DeepSeek joins OpenAI in move to custom inference silicon

Vertical integration is becoming the standard for frontier AI labs seeking to escape the margin squeeze of third-party hardware. For DeepSeek, building custom silicon is a survival necessity due to tightening U.S. export controls on Nvidia's H800 and Rubin architectures. Within the Chinese market, this project directly challenges Huawei's dominance as the primary local alternative. Success would allow DeepSeek to optimize performance-per-watt specifically for its reasoning models, theoretically offering a better price-to-performance ratio than rivals using generic hardware. Watch for a potential partnership with a domestic foundry like SMIC as the project moves toward a tape-out date.

The strategic pivot into hardware reflects a broader industry trend of large-scale model developers seeking to optimize 'the cost of intelligence' at the silicon layer. Per OpenAI and Broadcom in June 2026, their first custom Intelligence Processor, Jalapeño, achieved a 9-month design-to-production cycle using AI-assisted engineering. This chip is claimed to offer roughly 50% lower inference costs per token compared to general-purpose GPUs. In the same window, Microsoft and other hyperscalers reportedly committed to a 10-gigawatt infrastructure roadmap specifically built around these custom accelerators to support the next generation of agentic AI workflows.

In China, the push for self-reliance has accelerated as the U.S. Department of Commerce moved in June 2026 to close loopholes regarding the export of advanced Blackwell processors to overseas subsidiaries of Chinese firms, per Reuters. This regulatory pressure has forced domestic leaders to lean on local infrastructure; for instance, DeepSeek’s V4 series was specifically optimized for Huawei’s Ascend 950PR and 950DT chips. However, per Asia Tech Review in July 2026, leading startups like Zhipu AI are joining DeepSeek in pursuing bespoke processors to avoid becoming overly dependent on a single domestic vendor like Huawei.

Huawei remains the incumbent leader, confirmng in June 2026 that its new Ascend 950DT chip—featuring 4 TB/s memory bandwidth—will debut on its cloud ecosystem in August 2026 to support massive LLM clusters. Despite this domestic progress, custom chip ventures from startups like DeepSeek face significant manufacturing hurdles, as U.S. restrictions continue to bar Chinese designers from reaching the most advanced sub-3nm global foundries required for efficiency parity with Western peers.