DeepSeek is developing its first in-house AI inference chip, three people familiar with the matter told Reuters, a move that could reduce its dependence on NVIDIA and Huawei processors while advancing China's push for a domestic semiconductor ecosystem under US export restrictions.
Technical details
The chip is being designed specifically for inference - the stage where trained models generate responses for users - rather than for training new AI models. DeepSeek gained international attention over a year ago with highly efficient AI models that spread rapidly across global markets. The company's R1 reasoning model, introduced in early 2025, showed competitive performance at lower costs, drawing fresh scrutiny on China's AI progress. For professionals seeking hands-on understanding, Deepseek AI Training courses cover the company's model architecture and deployment.
Chip development strategy
DeepSeek began exploring chip development about a year ago and has held discussions with chip designers, foundries, and memory suppliers, according to the sources. The startup has quietly expanded its semiconductor engineering team in recent months, relying on private recruitment rather than public job postings. The effort marks a strategic shift for a company known for building efficient foundation models rather than commercializing a full technology stack.
Industry trend and China's push
The move follows a broader pattern across the AI industry. OpenAI introduced its first custom inference chip, Jalapeno, developed with Broadcom, last month. Reuters also reported in April that Anthropic is evaluating its own AI processor design. For DeepSeek, the strategy aligns with Beijing's encouragement for domestic technology companies to strengthen local semiconductor capabilities as export controls limit access to advanced foreign hardware.
Founder Liang Wenfeng acknowledged in a 2024 interview that export controls on advanced chips presented a significant challenge. DeepSeek has used both NVIDIA and Huawei processors. The company said the foundation model behind R1 was trained on NVIDIA's H800 chips, developed for the Chinese market before Washington prohibited their export in late 2023. More recently, DeepSeek expanded its use of Huawei's hardware. The V4 model, introduced in April, was optimized for Huawei's Ascend processors, and Huawei said its chips were used for portions of training behind V4-Flash, a lighter version.
Funding and business evolution
DeepSeek's semiconductor ambitions come as its business model shifts. In June, the company closed its first external funding round, a $7 billion round at a valuation between $52 billion and $59 billion, departing from its longstanding policy of avoiding outside investment.
Why this matters for IT and development
DeepSeek's entrance into inference chip design signals that AI model developers are increasingly seeking control over the hardware stack, not just the software. For IT and development professionals, this trend means deeper integration between model architecture and silicon, affecting deployment costs, performance tuning, and infrastructure planning. Understanding how domestic semiconductor efforts in China are reshaping hardware availability can inform technology roadmaps. To build relevant skills, AI for IT & Development Courses cover the intersection of AI models and infrastructure. Monitoring moves like DeepSeek's can help professionals anticipate hardware shifts that ripple through the development stack.
Your membership also unlocks:
