DeepSeek

AI Chips · China Tech · Data Centers

Milestone

DeepSeek reportedly plans 160,000 Huawei chips for new data center

The bet signals a widening split from Nvidia's CUDA software ecosystem, even though memory shortages could delay full deployment by over a year.

• DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT chips at a new gigawatt-scale data center it's building in Inner Mongolia, according to people familiar with the matter cited by Bloomberg.
• DeepSeek, the Chinese AI lab known for its low-cost open-weight models, intends to use the chips to run inference on its models rather than to train them.
• DeepSeek still trains its models on Nvidia hardware; Huawei's own memory-chip shortages could make fulfilling the full 950DT order take more than a year.
• Component and memory constraints are expected to cap Huawei's total 950DT output at the low hundreds of thousands of units this year, limiting how fast the order can be filled.
• The Inner Mongolia site sits within China's 'Eastern Data, Western Compute' program, which has already drawn roughly $27 billion in AI computing projects to the region for cheaper power.
• Nvidia CEO Jensen Huang has called a shift by DeepSeek to Huawei chips a "horrible outcome" for the US, reflecting concern that export controls are pushing China's AI software away from Nvidia's CUDA platform.
• Every major Chinese model tuned for Huawei's Ascend chips chips away at CUDA's software lock-in, a slower but more durable threat to Nvidia than any single hardware order.