DeepSeek Reportedly Builds Its Own AI Inference Chip to Reduce Reliance on NVIDIA and Huawei

DeepSeek is reportedly developing its own AI chip as China’s artificial intelligence ecosystem continues moving toward deeper hardware independence. According to Reuters, 3 people familiar with the matter said the Chinese AI company is working on an in house chip designed for inference, the stage where trained models generate responses for users. The report was also amplified in the hardware community through Underfox3, as the industry continues watching how Chinese AI firms respond to tighter access around advanced NVIDIA accelerators.

"developing its own AI chip"
—Reuters

The move does not mean DeepSeek is ready to replace NVIDIA or Huawei across its full AI stack. Reuters describes the project as being in an early stage, with DeepSeek reportedly reaching out to chip design, foundry, and memory partners while recruiting engineers privately. That makes this a strategic direction rather than a near term product launch. Still, the timing is important because DeepSeek has become one of China’s most visible AI companies, and its hardware choices now carry wider geopolitical and commercial meaning.

DeepSeek has already relied on both NVIDIA and Huawei at different points. The company previously gained global attention with efficient models trained on NVIDIA hardware, while its newer model strategy has increasingly aligned with domestic Chinese infrastructure. Reuters reported earlier this year that DeepSeek V4 was adapted for Huawei Ascend chips, and Huawei later said its Ascend SuperNode infrastructure based on Ascend 950 series chips would support DeepSeek V4 inference. That shift showed how quickly China’s AI stack is forming around local models, local accelerators, and local deployment platforms.

At the same time, NVIDIA remains deeply involved in DeepSeek V4 performance outside China’s restricted hardware environment. NVIDIA published official guidance for building with DeepSeek V4 on Blackwell, highlighting support for the 1.6T parameter DeepSeek V4 Pro model and the 284B parameter DeepSeek V4 Flash model, showing how software optimization, TensorRT LLM, Dynamo, NVFP4, NVLink, and model serving improvements can rapidly change token economics after a model launches.

That is exactly why DeepSeek’s reported chip project matters. Inference is now one of the most important battlegrounds in AI infrastructure. Training gets the spotlight because it requires massive clusters and expensive accelerators, but inference is where models are served repeatedly to users, developers, agents, and enterprise systems. If a company can lower inference cost, improve availability, and tune silicon around its own model architecture, it can gain long term control over margins and deployment scale.

DeepSeek is not alone in that direction. Alibaba, Baidu, Meta, Google, Amazon, Microsoft, OpenAI, and other major AI players are all investing in custom silicon or dedicated accelerators because the industry does not want to rely only on general purpose GPUs forever. NVIDIA still has the strongest full stack advantage, but the rise of in house inference chips shows that large AI companies want more control over cost, supply, and model specific optimization.

The China angle makes this even more urgent. US export controls have restricted China’s access to the most advanced NVIDIA accelerators, and Reuters recently reported that China may allow selected firms such as Alibaba, ByteDance, and DeepSeek to buy limited quantities of NVIDIA H200 chips. However, Blackwell class GPUs remain highly restricted for China, and access to top tier AI compute continues to be shaped by policy, licensing, supply limits, and local security concerns. This environment gives Chinese firms a powerful reason to build more domestic alternatives.

Huawei is currently the strongest local alternative, but moving from Huawei dependence to in house silicon would give DeepSeek more control over its own future. The tradeoff is difficulty. Designing an inference chip is only one part of the challenge. DeepSeek would still need a stable manufacturing path, advanced packaging access, memory supply, compiler support, runtime software, model serving tools, reliability validation, and enough production volume to make the chip economically worthwhile.

Recent research also shows how hard non NVIDIA deployment can be in practice. A new field study on Huawei Ascend inference found that serving large model workloads required multiple software patches, disabled features, and additional safeguards to preserve reliability and correctness. That does not mean domestic accelerators cannot work, but it does show that ecosystem maturity remains a major barrier when moving away from CUDA and NVIDIA’s long established software base.

Jensen Huang’s warning that NVIDIA’s China AI share had fallen to zero, with export controls pushing Chinese companies to form a domestic AI stack. DeepSeek’s reported inference chip is one of the clearest examples of that shift. It does not threaten NVIDIA globally overnight, but it shows how restrictions, supply pressure, and AI demand are pushing major players to design hardware around their own models.

DeepSeek building an inference chip would be a logical next step, but it should not be overread as an instant NVIDIA killer. NVIDIA’s advantage is not only the GPU. It is CUDA, networking, NVLink, TensorRT LLM, NIM, Dynamo, developer access, deployment recipes, and years of production experience across cloud scale AI systems. That ecosystem is hard to replicate, especially for a chip that has not yet reached commercialization.

The bigger story is that AI companies no longer want to be only software companies. The economics of inference are becoming too important. Every token generated costs power, memory bandwidth, compute, and infrastructure capacity. If DeepSeek can align its own models with its own accelerator, even only inside China, it could reduce cost and gain more independence from both NVIDIA and Huawei.

For NVIDIA, the risk is not losing the whole market to one DeepSeek chip. The risk is fragmentation. If every major AI company builds specialized silicon for its own models, the market becomes less dependent on one universal accelerator path. NVIDIA can still lead that world, but it will need to keep proving that its full stack is faster, cheaper, and easier to deploy than custom silicon alternatives.

For Huawei, the message is also clear. Domestic leadership in China is not guaranteed forever. If DeepSeek, Alibaba, Baidu, and others move deeper into in house chips, Huawei may become a bridge toward domestic AI independence rather than the final destination.

Do you think DeepSeek’s reported inference chip can become a real alternative to NVIDIA and Huawei inside China, or will software ecosystem and manufacturing limits keep NVIDIA ahead?