Deepseek is facing fresh scrutiny after a report said the company is training its next major model with thousands of Nvidia chips that are banned from export to China under US regulations.
According to The Information, which cited six people familiar with the matter, the chips at the center of the report are powerful Blackwell processors. The claim places Deepseek, Nvidia hardware and the broader US-China AI rivalry back in the same frame: advanced models still need advanced compute, and access to that compute remains politically sensitive.
What the report says happened
The Information’s report describes a route that begins outside China. According to the sources cited in the report, the Blackwell chips were first shipped legally to data centers in Southeast Asia, where Nvidia or its partners installed and tested them.
From there, dealers allegedly took the systems apart. The servers were then moved through customs under false pretenses and reassembled at Chinese facilities, according to the report.
The alleged operation reportedly uses smaller eight-chip servers. The reason given is practical: those systems are easier to transport and maintain.
The core claim is not simply that Deepseek has access to Nvidia hardware. It is that the company is reportedly relying on thousands of advanced chips whose export to China is restricted, and that the hardware may have reached Chinese facilities through a workaround involving dismantled servers and third-country data centers.
Nvidia’s response and the tracking question
Nvidia told Bloomberg that it had no evidence of the so-called phantom data centers described by The Information. At the same time, the company said it follows up on any reports of possible smuggling.
Reuters has also reported that Nvidia is developing a feature to track the location of Blackwell chips and future models. If that kind of technology is widely adopted, it could make the use of smuggled hardware in China significantly harder.
That point matters because the alleged supply chain described in the report depends on opacity. If chip locations can be tracked more effectively, then moving hardware from approved sites to restricted locations becomes a higher-risk operation.
The report therefore raises two linked questions. One is whether export controls can be enforced when high-value AI chips move through global data center networks. The other is whether hardware makers can add technical measures that make diversion more difficult after shipment.
Why Blackwell matters for Deepseek
The report is politically sensitive partly because it suggests Deepseek still depends on high-end Nvidia hardware to compete. That runs against portrayals that the company can already rely primarily on other sources of compute.
The source article notes that Deepseek has experimented with Huawei chips and likely has a stockpile of older Nvidia A100 and Hopper units. Even so, the new Blackwell processors are reportedly essential for its upcoming model.
The reason is tied to how Deepseek runs its AI systems. The Blackwell chips include specialized hardware designed to accelerate sparse attention, a technique Deepseek uses to reduce running costs by activating only parts of the model at a time.
Previous reports also suggest Deepseek has struggled with the performance of Huawei chips. Taken together, those points make the Blackwell processors important not just as raw hardware, but as infrastructure that may fit the way Deepseek is trying to build and operate its models.
The geopolitical stakes
The second sensitive part of the case is political. The source article says the situation gives the Trump administration more leverage in its tech conflict with Beijing and weakens the argument that China can quickly move its AI industry to domestic silicon.
Lawmakers in Washington have already labeled Deepseek a national security threat in a House China Committee report. That report accused the company of violating export rules.
Deepseek also puts pressure on Western AI labs by releasing similarly capable models as open source, including DeepseekMath V2. That makes its hardware access more consequential: the company is not only building models for itself, but also influencing the broader competitive field.
The criticism around Deepseek and other Chinese models is not limited to performance. The source article says these models, along with US models facing scrutiny in Europe, are criticized because beyond their technical capabilities, they convey political and cultural values shaped by state censorship.
What to watch next
The immediate issue is whether the reported smuggling route can be verified, disrupted or deterred. Nvidia’s statement to Bloomberg leaves room for investigation, while Reuters’ report on location tracking points to a possible technical response.
For Deepseek, the larger question is whether it can keep advancing without relying on restricted high-end Nvidia chips. The report suggests that Blackwell processors are important for its next major model, while Huawei chips and older Nvidia A100 and Hopper units may not fully solve the same problem.
For the wider AI market, the case shows how closely model progress, chip supply and export policy are now connected. Training frontier AI systems is not only a software challenge. It is also a contest over who can obtain, monitor and control the hardware needed to build them.