# DeepSeek is developing its own AI inference chip, Reuters reports > Source: > Published: 2026-07-07 19:58:46+00:00 [DeepSeek](https://www.deepseek.com/) founder Liang Wenfeng is pushing the Hangzhou AI lab into chip design, with Reuters reporting on July 7th that the company is developing an in-house AI chip aimed at inference, citing three people familiar with the effort. The report, shared Tuesday by [Andrew Curran (@AndrewCurran_)](https://x.com/AndrewCurran_/status/2074492384726159725), says DeepSeek's chip work is still early. Reuters reported that DeepSeek has been contacting outside partners and holding discussions with chip-design, foundry and memory companies. One source told Reuters the effort began about a year ago, while two sources said DeepSeek has increased private hiring of chip-design engineers in recent months. The target matters. Reuters said the chip is being designed for inference, the stage where a trained model generates responses for users. That is the workload DeepSeek must control if its models keep moving from research breakthrough to widely used product. Training grabs the headlines and consumes huge one-time clusters. Inference becomes the recurring cost center once users show up. Liang's move into silicon is a reversal of the image that made DeepSeek famous: a research-first lab that appeared to squeeze frontier performance out of constrained hardware and open model releases. The [DeepSeek-V3 GitHub repository](https://github.com/deepseek-ai/DeepSeek-V3) says V3 is a 671 billion parameter mixture-of-experts model with 37 billion active parameters per token, and that full training required 2.788 million H800 GPU hours. DeepSeek also publishes model weights, code and deployment notes through GitHub, including support paths for Nvidia, AMD and Huawei Ascend hardware. Liang came to AI from quant trading rather than the standard big-tech founder track. The Associated Press reported in 2025 that Liang founded DeepSeek in 2023 after building High-Flyer Quantitative Investment Management, a hedge fund that used machine-learning models for computerized trading and had accumulated a large Nvidia A100 cluster by 2022. AP also quoted Liang saying Chinese AI cannot remain in a position of following forever, a thesis that makes the chip push easier to read: DeepSeek is trying to reduce dependence at the hardware layer, where Chinese AI has faced its sharpest constraint. The constraint is policy as much as procurement. The U.S. Commerce Department's Bureau of Industry and Security said its October 17th, 2023 rules were designed to restrict China's ability to buy advanced computing chips and manufacture chips critical for military advantage. Those rules tightened the path for Chinese AI labs to acquire leading Nvidia accelerators, including China-market parts that had been designed around earlier controls. Reuters reported that DeepSeek has depended on Nvidia and Huawei chips to train and run its models. It also reported that DeepSeek said the foundation model behind R1 was trained on Nvidia's H800, a China-market chip Washington banned in late 2023. Since then, Reuters said, DeepSeek has leaned more on Huawei: in April, it released its V4 model adapted for Huawei Ascend chips, and Huawei said its processors were used in part of V4-Flash training. A DeepSeek inference chip would put Liang in the same strategic lane as the largest AI labs, though with a different set of bottlenecks. Reuters noted that OpenAI last month unveiled Jalapeno, its first custom inference chip developed with Broadcom, and that Anthropic has weighed building its own AI chips. The reason is direct: model companies want more control over the hardware curve as inference demand rises and Nvidia supply remains expensive, contested and politically exposed. For DeepSeek, the move also pressures Huawei inside China. Reuters reported that Huawei has gained around half of China's estimated $50 billion domestic AI chip market because U.S. export controls blocked Chinese access to Nvidia's most advanced chips. Huawei's position is already being contested by Alibaba and Baidu, which Reuters said are developing their own AI chips and gaining share. If DeepSeek can build usable inference silicon, Huawei loses leverage over one of China's most visible AI customers. The gap between designing a chip and deploying one at scale remains large. Reuters reported no tape-out date, manufacturing partner, memory supplier, performance target or production plan. Those omissions are the center of the story. AI inference chips depend on [advanced manufacturing](/article/vulcanforms-matthew-rose-mission-systems), packaging, memory bandwidth and software support. U.S. controls have also limited Chinese access to high-bandwidth memory and the most advanced overseas foundries, according to Reuters. DeepSeek can hire chip engineers and talk to suppliers; turning that into a competitive accelerator is a multi-year capital project. The timing lines up with a second shift inside DeepSeek: outside capital. Reuters reported on June 3rd that DeepSeek was set to raise about 50 billion yuan, or $7.4 billion, in its first funding round, with Tencent and CATL among the expected investors and a possible post-money valuation between 350 billion yuan and 400 billion yuan, or $52 billion to $59 billion. That would mark a break from Liang's earlier posture of keeping DeepSeek funded through High-Flyer and focused on research. A chip program gives that capital an obvious use. It also changes what investors are underwriting. DeepSeek is no longer only a model lab trying to prove that Chinese teams can reach frontier performance with fewer chips. It is starting to behave like an AI infrastructure company, where model architecture, serving stack, silicon access and state industrial policy are tied together. That is why the inference detail matters more than the generic phrase "own chip." A training chip would chase the most expensive and technically punishing part of AI infrastructure. An inference chip targets the workload that scales with every prompt, every enterprise deployment and every API call. If DeepSeek's models remain popular, Liang's cost problem moves from the research cluster to the serving fleet. Reuters' report does not show that DeepSeek has solved that problem. It shows Liang is trying to move the bottleneck inside the company before Huawei, Nvidia restrictions, memory suppliers or foundry access define DeepSeek's ceiling for him.