China's AI Hardware Independence
DeepSeek develops inference AI chip, aiming to reduce Nvidia reliance in China
Key Points
V4 training activity incorporated Huawei Ascend chips, with Huawei indicating some V4 training used its hardware, and the April rollout showing a push to optimize and benchmark on domestic hardware.
The project remains in an early stage, with a year-long timeline noted and a marked ramp-up in private hiring of chip-design engineers alongside discussions with foundries and memory suppliers.
The R1 foundation model was trained on Nvidia’s H800, a China-market chip that Washington has blocked export for, highlighting how export controls are shaping DeepSeek’s hardware choices.
The project is described as secretive, with covert recruitment of specialists and engagement with external design firms, foundries and memory makers, underscoring geopolitical considerations and possible shifts in global AI hardware dynamics.
DeepSeek is building its own AI chip, moving beyond software models to take on hardware giants like Nvidia and Huawei. The inference-focused chip project aims to cut the cost of running AI queries and reduce the Chinese startup's dependence on foreign and domestic suppliers alike, according to Reuters.
The project started about a year ago and has quietly picked up speed. DeepSeek has been recruiting chip-design engineers and holding talks with foundries and memory suppliers, according to TechNode. The move signals a major strategic shift — from a company known for efficient AI models to one trying to control its own silicon.
Why Inference Chips Matter More Than Training
AI chips do two main jobs: training models and running them. Training is done once. Inference — answering user queries — happens millions of times a day. DeepSeek's new chip targets inference specifically, TechNode reported. That focus makes sense: as usage scales, inference costs become the biggest expense.
A custom inference chip could give DeepSeek tighter control over cost per query and performance consistency. Other large tech companies, including Google and Amazon, have built their own chips for the same reason. DeepSeek appears to be following that playbook, according to World of Software.
DeepSeek's Messy Hardware Stack Reveals the Problem
Right now, DeepSeek relies on a mixed bag of chips. Its R1 model was trained on Nvidia's H800 — a chip Washington has since blocked from export to China. Its newer V4 model used Huawei Ascend chips for some training, with Huawei itself confirming that involvement, according to Yahoo Finance.
That patchwork setup creates risk. Export controls can cut off Nvidia hardware at any time. Huawei's Ascend chips are improving but still trail Nvidia in performance. Building its own chip would let DeepSeek sidestep both dependencies and tune silicon directly to its software needs.
Secretive Recruitment and Quiet Talks With Foundries
DeepSeek has kept the chip project tightly under wraps. The company has covertly recruited chip-design specialists and engaged with outside design firms, foundries, and memory makers, Reuters reported. The secrecy reflects both competitive pressure and the geopolitical sensitivity around China's access to advanced chip-making tools.
The project is still early-stage. No chip has been produced or announced publicly. But the hiring ramp and supplier talks show the effort is real and growing. Analysts see it as part of China's broader push to build a self-sufficient AI hardware stack amid tightening U.S. export controls.
A Threat to Nvidia's China Business and Global AI Dynamics
Nvidia still dominates AI chip sales globally. But DeepSeek's move is the latest signal that China's AI firms are serious about reducing that dependence. If DeepSeek ships a working inference chip, it could pressure Nvidia's remaining China-market revenue and inspire other Chinese AI labs to follow suit, according to World of Software.
The broader implications stretch further. DeepSeek has already rattled competitors like OpenAI with efficient, low-cost models. Adding custom hardware to that equation could sharpen its edge. The shift from model-maker to chip-maker, if successful, would mark one of the most significant moves in the global AI race so far.
Sentiment
Mostly NegativePublishers
54
Articles
201
Reach
255
