In short: China is in a frenzy to get its arms on the Nvidia RTX 5090 graphics card – however not simply because persons are determined for extra FPS of their video games. The demand is the results of DeepSeek’s AI fashions, which firms notice they’ll run on Nvidia’s shopper GPUs at a fraction of what it might value them to purchase AI accelerators.
DeepSeek’s industry-disrupting R1 fashions can effectively run on shopper GPUs, fairly than costly AI accelerators such because the H20. As reported by The Info (paywalled article), the state of affairs has led to demand for the China-specific RTX 5090D reaching some extent the place the playing cards are being bought on the black marketplace for as much as $5,000. And even at that value, the possibilities of discovering one are fairly slim.
It isn’t simply China the place the RTX 5090 and 5090D are being purchased for greater than 100% of their $2,000 MSRP. eBay.com has a number of listings for the playing cards ranging between $5,000 and $10,000, lots of that are actaully shipped from China. There are additionally numerous models from US sellers priced between $2,500 and $5,000.
Given the low to non-existent inventory ranges and the massive demand for the flagship, shopping for an RTX 5090, or any Blackwell card, straight from a retailer proper now could be like attempting to buy a card throughout the peak of the pandemic/crypto-boom disaster in 2020 and 2021.
Whereas DeepSeek’s R1 with its 671 billion parameters is designed for high-performance environments, the corporate has developed distilled variations of the mannequin which can be extra accessible for shopper {hardware}. The 8 billion parameter Llama 8B, for instance, recommends a GPU with 21GB of VRAM.
Final month, Nvidia and AMD have been embroiled in a public argument over which of their respective shopper graphics playing cards provided higher efficiency when working DeepSeek’s AI fashions. Workforce Pink claimed the 7900 XTX outperformed the RTX 4090, prompting Workforce Inexperienced to launch benchmarks exhibiting each the 4090 and RTX 5090 are quicker – measured with llama-bench – than AMD’s card when utilizing Distill Qwen 7B, Llama 8B, and Qwen 32B.
