DeepSeek Is Building Its Own AI Chips to Dodge US Export Controls

DeepSeek, the Chinese AI startup known for building language models that rival OpenAI's and Anthropic's, is getting into the chip business. According to Reuters, which cited three people familiar with the matter, the company has spent roughly a year working toward this move, meeting with potential hardware and silicon partners while hiring engineers to build out the effort.

The plan centers on data center chips built for inference, meaning the computing work involved in actually running an AI model and generating responses, rather than training new models from scratch. That's a telling detail. It suggests DeepSeek isn't trying to build a rival to Nvidia's most advanced training hardware so much as securing the supply of chips it needs to keep its products running at scale, on its own terms.

The underlying motive isn't hard to figure out. US export controls have blocked Nvidia, the dominant chip supplier for AI companies in North America and Europe, from maintaining anything close to that same footprint in China. That restriction has created an opening that Huawei has largely filled, now commanding around half of China's data center chip market. DeepSeek's push into silicon looks like an attempt to avoid depending too heavily on either company: Nvidia because of the export ban, and Huawei because relying on a single domestic supplier carries its own risks.

DeepSeek isn't alone in reading the situation this way. Chinese tech giants Alibaba and Baidu have also been moving into chip development, all racing to build alternatives as the country's AI ambitions collide with Washington's restrictions. The chip export fight has effectively forced a wave of Chinese companies to treat silicon self-sufficiency as a strategic necessity rather than a side project.

The trend isn't just a China story

What makes this notable is that it's not happening in isolation. US-based AI companies are making remarkably similar moves, just for a different set of reasons. OpenAI and Broadcom recently announced Jalapeño, OpenAI's first chip built specifically for large-scale inference, unveiled only a couple of weeks before the DeepSeek report surfaced. Anthropic has reportedly been exploring its own custom chip work as well, though nothing publicly concrete has emerged yet.

For OpenAI, the motivations run in two directions. One is the familiar goal of reducing reliance on Nvidia, whose chips remain the industry standard and are in constant, heavy demand. The other is more ambitious: building the kind of end-to-end control over hardware and software that Apple has long used as a competitive edge. Owning more of the stack, from silicon up through the model, gives a company leverage over both cost and performance that it doesn't have when renting someone else's chips.

There's also a more practical calculation at play across the industry. Data center capacity is shaping up to be one of the tightest bottlenecks in AI for the foreseeable future, with multiple companies scrambling for compute as they scale up models and services. Getting a foothold at the chip and data center level now is a hedge against a future where that access only gets harder to secure.

DeepSeek's move fits neatly into that same logic, just with an added layer of urgency. Export controls aren't a hypothetical risk for a Chinese company, they're an active constraint already shaping what hardware it can buy and from whom. Designing its own inference chips would give DeepSeek a way to keep scaling its models without being fully at the mercy of Huawei's supply or the shifting boundaries of US policy.

None of this means DeepSeek is on the verge of matching Nvidia's engineering, and the reporting doesn't suggest any product is close to shipping. What it does show is how thoroughly the chip export fight has reshaped the strategic thinking of AI companies on both sides of the divide. In the US, the push into custom silicon is largely about control and cost. In China, it's increasingly about necessity. Either way, the era of AI labs simply buying whatever chips they need off the shelf looks like it's ending, replaced by a scramble to own more of the hardware stack outright.

What happens next will hinge on how far DeepSeek can actually get with a hardware effort still in its early stages, and whether it can find the manufacturing partners needed to turn chip designs into real, shippable silicon.