DeepSeek’s Push into AI Inference Chips – Could It Become the Next OpenAI or Anthropic?

DeepSeek AI inference chip development

DeepSeek is developing a proprietary AI inference chip aimed at reducing dependence on Nvidia and domestic silicon providers. The chip is designed exclusively for the inference stage of large‑language‑model workloads, i.e., generating responses from a trained model, not for training new models.

Strategic context

Geopolitical factors have driven Nvidia’s market share in China to near zero; Huawei’s Ascend processors now dominate the domestic market and supply DeepSeek and other firms.

DeepSeek’s R1 model was trained on Nvidia H800 accelerators; US export controls at the end of 2023 cut off access to Nvidia’s high‑end compute, prompting a shift to Huawei Ascend chips. In April 2026 DeepSeek released the V4 model optimized for Ascend, and Huawei confirmed Ascend contributed to training the lightweight V4‑Flash variant.

DeepSeek’s chip effort is in an early stage. Sources say the company has been engaging external chip designers, wafer fabs and memory suppliers for about a year and has quietly increased hiring of chip‑design engineers.

Industry trend

In June 2026 OpenAI launched its first custom inference chip, Jalapeño, co‑developed with Broadcom and fabricated by TSMC. The chip targets large‑language‑model inference, improving response speed, reducing operating cost and lowering API prices.

Anthropic disclosed early‑stage chip work and discussions with Samsung on 2 nm processes and advanced packaging, making two of the world’s three leading LLM companies active in chip development.

For leading model providers, custom chips enable deep integration of model architecture, inference patterns and service systems into hardware, achieving efficiencies unattainable with general‑purpose GPUs and providing a decisive edge as AI competition shifts from algorithms to infrastructure.

Uncertainties and challenges

Technical and cycle barriers: Competitive AI chips typically require several years and massive capital investment; tape‑out failures and unmet performance expectations are common. As a newcomer, DeepSeek must build team cohesion and engineering expertise over time.

Manufacturing constraints: US export restrictions limit Chinese designers’ access to the most advanced overseas fabs, narrowing process options. High‑bandwidth memory, a core component of inference chips, is also subject to US controls, creating supply‑chain risk.

Ecosystem and scale challenges: Nvidia’s dominance stems from its mature CUDA software ecosystem. Domestic chips often face immature software stacks and high integration costs. DeepSeek can tightly couple its own models to a custom chip, but a chip serving only internal workloads may lack the volume needed for economies of scale, driving up per‑unit cost.

Funding context

Media reports indicate DeepSeek is seeking $7 billion in its first external financing round, valuing the company between $52 billion and $59 billion, marking a shift from its previous policy of rejecting outside investment.