DeepSeek, the Chinese startup developing large language models that are competitive with those from OpenAI and Anthropic, has been working for roughly a year on entering the AI chip business, according to Reuters, which cited three people familiar with the matter. The company has been meeting with potential hardware and silicon partners and is hiring engineers for the project.
Key takeaways
- DeepSeek has been working on its own silicon for about a year — currently at the partner meeting and hiring stage
- Focus is on inference chips, not training
- Primary goal: reduce dependence on Nvidia (blocked by US export controls) and Huawei (~50% of China DC chip market)
- Alibaba and Baidu are pursuing similar initiatives in China
- OpenAI and Anthropic are running parallel custom-chip projects in the US
Export controls, Huawei, and the infrastructure gap
Nvidia supplies the majority of AI chips for companies in North America and Europe, but US export controls have blocked it from achieving a similar position in China. Huawei has filled part of that gap, controlling roughly half of the Chinese data center chip market for AI. But Huawei's processors are not a full substitute for Nvidia's H100 and H200 — the performance gap is real, and the software ecosystem around Huawei chips is thinner.
DeepSeek AI joins a growing list of Chinese tech companies working on custom silicon. Alibaba and Baidu have been making moves in this direction for some time. Nor is this dynamic limited to China: in the US, OpenAI and Broadcom jointly announced Jalapeño, a chip designed for inference at scale, a few weeks ago. Anthropic has also been exploring custom chip design, though without any publicly confirmed milestones.
Reuters' account places the project at an early stage: partner meetings and engineering hires are underway, but the company has not shared a timeline or target production date. DeepSeek has not confirmed or denied the project. The reporting rests on three anonymous sources.
Inference, not training — a critical distinction
The focus on inference chips rather than training chips is technically and economically significant. Training large models demands massive parallelism, high memory bandwidth, and fast interconnects between accelerators — which is why Nvidia's specialized GPUs dominate that workload. Inference is a different profile: lower peak compute requirements, but much higher query volumes and strict latency targets.
| Dimension | Training | Inference |
|---|---|---|
| Parallelism | Massive | Lower |
| Memory bandwidth | High | Lower |
| Query volume | Lower | Very high |
| Priority | Raw compute | Low latency |
Building custom inference chips would give DeepSeek hardware-level independence for the operation it is already running at scale: serving model responses to millions of users. The cost savings and the elimination of supply-chain risk — whether from export controls or from Huawei's constrained availability — are concrete, immediate motivations for the investment.
DeepSeek gained wide attention in early 2025 when it released models R1 and V3, which achieved competitive results at a fraction of the training budget spent by rivals in the US. That philosophy of doing more with less is a natural fit for hardware: a company that learned to train models on a lean infrastructure stack has a strong incentive to optimize the inference layer as well.
East-West symmetry
The parallel between China and the US is worth noting. In both markets, leading AI companies have concluded that dependence on a single chip supplier creates strategic risk. In the US, the motivation is partly financial: Apple-style vertical integration gives margin advantages and insulates companies from Nvidia's pricing cycles. A second driver is competition for data center capacity, where owning the silicon layer improves a company's negotiating position as compute demand grows faster than supply.
In China the argument is sharper and more immediate: the export ban is a fact, and Huawei is not a complete substitute for Nvidia. DeepSeek, which built its reputation on efficiency under constraint, is applying the same logic to hardware. The question is not whether the move makes sense but when it will produce results and which partners will be involved.
Why it matters
DeepSeek's move into silicon marks a maturation of China's AI sector — from infrastructure importer to a player that wants to control its own stack. For Huawei, it is a challenge from within the market it just built. For Nvidia, it is further evidence that its AI dominance is being eroded from both directions simultaneously: custom chip projects in the US and a push for domestic alternatives in China.
For the global AI industry, the broader pattern is a fragmentation of the hardware layer. As more companies develop their own silicon, the AI chip landscape will become less unified than it is today. In a five-year view, that means meaningfully different infrastructure architectures across regions — with corresponding differences in the cost structure, performance envelope, and software compatibility of AI deployments.
What's next?
- Reuters places the project at an early stage — no confirmed partners, timeline, or production date have been announced
- The OpenAI/Broadcom Jalapeño chip will be the first public test of whether custom AI inference silicon delivers on its promise — results will inform the market's view of DeepSeek's project
- Any changes to US export control policy could accelerate or slow Chinese silicon initiatives significantly
