DeepSeek Reportedly Developing In-House AI Inference Chip, Global LLM Firms Accelerate Compute Autonomy

According to a Reuters report on July 7, 2026, Chinese AI company DeepSeek is secretly developing its own AI inference chip to reduce reliance on Nvidia and domestic chips. The project, launched about a year ago, is in early stages; DeepSeek has engaged with chip design, foundry, and memory chip companies, and privately recruited chip design engineers. On the same day, Zhipu AI was also reported to be considering custom AI chips. Both companies position their chips for inference scenarios, responding to surging inference compute demand from AI application growth.

Background: Compute Constraints and Demand Surge

• External environment: Due to U.S. export controls, supply of Nvidia's high-end chips (e.g., H100/H200) to China is restricted, with Huawei Ascend as the main alternative. Analysts note Nvidia's market share in China has approached zero.
• Internal demand: DeepSeek's R1, V4 and other models attract global API calls with low cost and high performance; Zhipu AI's GLM-5.2 saw daily token usage on overseas platforms surge 27x in a week, with inference compute consumption far exceeding training. Institutional data shows inference can account for 80-90% of a model's lifecycle computing costs.

Key Details: In-House Inference Chip Strategy

• Chip positioning: Designed specifically for AI inference—the stage where trained models generate answers for users—rather than training new models. Compared to general-purpose GPUs, dedicated inference chips (ASICs) can strip redundant circuits for higher energy efficiency and lower mass production costs. Trendforce predicts ASIC growth rate will reach 44.6% by 2026, far exceeding GPU's 16.1%.
• Project progress: DeepSeek's chip project began about a year ago, with intensified hiring in recent months but no public job postings. Zhipu AI has consulted local chip design firms; from design to tape-out is expected to take over two years.
• Funding support: In June 2026, DeepSeek completed its first external funding round, raising approximately $7 billion (Reuters reported $7 billion; QbitAI reported 51 billion yuan, about $7.4 billion), with a valuation of $52-59 billion. Funds will be used for compute center expansion, in-house chips, and talent recruitment.

Reactions and Industry Trends

• Market reaction: Following the news, Nvidia's stock fell about 1.6% in pre-market trading.
• Global trends: OpenAI released its first custom inference chip, Jalapeño, in June 2026 (in collaboration with Broadcom and manufactured by TSMC); Anthropic is also reported to have started early work on in-house chips, engaging with Samsung's 2nm process. Among the world's top three LLM companies, two have clearly advanced chip self-development.
• Challenges: DeepSeek faces long technology cycles (several years), manufacturing process limitations (cannot use advanced overseas processes), HBM supply chain risks, and Nvidia's CUDA ecosystem barriers. Its advantage lies in deep adaptation to its own models, but if chips only serve internal needs, insufficient scale may drive up costs.

Impact and Outlook

• Strategic significance: In-house chips will transform DeepSeek from a pure algorithm company to a software-hardware integrated one, gaining control over underlying model hardware and reducing reliance on external supply chains. DeepSeek has already optimized at the software level (e.g., UE8M0 FP8 data format, DSpark speculative decoding module), laying groundwork for hardware synergy.
• Industry impact: Chinese AI companies entering chip development signals competition shifting from algorithms to infrastructure. If successful, DeepSeek and Zhipu AI will achieve a full-chain closed loop of "in-house chips + proprietary models + domestic manufacturing," but in the short term, they still rely on existing domestic chip solutions.