U.S. Says DeepSeek Trained New Model on Nvidia Blackwell Chips, Raising Export Control Alarm

U.S. officials say DeepSeek trained its upcoming model on Nvidia’s Blackwell chips in China, a claim that, if verified, would signal a breach of export controls and intensify the policy split in Washington over AI chip sales.

U.S. authorities have said that Chinese AI startup DeepSeek trained its upcoming artificial intelligence model using Nvidia’s most advanced AI processor, Blackwell.

The development could constitute a breach of U.S. export controls and deepen an already tense debate in Washington over Chinese access to cutting-edge AI technology.

According to a senior Trump administration official who spoke to Reuters, the chips were likely concentrated in DeepSeek’s data center in Inner Mongolia, where the company has reportedly removed technical indicators that might reveal their use. The official emphasized that U.S. policy prohibits shipments of Blackwell processors to China.

The official said U.S. authorities believe Blackwell chips were clustered at DeepSeek’s data center in Inner Mongolia and used to train a model expected to be released as soon as next week. The person declined to disclose how the U.S. obtained the information or how the chips reached China.

The Chinese embassy in Washington said Beijing opposes “drawing ideological lines, overstretching the concept of national security, expansive use of export controls and politicizing economic, trade, and technological issues.” At a regular briefing, foreign ministry spokesperson Mao Ning said she was not aware of the specific circumstances but reiterated China’s longstanding objections to U.S. restrictions on chip exports.

Confirmation by U.S. officials that DeepSeek obtained and used Blackwell chips, first reported by Reuters, is likely to deepen divisions in Washington over how tightly to restrict China’s access to cutting-edge AI hardware.

President Donald Trump has shifted positions over the past year. In August, he signaled openness to allowing Nvidia to sell a scaled-down version of Blackwell in China. He later reversed course, stating that the most advanced chips should be reserved for U.S. companies.

In December, the administration allowed Chinese firms to purchase Nvidia’s second-most advanced AI chip, the H200, drawing criticism from national security hawks. Shipments have since stalled over approval conditions and compliance guardrails.

White House AI adviser David Sacks and Nvidia CEO Jensen Huang have argued that permitting some advanced chip sales to China reduces incentives for Chinese firms such as Huawei to accelerate domestic alternatives that could eventually challenge U.S. technological leadership.

Others take the opposite view. Chris McGuire, who served on the National Security Council under former President Joe Biden, said the development demonstrates the risk of any advanced AI chip exports to China.

“Given China’s leading AI companies are brazenly violating U.S. export controls, we obviously cannot expect that they will comply with U.S. conditions that would prohibit them from using chips to support the Chinese military,” he said.

Blackwell represents Nvidia’s latest-generation AI architecture, designed to power large-scale model training and inference workloads. Its performance gains over prior chips significantly reduce training time for frontier models and lower energy consumption per computation — advantages that can accelerate iteration cycles and narrow competitive gaps.

If DeepSeek successfully trained a major new model on Blackwell hardware inside China, it would suggest either a breakdown in export enforcement, diversion through third countries, or access to previously shipped inventory before controls tightened.

The U.S. official said Washington believes DeepSeek may attempt to remove technical indicators that could reveal the use of American AI chips. Such indicators can include firmware signatures, performance characteristics, or configuration traces embedded in model training logs.

Distillation and model replication

The administration official added that the DeepSeek model likely relied in part on “distillation” of leading U.S. AI systems, echoing prior allegations from OpenAI and Anthropic.

Distillation involves using outputs from a larger, more advanced model to train a smaller or newer model, effectively transferring learned behavior without replicating the original training dataset or architecture from scratch. If combined with access to top-tier hardware like Blackwell, distillation can compress development timelines and reduce compute costs.

Hangzhou-based DeepSeek unsettled global markets last year with a series of AI releases that approached the performance of leading U.S. systems at lower reported training costs. The prospect that it may now have leveraged Blackwell chips — the same hardware underpinning frontier U.S. models — raises the stakes.

Export control credibility

At issue is not only competitive positioning but also enforcement credibility. U.S. export controls are designed to limit China’s ability to train or deploy frontier AI systems with potential military applications. Blackwell is among the most tightly controlled chips due to its capability to handle massive parallel workloads required for advanced AI.

If China-based firms can access such hardware despite restrictions, policymakers may push for tighter secondary sanctions, expanded entity listings, or broader licensing requirements for cloud-based compute services.

At the same time, stricter controls carry trade-offs. Nvidia derives significant revenue from international markets, and curtailing overseas sales can reduce scale advantages and funding for future research. Proponents of selective access argue that engagement preserves U.S. influence over standards and supply chains.

The immediate question is whether Washington adjusts its stance on H200 approvals or broadens enforcement mechanisms.