DeepSeek Begins Development of Proprietary Chip for AI Inference: Goal is Independence from NVIDIA and Huawei

The Chinese startup DeepSeek, which gained global prominence for the high efficiency of its generative artificial intelligence models, has initiated the development of a proprietary processor. According to industry sources, the chip will not be intended for model training but for inference operations, the computational phase in which the already structured algorithm responds to user queries.

This move marks a shift for the Hangzhou-based company, which has historically focused on software optimization and algorithmic research rather than commercialization or the development of physical infrastructures. The initiative reflects a global trend already undertaken by other players in the sector: OpenAI recently unveiled the custom chip "Jalapeno" (developed in collaboration with Broadcom), while Anthropic is evaluating similar solutions to gain greater control over costs and development.

Currently, DeepSeek relies on a heterogeneous fleet of accelerators. The reasoning model R1, whose reduced operational costs shook the financial markets in early 2025, was trained on NVIDIA H800 GPUs (a downgraded variant for the Chinese market, subsequently affected by U.S. government export restrictions at the end of 2023). In response to supply constraints, the startup has progressively integrated Huawei's domestic hardware: the recent V4 model has been optimized for Ascend series accelerators, and Huawei processors have also been employed in the training pipeline of the lightweight V4-Flash version.

DeepSeek's entry into semiconductor design could alter the balances of the domestic Chinese market, estimated at about $50 billion. Huawei currently holds about half of this share following the substantial "ban" on NVIDIA solutions, but Huawei's monopoly is already threatened by custom solutions developed internally by giants like Alibaba and Baidu.

DeepSeek's project reportedly began about a year ago and is still in the early stages of development. The company has initiated consultations with external partners, including chip-design specialists, foundries, and memory manufacturers, while simultaneously ramping up the recruitment of electronic engineers through private channels not indexed on public platforms.

Alongside the development of the silicon, DeepSeek is restructuring its corporate structure. The company is about to close an initial funding round of $7 billion, which would certify its valuation between $52 and $59 billion. This marks a sharp turnaround from the previous corporate policy, which for years rejected the influx of external venture capital.