Report: China’s DeepSeek follows OpenAI in developing its own custom inference chips
DeepSeek Ltd., one of China’s most visible artificial intelligence companies, is pushing to design its own, in-house silicon aimed at inference workloads, according to a report by Reuters today.
The report cites three people familiar with the company’s plans as saying that it has been exploring the concept of developing its own AI accelerators for about a year, but the initiative is still at the early stages. So far, DeepSeek has contacted outside partners and held discussions with chip design, foundry and memory companies, while also trying to hire a number of experienced chip designers to join its engineering teams.
The focus on inference chip is telling, because this is the stage where a trained model generates responses for users, and it’s an important workload that AI companies must control if they want to be able to commercialize their most advanced models. While AI training is what made DeepSeek famous, inference is what becomes the recurring cost center that drives revenue once people start using those models.
DeepSeek’s goal seems to be to reduce its reliance on Nvidia Corp. and Huawei Technologies Co. Ltd., which until now have been its two primary suppliers of AI chips. When the company shot to fame in January 2025 with its low-cost yet high-performing R1 model triggering a rout in U.S. technology stocks, it said it had trained that system on Nvidia’s H800 chips. That hardware was designed specifically for the Chinese market and had been throttled compared to what was being sold to U.S. firms, but it was later banned by the U.S. under President Donald Trump’s administration.
