia CEO Jensen Huang was dismissed on November 12 following revelations of extramarital affairs, including one with Alice Huang (Huang Le), who is suspected of being a Chinese spy. Authorities suspect she exploited her relationship with the former CEO to funnel restricted AI chips into China through Singapore-based Megaspeed and its Malaysian subsidiary. Still, the researchers were careful to note the limits of their findings. They emphasized that their conclusions are based strictly on hardware traces embedded in the AI models themselves, rather than physical verification of where the chips are located or direct legal proof of export-control violations. Industry experts have long suspected that Chinese firms either rent Nvidia compute remotely through Southeast Asian data centers or obtain the hardware via smuggling networks.
Sponsored · Ad