The Takeaway
• Chinese companies are using Nvidia gaming chips to deploy DeepSeek
• Blackmarket price of Nvidia’s newest gaming chip has soared 150%
• Nvidia CEO Huang said DeepSeek was Nvidia’s most important customer in China
When DeepSeek’s success in developing competitive artificial intelligence models at a low cost prompted a $750 billion sell-off in Nvidia shares in late January, Nvidia CEO Jensen Huang wasn’t worried. He told some employees a few days later that DeepSeek had just become Nvidia’s most important customer in China, since any progress in AI will only increase demand for the company’s chips, according to a person who heard his remarks.
Demand for Nvidia chips in China has been soaring in the last several weeks, but not necessarily in the way Huang predicted. DeepSeek’s skyrocketing popularity has boosted demand for Nvidia’s gaming chips in the underground market as tech firms realize they can run DeepSeek’s models at a fraction of the price of AI chips, according to five chip smugglers.
The black market price for Nvidia’s latest and most powerful gaming chip, RTX 5090, has jumped 150% to more than $5,000 in recent weeks, according to the smugglers, compared to the chip’s $2,000 sticker price. Even at $5,000 apiece, the price is a sixth of the normal black market price for one of Nvidia’s most popular AI chips currently in use, the H100.
The episode is a reminder of the constantly shifting tactics used by buyers in the illicit market for Nvidia’s chips in China, where U.S. export controls ban purchase of the company’s most cutting-edge chips. Tech startups and universities have been buying Nvidia’s AI chips on the black market for more than a year to get around the export bans, using chip smugglers operating out of Singapore, Malaysia and Japan.
But the smugglers have lately shifted their attention to Nvidia’s gaming chips, responding to growing demand for consumer products that can run DeepSeek models, which don’t require the high-end AI chips.
To be sure, demand for Nvidia’s AI chips also surged after DeepSeek’s models went viral. China’s internet giants significantly increased their orders for less powerful Nvidia chips that U.S. regulations allow to be sold to China, according to five people with direct knowledge of the chip procurement. But orders for those chips placed now won’t be delivered for at least two quarters.
Universities, tech startups, and technology departments of state-owned enterprises meanwhile have turned to Nvidia gaming chips, which are cheaper and have the necessary computing power for DeepSeek model reasoning. And they’re available on the black market.
In January, for instance, a tech company in China’s northern Hebei province announced plans to invest approximately $2.6 million in an AI reasoning computing facility comprising initially only Nvidia gaming chips.
Meanwhile, in early February, a research team at China’s prestigious Tsinghua University announced that in collaboration with Chinese startup Approaching.AI they had managed to run DeepSeek’s full-powered models on just one Nvidia gaming chip.
Some other Chinese tech companies are running servers using eight modified Nvidia gaming chips instead of the standard eight AI chips, according to two smugglers and an engineer involved in their construction. These servers significantly lag behind regular Nvidia AI servers in performance, but their affordable pricing and quick availability have made them quite popular in China.
This isn’t the first time Chinese companies have switched to Nvidia’s gaming chips for AI work in response to U.S. export bans, although the surge in gaming chip prices in the black market is unusual.
It’s unclear whether Nvidia’s management is aware that its gaming chips are increasingly used as AI reasoning chips in China, but some of its employees in China certainly are. A couple of Nvidia employees visited an AI studio in China this month, witnessed how an engineer assembled an AI server with eight of Nvidia’s gaming chips and gushed over the craftsmanship showcased in the process, according to the engineer who did the assembly.
An Nvidia spokesperson said gaming cards are not viable for use in data center computing clusters for AI because the products are “designed, manufactured and marketed for individual gamers and consumers.”
Gaming chips are primarily designed for rendering graphics in videogames, but their powerful processing capabilities and architecture make them capable of handling AI tasks.
Before Nvidia developed AI-specific chips, academics and researchers in the early 2010s used the company’s graphics processing units designed for image processing to perform machine-learning tasks.
Just as it does with AI chips, Nvidia uses authorized distributors to sell its gaming chips rather than selling them directly. Distributors often sell some chips to other dealers, according to the five chip smugglers. Nvidia said its gaming chips are sold in compliance with U.S. export control laws. But its layered distribution network can make it difficult for the company to track the smuggling of its chips into China for use in AI data centers.
Gaming chips used to make up nearly half of Nvidia’s revenue until the debut of ChatGPT in late 2022 fueled demand for AI data center chips. But the share of Nvidia’s revenue gaming chips comprise has gradually declined, amounting to less than 10% in the quarter ending on January 26.
These chips typically feature less memory than dedicated AI chips, as they are primarily designed for gaming rather than large-scale AI tasks. To bridge the gap, many dealers in China have resorted to salvaging usable storage modules from older-generation gaming chips and incorporating them into the newer units, said two of the five gaming chip dealers.
This process enhances the storage capacity and data-handling capabilities of the modified chips, enabling them to run larger AI models more effectively.