DeepSeek Trained New AI Model on Restricted Nvidia Blackwell Chips

The disclosure that DeepSeek, a prominent Chinese artificial intelligence laboratory, successfully trained its latest model using Nvidia’s state-of-the-art Blackwell architecture represents a significant challenge to the efficacy of U.S. export controls. According to a senior official within the Trump administration, the upcoming model—expected to debut as early as next week—was developed using hardware that is explicitly prohibited for export to China under current Department of Commerce regulations. This revelation underscores the persistent difficulty in sealing off global supply chains from determined state-backed or well-funded private entities in the high-stakes AI arms race.

Nvidia’s Blackwell chips, which offer a massive leap in compute density and energy efficiency over the previous H100 generation, were designed to be the cornerstone of the next wave of generative AI. By restricting these chips, the U.S. government aimed to maintain a 'compute moat' that would keep Chinese AI capabilities several generations behind Western leaders like OpenAI and Anthropic. However, the presence of these chips in DeepSeek’s data centers suggests that either the 'gray market' for high-end silicon is more robust than previously estimated, or that Chinese firms are successfully utilizing cloud-based loopholes and third-party intermediaries to access restricted compute power.

“

The disclosure that DeepSeek, a prominent Chinese artificial intelligence laboratory, successfully trained its latest model using Nvidia’s state-of-the-art Blackwell architecture represents a significant challenge to the efficacy of U.S.

The timing of this disclosure is particularly sensitive as it coincides with separate allegations from Anthropic, which recently accused DeepSeek of 'industrial-scale distillation' attacks. Anthropic claims that DeepSeek systematically harvested data from its Claude models to train its own systems, a technique known as model distillation that allows a smaller or less advanced firm to 'steal' the reasoning capabilities of a superior model. When combined with the news of Blackwell hardware usage, a picture emerges of a multi-pronged strategy by Chinese labs: acquiring the best hardware through illicit channels while simultaneously harvesting the best software outputs from Western competitors.