DeepSeek built AI models that matched the best American ones while spending a small fraction of what US labs spent. Now it wants to build the chip those models run on.
According to Reuters, the company is in early discussions with chip design, manufacturing, and memory firms, and has been quietly adding chip engineers to its team. No public job postings. The goal is an inference chip: hardware specifically designed for the moment when an AI model responds to a user, rather than the much longer and more expensive process of training the model in the first place.
That distinction matters. Training a model happens once. Inference happens billions of times a day, every time someone sends a message and gets a reply. As AI use grows, inference costs become the biggest ongoing expense for any company running AI at scale. A chip built for that one job, and nothing else, can do it more cheaply than a general-purpose chip built to handle everything.
This is exactly why every major AI lab is now moving in the same direction. OpenAI unveiled its first custom inference chip, called Jalapeño, built with Broadcom, in late June. Google has been building its own chips for a decade. Amazon has its own too. The pattern is consistent: once a lab reaches a certain scale, buying chips from Nvidia stops making economic sense and building your own becomes worth the effort.
DeepSeek faces a harder version of that path. US export controls bar Chinese companies from accessing the most advanced chip factories outside China, and separate restrictions have cut off China's supply of the high-speed memory that inference chips require. DeepSeek's founder said in 2024 that chip restrictions force Chinese companies to use two to four times the computing power to achieve what US labs do with the best hardware. Designing your own chip is partly a response to that reality.
It is also worth noting the context inside China. Huawei currently controls roughly half of the domestic AI chip market, partly because US sanctions on Nvidia have left a gap it can fill. But Alibaba and Baidu are already building their own chips and taking share from Huawei. DeepSeek entering that race would put pressure on Huawei from another direction.
Meanwhile, DeepSeek has also abandoned one of its founding principles. The company had operated since 2023 entirely on funding from its founder Liang Wenfeng's quantitative hedge fund, High-Flyer, and had repeatedly said it had no plans to raise external capital. That changed last month. The company closed a funding round of roughly $7.5 billion, the largest single AI fundraising round in China's history, at a valuation approaching $59 billion. Tencent led with around $1.5 billion; CATL, one of the world's largest battery manufacturers, put in roughly $735 million. The deal was structured unusually: investors put money into a limited partnership managed by Liang himself, not into the company directly, and they have no voting rights and a five-year lock-up.
The reason for finally raising money was not a shortage of cash. It was a people problem. As rival Chinese AI companies went public or saw their valuations surge, DeepSeek's employees holding stock options were watching those options go nowhere without an external valuation to anchor them to. Raising outside capital put a number on the company and gave employees a reason to stay.
The chip project and the funding round are two sides of the same decision. DeepSeek built something extraordinary by staying small, scrappy, and reliant on whatever hardware it could get. That approach has now hit a ceiling. Scaling to the next level requires owning more of the infrastructure and having the money to do it. Both moves point in that direction.