Three independent sources confirm that DeepSeek has been working for about a year on its own AI inference chip. The Chinese lab wants to reduce its combined dependence on Nvidia and Huawei, against a backdrop of US export rules that keep advanced GPUs out of Chinese hands. The project is still early, with active outreach to partners in design, manufacturing and memory. It fits inside a broader silent race where OpenAI and Anthropic have already opened their own silicon paths.
Key Takeaways
- DeepSeek is building an in-house inference chip, not a training chip, and has been working on it for about a year.
- The stated goal is to reduce the joint dependence on Nvidia and Huawei.
- The lab is looking for partners in chip design, fabrication and memory, under strict US export controls.
Have an AI Sum Up This Article
ChatGPTAn inference chip, not a training one
The choice to target inference rather than training is strategic. Inference is the daily compute that runs a model in production. It is the visible chunk of the GPU bill for a lab that serves millions of daily requests.
Training remains dependent on high-end Nvidia GPUs. An in-house chip meant for training would take extra years of development and access to advanced packaging that Chinese players no longer have since the latest US restrictions. Focusing on inference lets DeepSeek bite into its operational bill right away. For context, see our earlier piece on Horizon: DSpark: DeepSeek Boosts AI Speed 85% on GPUs.
DeepSeek is not alone here. OpenAI has openly discussed an in-house accelerator, Anthropic is weighing options with Samsung and Broadcom. The difference is regulatory pressure. US labs want to optimize unit economics. DeepSeek wants to survive sanctions that keep hardening.
The timeline is unclear. Development has been running for a year, the stage is early, and no major industrial partner has been announced publicly. We have no production date and no confirmed working first prototype.
The industrial wall stands tall
Designing a competitive AI chip typically takes years and heavy capital. The real obstacle for DeepSeek is not design, it is fabrication. US rules bar Chinese designers from using the most advanced foundries in Taiwan and South Korea.
High bandwidth memory, the famous HBM, is the second bottleneck. Without HBM, an inference chip loses all interest against an Nvidia GPU. US curbs cut China off from the newest HBM suppliers. DeepSeek will either have to rely on grey channels or contribute to building a domestic HBM industry, two costly and slow options.
The lab is openly recruiting partners in chip design, semiconductor manufacturing and memory. It is in preliminary talks with several players. The internal team headcount is rising, a signal that suggests a real multi-year commitment rather than a trial balloon.
Political pressure pushes the project forward. Beijing is nudging its tech champions to build domestic alternatives in chips as much as in models. DeepSeek has been the archetype Chinese AI champion since DSpark. The Party has a stake in seeing the lab succeed.
Also on Horizon:
- DeepSeek Builds Its Own AI Chip to Cut Nvidia Reliance
- iOS 27 Finally Lets You Reshape Siri’s Voice
- Nvidia Pushes Its New AI Rack to 2028, Over a Year Late
What it changes in the global AI game
In the short term, this news changes little. DeepSeek will keep running on grey-market Nvidia and Huawei Ascend for its immediate production loads. First in-house silicon will not land before 2027 or 2028 in the most optimistic scenario.
In the medium term, the stake is strategic. An in-house inference chip, even modest in performance, reduces US leverage over Chinese labs. Every inference dollar that no longer flows through Nvidia is one less sanction dollar.
Nvidia is already losing China market share mechanically because of restrictions. A DeepSeek that pulls off its silicon accelerates the pattern for other Chinese labs looking to copy the move. Alibaba, Baidu and ByteDance are watching this file closely.
Huawei is the closest direct rival for DeepSeek in that approach. The Ascend line already plays this role for several Chinese players. A DeepSeek chip aimed at inference nibbles at the pie Ascend was after, and the lab could get better prices for its own needs by negotiating with several internal suppliers.
The wider lesson for US industry is clear. Every restriction cycle speeds up the build-out of a parallel Chinese silicon stack. What Washington hoped to block through sanctions, it ends up accelerating through constraint. The open question is how the first DeepSeek chip actually performs when it ships, and how quickly other Chinese champions follow its lead.
Follow the story on Horizon.
