DeepSeek Targets $50B AI Chip Market with In-House Inference Silicon

For AI practitioners, inference is where the rubber meets the road: it accounts for the vast majority of operational costs as models scale. DeepSeek, already renowned for squeezing maximum performance from minimal compute, is now taking the next logical step—building custom silicon to optimize the full stack. This could redefine the economics of deploying large language models and set a precedent for software-centric AI companies moving into hardware.

Chinese AI startup DeepSeek is developing its own inference chip, marking a dramatic pivot from pure model development to vertical integration. The project, still in early stages, aims to reduce the company’s dependence on Nvidia and Huawei—the two dominant forces in AI silicon. According to three anonymous sources cited by Reuters, the chip is being designed specifically for inference, the phase where trained models generate responses, rather than for training. This distinction matters: inference workloads represent the bulk of operational AI costs at scale, and custom silicon tailored to DeepSeek’s efficient models could drastically reduce latency and per-query expenses. If successful, DeepSeek would join a growing list of Chinese tech giants—including Alibaba and Baidu—that are designing their own AI chips to sidestep U.S. export restrictions and capture greater value from the AI stack.

“

DeepSeek has almost no chance of selling silicon outside of China unless it gets access to leading-edge manufacturing,” Richard Windsor of Radio Free Mobile said, effectively dismissing the global threat to Nvidia.

The strategic significance cannot be overstated. DeepSeek first gained international prominence over a year ago by releasing two highly efficient AI models that rivaled Western counterparts while using far fewer computational resources. That efficiency ethos now extends to hardware. By controlling both the algorithm and the chip, DeepSeek could offer a fully optimized AI service that outperforms competitors on cost and speed. However, chip design is a notoriously capital-intensive and expertise-dependent endeavor, and DeepSeek’s quiet recruitment of chip engineers—without public job postings—suggests both determination and the challenge of scaling quickly. The company has spent about a year working with foundries and memory suppliers, indicating that it is serious about moving beyond proof-of-concept, but access to cutting-edge manufacturing nodes remains a major hurdle given U.S.-led export controls on advanced semiconductor equipment.

The development unfolds against the backdrop of a $50 billion AI chip market in China, reshaped by U.S. sanctions. Huawei, whose own Ascend chips are less advanced than Nvidia’s top products, has capitalized on the restrictions to capture roughly half of that market, supplying DeepSeek and other domestic AI developers. A shift in which DeepSeek becomes a chip competitor rather than a customer could erode Huawei’s market share and intensify price pressure. Meanwhile, Nvidia—already effectively locked out of China—saw its shares slip about 1.6% in premarket trading, but analysts were quick to downplay the impact. “Nvidia is at zero in China and staying there. DeepSeek has almost no chance of selling silicon outside of China unless it gets access to leading-edge manufacturing,” Richard Windsor of Radio Free Mobile said, effectively dismissing the global threat to Nvidia.