RTX 5090 cards selling for $5,000 on China's black market as DeepSeek AI models drive demand

In brief: China is in a frenzy to get its hands on the Nvidia RTX 5090 graphics card – but not just because people are desperate for more FPS in their games. The demand is the result of DeepSeek's AI models, which companies realize they can run on Nvidia's consumer GPUs at a fraction of what it would cost them to buy AI accelerators.

DeepSeek's industry-disrupting R1 models can efficiently run on consumer GPUs, rather than expensive AI accelerators such as the H20. As reported by The Information (paywalled article), the situation has led to demand for the China-specific RTX 5090D reaching a point where the cards are being sold on the black market for up to $5,000. And even at that price, the chances of finding one are pretty slim.

It's not just China where the RTX 5090 and 5090D are being bought for more than 100% of their $2,000 MSRP. eBay.com has several listings for the cards ranging between $5,000 and $10,000, many of which are actaully shipped from China. There are also a number of units from US sellers priced between $2,500 and $5,000.

Given the low to non-existent stock levels and the huge demand for the flagship, buying an RTX 5090, or any Blackwell card, directly from a retailer right now is like trying to purchase a card during the height of the pandemic/crypto-boom crisis in 2020 and 2021.

While DeepSeek's R1 with its 671 billion parameters is designed for high-performance environments, the company has developed distilled versions of the model that are more accessible for consumer hardware. The 8 billion parameter Llama 8B, for example, recommends a GPU with 21GB of VRAM.

Last month, Nvidia and AMD were embroiled in a public argument over which of their respective consumer graphics cards offered better performance when running DeepSeek's AI models. Team Red claimed the 7900 XTX outperformed the RTX 4090, prompting Team Green to release benchmarks showing both the 4090 and RTX 5090 are faster – measured with llama-bench – than AMD's card when using Distill Qwen 7B, Llama 8B, and Qwen 32B.