Nvidia Corp. on Wednesday challenged allegations that Chinese artificial intelligence (AI) startup DeepSeek obtained thousands of smuggled Blackwell chips to develop its next-generation model, calling the reported scheme "far-fetched" while vowing to pursue any substantiated tips. The chipmaker responded to a report by The Information claiming DeepSeek accessed several thousand of the U.S.-banned processors through an elaborate smuggling operation involving phantom data centers across Southeast Asia.
AI Generated Image
"We haven't seen any substantiation or received tips of 'phantom data centers' constructed to deceive us and our [original equipment manufacturer] partners, then deconstructed, smuggled and reconstructed somewhere else," an Nvidia spokesperson said in a statement Wednesday. The comments came hours after The Information published details of the alleged smuggling ring, citing six unnamed sources with knowledge of the matter.
The report reignites concerns about enforcement of U.S. export controls on advanced semiconductors, which Washington has positioned as critical to maintaining technological superiority over China. DeepSeek gained global attention in January when its R1 reasoning model rivaled Silicon Valley's best offerings at a fraction of the reported cost, demonstrating China's ability to advance AI capabilities despite restrictions on accessing cutting-edge chips.
The allegations come as the Trump administration navigates a delicate balance between restricting China's access to advanced chips and preserving Nvidia's business interests. Trump announced Monday that Nvidia could ship its H200 chips—one generation older than Blackwell—to approved Chinese customers, with the U.S. government receiving 25% of those sales.
Reported Smuggling Operation
According to The Information's report, the smuggling scheme involves shell companies purchasing Nvidia servers for data centers in Southeast Asian countries where sales are permitted. After Nvidia's Original Equipment Manufacturer(OEM) partners inspect and verify the installations comply with export regulations, smugglers allegedly dismantle entire data centers rack by rack and ship graphics processing unit (GPU) servers in suitcases across borders into mainland China.
Sources familiar with smuggling operations told The Information that operatives prefer eight-chip rack servers like the HGX B200 over the powerful 72-chip GB200 NVL72 systems for their smaller size and ease of covert transportation. The elaborate method ensures no paperwork can be traced to the end user in China, according to four people familiar with chip smuggling cited in the report.
The Justice Department announced Monday that three businessmen were charged with a scheme to smuggle advanced Nvidia AI chips to China, Hong Kong and other countries in violation of U.S. export laws. One has pleaded guilty. However, prosecution remains difficult. "The burden of proof for enforcing and prosecuting chip smuggling deals is fairly high. There needs to be clear and compelling evidence," said Jacob Feldgoise, senior data research analyst at Georgetown University's Center for Security and Emerging Technology.
DeepSeek's Chip Requirements
DeepSeek's development priorities align closely with Blackwell's technical strengths, particularly for pre-training compute. In a whitepaper released December 2 on DeepSeek V3.2, the company identified pre-training compute as a bottleneck compared to frontier models like Gemini-3.0-Pro, stating plans to address this "knowledge gap in future iterations by scaling up the pre-training compute."
The startup has focused on sparse attention methods, which use only certain parts of a model rather than the entire system to generate responses. Blackwell chips include specialized hardware designed to accelerate sparse computing, running such calculations up to twice as fast as traditional methods, according to one person with direct knowledge of DeepSeek's development process cited by The Information.
DeepSeek trained its R1 model using 2,048 Nvidia H800 chips stockpiled by its parent company High-Flyer Capital Management before U.S. export restrictions took effect in 2022. Despite Beijing's push to use domestic alternatives, reports in August indicated Huawei's Ascend GPU servers proved unable to run necessary training workloads, prompting a return to Nvidia hardware. DeepSeek did not respond to requests for comment.
DeepSeek's Chip Requirements
DeepSeek's development priorities align closely with Blackwell's technical strengths, particularly for pre-training compute. In a whitepaper released December 2 on DeepSeek V3.2, the company identified pre-training compute as a bottleneck compared to frontier models like Gemini-3.0-Pro, stating plans to address this "knowledge gap in future iterations by scaling up the pre-training compute."
The startup has focused on sparse attention methods, which use only certain parts of a model rather than the entire system to generate responses. Blackwell chips include specialized hardware designed to accelerate sparse computing, running such calculations up to twice as fast as traditional methods, according to one person with direct knowledge of DeepSeek's development process cited by The Information.
DeepSeek trained its R1 model using 2,048 Nvidia H800 chips stockpiled by its parent company High-Flyer Capital Management before U.S. export restrictions took effect in 2022. Despite Beijing's push to use domestic alternatives, reports in August indicated Chinese telecom giant Huawei's Ascend GPU servers proved unable to run necessary training workloads, prompting a return to Nvidia hardware. DeepSeek did not respond to requests for comment.
Export Control Challenges
The allegations underscore difficulties the U.S. faces enforcing export controls in the globalized semiconductor industry. Most Nvidia chips are manufactured in Taiwan and distributed through a complicated network of resellers worldwide, making tracking and enforcement complex.
Reuters reported Wednesday that Nvidia has developed a software feature to track chip locations, a capability that could combat smuggling if implemented. The U.S. banned export of Blackwell chips—which Trump described in October as "10 years ahead of every other chip"—to China as part of efforts to maintain AI leadership. Nvidia began shipping Blackwell in late 2024 to customers including Google, Microsoft and OpenAI.
The Trump administration's recent decision to permit H200 sales to China with a 25% revenue share for the U.S. drew criticism from some Republicans and marks a policy shift from complete export isolationism. The move reportedly reflects both Nvidia's lobbying efforts and concerns about Huawei's developing CloudMatrix 384 and Ascend 910C systems, which some claims suggest match certain H200 performance metrics.
Explore more exclusive insights at nextfin.ai.
