China's DeepSeek Trained on Banned Chips, Now Locks Out US Companies

China's DeepSeek is launching its V4 AI model in April — trained on chips the US banned from export, then locking American chipmakers out of the optimization process entirely.

US officials confirmed DeepSeek trained V4 on Nvidia's Blackwell chips, the most advanced AI processors currently prohibited from export to China. The chips are believed to be located in a data center in Inner Mongolia. That's a potential violation of US export controls, but here's where it gets sharper: after training on American hardware, DeepSeek gave Huawei and other Chinese chipmakers several weeks of early access to optimize V4 for their processors. Nvidia and AMD? Shut out completely.

This isn't just about one AI lab bending the rules. It's the playbook: acquire restricted technology through channels the US hasn't closed yet, use it to build competitive models, then lock out the companies whose hardware made it possible. US labs like OpenAI, Anthropic, and Google are now bracing for a release that could rattle the market again, just like DeepSeek's R1 model did earlier this year.

The irony here is thick. The US spent years building an export control architecture meant to keep advanced AI chips out of Chinese hands. DeepSeek's V4 proves those controls have gaps you can drive a data center through — literally. Singapore has emerged as a known transshipment point, but Inner Mongolia suggests the supply chain is more complex than regulators assumed.

And now the retaliation. By withholding V4 from US chipmakers for optimization, DeepSeek is forcing Nvidia and AMD to compete on Chinese silicon's terms when the model hits the market. That's a deliberate competitive disadvantage: Huawei's chips will run V4 better on day one because they had weeks to tune performance. American hardware? Playing catch-up.

The April launch date matters because it lands right before the summer conference season, when AI labs typically showcase their capabilities. DeepSeek is timing this to maximum effect — proving China can build frontier models despite US restrictions, then demonstrating those models run better on Chinese chips than American ones. That's not just technological competition. It's narrative warfare.

The Albis Perception Gap Index scored the DeepSeek chip controversy 6.4 in February, with US media framing it as sanctions evasion and national security threat, while Asia-Pacific coverage portrayed it as Chinese innovation triumph. The gap reveals how the same facts feed opposite stories: rule-breaking versus resourcefulness, depending on which side of the Pacific you're reading from.

US export controls were supposed to slow China's AI development by choking off access to cutting-edge hardware. Instead, they've accelerated China's push for silicon independence while exposing how leaky the control regime actually is. DeepSeek didn't just circumvent the ban — they trained a competitive model, then used it to disadvantage the companies whose chips made it possible.

The question isn't whether DeepSeek broke the rules. US officials have already confirmed that. The question is what happens when the export control strategy turns out to have more holes than enforcement, and China's AI labs learn to weaponize that gap against the companies the controls were meant to protect.

April's going to be interesting.