DeepSeek uses smuggled Nvidia chips for its AI model

China’s AI scene was never going to slow down just because Washington told it to. And DeepSeek—the upstart that embarrassed half of Silicon Valley by training a frontier-class model on pocket change—appears to be proving that point again.

According to The Information, DeepSeek has been building its next-gen model using thousands of Nvidia Blackwell chips that were never supposed to cross China’s border. These GPUs, banned under U.S. export controls, allegedly slipped in through a Rube Goldberg–style smuggling pipeline:ship the servers to a “friendly” country, install and test the hardware, tear everything apart, mislabel the parts, and reassemble it all in China as if nothing happened.

If true, it’s a stunning reminder of something obvious yet often ignored: you can restrict chips, but you can’t restrict ambition.

Why smuggle Blackwells? Because Chinese chips still can’t match Nvidia.

DeepSeek has plenty of patriotic swagger, but its engineers quietly admit the uncomfortable truth—China’s domestic chips still aren’t good enough to train the models they want to build.

Blackwell GPUs, launched in late 2024, aren’t just fast. They’re optimized for sparse computing—precisely the architecture DeepSeek uses in its R1 and V-series models. Sparse attention activates only the useful parts of a model during inference, slashing costs.When you’re trying to out-compute U.S. giants with a fraction of the budget, that matters.

So DeepSeek turned to the only hardware that can keep them competitive… even if they had to go through the AI equivalent of a speakeasy to get it.

Remember: DeepSeek built its reputation by doing more with less.

This is the same startup that shocked the West earlier this year with its R1 model—delivering near-frontier performance on what analysts estimate was a shoestring training budget.

Before the chip clampdown, DeepSeek relied heavily on nearly 10,000 A100s stockpiled by its parent firm, High-Flyer Capital, back when U.S. export restrictions were merely a rumor. They later mixed in Hopper chips and experimented with Huawei hardware on the side. But Nvidia remained the backbone.

Their next model—supposedly aiming for a mid-February internal deadline that founder Liang Wenfeng refuses to confirm—is rumored to be the company’s most ambitious yet.

Washington isn’t amused.

Several lawmakers have now singled out DeepSeek as a national-security concern, accusing it of gaming the system. Washington expected to starve China’s AI labs of compute. Instead, it seems to have built a thriving underground GPU economy.

Meanwhile, President Trump threw a wrench into the narrative on Monday by signaling that H200 sales to China might be allowed—if Beijing approves.

That would punch a hole in the smuggling market overnight.

It would also raise an awkward question: Did we just spend two years building walls around the house only to unlock the front door?

Nvidia, caught in the middle, is trying to plug the leaks.

Nvidia insists it has seen no evidence of “phantom data centers” created solely to slip chips through export controls. But the company also says it’s developing new software to track GPU locations—effectively a digital leash for hardware.

If rolled out, it could prevent future smuggling schemes.But that still wouldn’t answer the bigger question: What happens when the world’s most powerful technology becomes too valuable—and too strategic—to control?