Researcher claims Chinese AI models were trained on banned Nvidia Blackwell chips
Tracking: Export-controlled Blackwell chip access by Chinese AI labsPolicy matter over whether Chinese frontier AI developers such as DeepSeek, Moonshot, MiniMax and Ant Group obtained or trained on export-controlled NVIDIA Blackwell-class GPUs, including forensic evidence from model weights, Commerce export rules under 15 CFR 742.6, enforcement against diversion networks, and chip export policy responses.
- Earliest seen
- Nov 11
- Unemployment Labs · Wednesday, Nov 11
- Articles · 30d
- 4
- 1.3 per story
- Outlets
- 4
- Most: Scientific American (1)
- Hill · 30d
- 0
- None in 30 days or ahead
- Contracts · 30d
- 0
- No awards or solicitations
- Momentum
- 0.1
- Articles ▼2 vs prior week
Activity
30 days back, 14 ahead · hollow = scheduledWhat's happening
Updated yesterdayAn independent weight-fingerprinting study by Will DePue's Unemployment Labs says six recent models from DeepSeek, Moonshot AI, MiniMax and Ant Group likely trained on export-banned Blackwell GPUs. The study itself says it does not establish location, acquisition or legal violation. No upcoming hearings or votes appear in the items; OpenAI is separately touting defenses against adversarial distillation and backing chip-security bills.
Latest
2d agoOur first signal
Hill
From news coverageJim Banks and Elizabeth Warren are cited as supporting the AI Overwatch Act; no scheduled Hill action appears in the items.
Contracts
From news coverageNo solicitations or awards in the items; coverage notes Nvidia's China revenue exposure, the Megaspeed smuggling scandal and Jensen Huang's departure from Nvidia.
Pressure
4 articlesOpenAI backs the Chip Security Act, AI OVERWATCH Act and MATCH Act plus stronger anti-distillation protections; analysts such as Sam Bresnick and Culper criticize offshore chip access and Nvidia's China exposure.
Who covered it, when
A dot per article · ringed = earliest in our sourcesForces at work
share of articles- Threats & findings3 · 60%Weight-initialization fingerprints are said to show six Chinese models likely trained on Blackwell chips, with outside experts split on the method and the study disclaiming proof of a legal violation.Will DePue · Unemployment Labs · Rohith Kuditipudi · Tor Aamodt
- Technical milestones1 · 20%OpenAI reports its defenses cut reasoning-extraction attack success on GPT-6.2 Astra from 54.0% to 0.2%, aimed at adversarial distillation by rivals.OpenAI
- Oversight1 · 20%Coverage links the forensic study to the Megaspeed smuggling scandal and concerns over offshore data center access and Nvidia's China revenue.Sam Bresnick · Culper · Unemployment Labs