DeepSeek developing in-house inference chip to cut Nvidia, Huawei reliance

Chinese AI startup DeepSeek is developing its own inference-focused AI chip, according to Reuters reporting from three sources familiar with the effort. The chip is designed for inference—the stage where a trained model answers user queries—rather than for training. The project began approximately a year ago and DeepSeek has recently intensified hiring of chip-design engineers and is holding discussions with external foundry and memory partners. The effort aims to reduce dependence on NVIDIA (whose export controls bar access to advanced chips for Chinese customers) and Huawei (whose Ascend accelerators currently power DeepSeek's production inference workloads).

DeepSeek's move follows the playbook of Western frontier labs: OpenAI unveiled Jalapeno, its custom inference chip with Broadcom, in June 2026, while Anthropic has explored in-house silicon. For DeepSeek, the strategic logic is sharper: US export controls bar Chinese companies from buying NVIDIA's most advanced chips, and Beijing has directed domestic AI champions to build domestic alternatives. Inference now represents approximately 70% of AI compute demand, and custom chips can deliver lower cost-per-token and better power efficiency than general-purpose GPUs running trained models.

Manufacturing constraints are formidable: US export controls block Chinese designers from advanced foundries and restrict access to high-bandwidth memory critical for competitive inference chips. DeepSeek is simultaneously seeking its first external funding round at a reported $52–59 billion valuation (raising ~$7 billion), which would provide capital for chip development. Architecturally, the effort signals that Chinese AI labs are consolidating their stack from model down to silicon, reducing foreign dependencies. For enterprises and architects, this chip announcement, combined with DeepSeek's cheap and capable open-weight models and funding momentum, reinforces the geopolitical bifurcation of the AI supply chain.