DeepSeek’s Secret AI Inference Chip: A Year‑Long Project Aimed at Reducing Nvidia Dependence

DeepSeek has begun developing its own AI inference chip, a project that started roughly a year ago. The company has kept recruitment for chip design engineers extremely low‑key, posting no positions on public hiring platforms.

The chip is intended solely for inference workloads, reflecting a market trend where AI compute demand is shifting from training to continuous inference. Inference‑optimized silicon can deliver lower power consumption and unit cost compared with general‑purpose GPUs, which is crucial for a commercializing AI firm.

DeepSeek’s strategy aligns with a broader industry movement: OpenAI recently announced a custom inference chip (Jalapeño) built with Broadcom, and Anthropic is reported to be considering its own chip. The company’s founder, Liang Wenfeng, previously warned that chip shortages pose a strategic challenge.

Historically, DeepSeek’s R1 model was trained on Nvidia H800 GPUs, while its later V4 model was adapted to Huawei’s Ascend processors, with Huawei confirming partial training on Ascend chips. Nonetheless, DeepSeek believes diversifying beyond two vendors is insufficient.

At the algorithm level, DeepSeek‑V3.1 introduced the UE8M0 FP8 data format, reportedly designed for the hardware characteristics of upcoming domestic chips, indicating that model design already anticipates future silicon.

Designing a competitive AI chip typically requires several years and massive capital; success is not guaranteed. To support its chip effort, DeepSeek completed a first external financing round in June 2026, raising about 510 billion RMB (≈ 74 billion USD) and valuing the company between 520 billion and 590 billion USD.

The funding will be used to expand domestic‑chip‑focused compute centers, continue chip development, and recruit top global talent. The company also announced hiring for IDC design engineers to plan megawatt‑ to gigawatt‑scale data centers in locations such as Inner Mongolia’s Ulanqab.

Despite these moves, DeepSeek remains low‑profile; three sources requested anonymity and the company has not issued a public comment. Nonetheless, the firm that reshaped global AI competition is now working to reshape its hardware foundation.

Reference: Reuters, "China's DeepSeek developing its own AI chip," 2026‑07‑07.