DeepSeek is reportedly developing its own inference chip, a move that shows how China’s AI market is becoming more vertically integrated.
(Source: JPR)
DeepSeek is developing its own AI chip, according to Reuters, in a move that would reduce the Chinese AI company’s reliance on Nvidia and Huawei hardware.
Insiders say that the chip is being designed for inference rather than training. Training still gets most of the attention, but inference is where AI becomes a recurring cost, and we are increasingly seeing companies decide to only compete on inferencing. Every user interaction has to run somewhere, and the economics of that workload are becoming central to the AI business.
According to Reuters, DeepSeek’s chip effort began about a year ago and remains at an early stage. The company has been speaking with external partners, including chip-design, foundry, and memory companies, and has increased hiring of chip-design engineers in recent months. The hiring has reportedly been done privately rather than through public job postings.
For DeepSeek, this makes strategic sense. The company uses both Nvidia and Huawei chips today, but US export controls have limited China’s access to Nvidia’s most advanced GPUs, and even where they are available, it is no longer culturally as acceptable. There is strong government pressure to have a Chinese supplier in your chain.
DeepSeek trained the foundation model behind R1 on Nvidia’s H800, a China-market chip that was later banned by Washington. Since then, the company has leaned more heavily on Huawei’s Ascend platform, including adaptations of its models for Huawei hardware.
A DeepSeek inference chip would put the company into the same broad trend as other major AI model companies seeking more control over their hardware stack. OpenAI, Anthropic, Alibaba, Baidu, and others are all moving toward greater custom silicon involvement. The reason is not fundamentally performance, where the US suppliers still have an edge, rather it is supply and control.
AI cost pressure is already changing customer behavior. DeepSeek’s January 2025 model release briefly shifted global attention toward cheaper Chinese models, and subsequently Chinese AI models’ share of global Web traffic rose from about 3% to 13% within two months. US labs responded quickly with cheaper models of their own, but the signal was clear: If performance is close enough, people will choose based on price.
That said, these lower-cost Chinese models are gaining only limited traction with US businesses. DeepSeek adoption among 70,000 US firms using Ramp rose from 0.1% to 0.3% over the first five months of 2026. That is still tiny, but as AI usage scales, model choice becomes less ideological and more economic, so we expect to see it grow further, and even single-digit growth is significant money in this market.
We should not read this story as only about competition between the West and China. The main effect of this move may be on Huawei, which has benefited from Nvidia’s restricted position in China, but will see its hold on the domestic AI chip market weakening if Chinese hyperscalers develop their own AI processors.
A model company can decide it wants its own chip, but that does not make it a semiconductor company overnight. Designing a competitive AI inference chip is expensive and slow. Manufacturing is also difficult. Chinese chip designers remain constrained by access to leading-edge foundries and high-bandwidth memory, both of which are affected by US restrictions.
Still, the strategy is significant. DeepSeek became famous for model efficiency, not for commercial infrastructure. A move into chip design would suggest that China’s AI leaders increasingly see hardware as part of the core model strategy, and that they do not expect supply-chain problems to go away any time soon.
What do we think?
China is a fascinating parallel market. It has its own active AI ecosystem, its own desktop GPU market, and a growing set of domestic chip suppliers shaped by export controls, national policy, and local demand.
DeepSeek’s chip effort is not just about replacing Nvidia or weakening Huawei. It is about the economics of inference. If Chinese model companies can deliver good-enough performance at lower cost, and then tune domestic silicon around those models, they create a market that looks increasingly distinct from the US-led AI stack.
An interesting opportunity may sit outside China. Much of Africa and Asia remains AI-poor: not because demand is absent, but because cost, cloud access, infrastructure, language support, and sovereign-control concerns make frontier US AI difficult to adopt at scale. Low-cost Chinese models, running on domestic Chinese silicon, could become attractive in those markets if they are good enough, cheap enough, and accessible.
Europe is a different case, but the same question is emerging. The sudden US restriction on foreign access to Anthropic’s most advanced models unnerved customers and governments because it made dependency visible. EU sovereign models remain overly specialized for areas like defense and science. For countries and companies outside the US, the issue is no longer only who has the best model. It is who can be trusted to keep providing access. We think that trust issue may turn out to be more actioned than the trust issue around whether China might snoop on Western data.
Still, we don’t think Europe will simply turn to Chinese AI. Data governance, security, political trust, and regulatory alignment all make that difficult. But it does mean the market for alternatives has become more credible. Sovereign AI, open models, domestic deployment, and non-US infrastructure all look more important than they did before.
In China, there will still be opportunities for Western companies in selected areas, but the larger long-term value may not sit with chip exporters. Five years from now, more of the value may accrue to enabling IP suppliers that help Chinese companies and other regional players build domestic silicon. That points to companies such as Imagination Technologies, which, sometimes controversially, supports local Chinese GPU and accelerator development without being the finished-chip supplier.
DeepSeek shows where the market is heading. In China, as elsewhere, AI is becoming more vertically integrated. The model companies want more control, the chip companies want more of the AI value chain, and IP suppliers may become increasingly important in the middle.
DeepSeek’s entrance into the AIP chip market raises the population count to 150 (this week). We track all the details about the products, companies, funding, sales, and segments in our AI processor Tracking Service. If you are an AIP supplier or interested in the market, you should check it out.
Take a look at our AI library, where, among things, we keep track of the 150 companies offering 292 AI processors. JPR puts the “I” in AI (and because you will ask—Intelligence, market intelligence).
LIKE WHAT YOU SAW HERE? SHARE THE EXPERIENCE, TELL YOUR FRIENDS.