DeepSeek Is Building Its Own AI Chip for Inference
The Chinese startup behind last year's viral efficient models is designing custom silicon for inference, the part of AI that answers user queries, joining OpenAI, Anthropic, Alibaba, and Baidu.
The Chinese startup behind last year's viral efficient models is designing custom silicon for inference, the part of AI that answers user queries, joining OpenAI, Anthropic, Alibaba, and Baidu.
DeepSeek, the Chinese AI startup whose ultra-efficient models went viral more than a year ago, is developing its own AI chip. The chip is designed for inference, the stage where an already-trained model generates answers, not for training new models, according to three people familiar with the reporting by DealStreetAsia.
Inference is where AI's cost stack actually accumulates. Training a frontier model is a one-time capital bill; running one means serving billions of queries a day and paying for every token a model generates. For a startup built on doing more with less, owning the inference layer is a structural bet: cost control on the part of the stack where compute bills recur.
DeepSeek is not the only one reaching for inference silicon. OpenAI has unveiled Jalapeno, its first custom inference chip, built with Broadcom. Anthropic has weighed building its own AI chips. In China, Alibaba (9988.HK) and Baidu (9888.HK) are already developing their own. DeepSeek is moving in the same direction without saying so on the record.
According to the DealStreetAsia report, the program began about a year ago and DeepSeek has reached out to external chip-design, foundry, and memory partners without signing any of them. The company has privately increased hiring of chip-design engineers, none of which has appeared in public job postings. DeepSeek did not respond to a request for comment.
The most immediate pressure point is Huawei. U.S. export bans on Nvidia's top chips to China helped Huawei capture roughly half of an estimated $50 billion domestic AI-chip market, supplying DeepSeek and other leading Chinese AI players, per the reporting. Huawei's share is already slipping as Alibaba and Baidu bring their own silicon to market. A DeepSeek chip, even one that never matches Nvidia on raw performance, would deepen that erosion by giving one of China's most-watched AI labs a reason to buy less from the two incumbents.
DeepSeek has not publicly named a foundry or memory partner, and the reporting describes outreach, not a signed deal. The choice matters because U.S. export controls on advanced chipmaking equipment, the same rules that pushed Huawei into the Chinese AI-chip market, set a ceiling on what any Chinese-designed AI chip can reach. A trailing-edge process node could remain cost-competitive for inference, where memory bandwidth and per-query efficiency matter more than peak training throughput, but it would not close the gap to Nvidia's most advanced silicon.
Nvidia's signal landed first: shares slipped about 2% in premarket trading on the report. The investor read was rough but premature. DeepSeek has not shipped a chip, named a partner, or committed to a timeline. It has joined OpenAI, Anthropic, Alibaba, and Baidu in reaching for the inference layer.
Three things to watch: a confirmed foundry or memory partner, any disclosed specifications, and whether Alibaba's or Baidu's chips ship first and set the benchmark a DeepSeek part would have to clear.