DeepSeek books 160,000 Huawei chips for a 1-gigawatt site. Inference is where the Nvidia exit happens first

DeepSeek, the Hangzhou lab behind the cheapest frontier-class open models, placed an order worth reporting attention in September.

Bloomberg reported the deal: at least 160,000 Huawei Ascend 950DT accelerators destined for a roughly one-gigawatt data center in Ulanqab, Inner Mongolia.

A gigawatt-scale site in Ulanqab, 160,000 Ascend 950DT accelerators, inference only, the company writeup said in describing the build.

The order is the largest known deployment of Huawei AI chips.

The economics behind the number

Follow the money in one line: the order prices at roughly $2.56 billion, analyst writeups said, about $16,000 per accelerator at the 160,000-unit floor.

That buys inference only, a key constraint. Ascend 950DT units are accelerators built to serve pre-trained models rather than to train them. Training capacity still leans heavily on Nvidia architecture across the market.

Delivery runs to the end of 2027 or early 2028, the reporting said, and depends on Huawei production.

The demand pressure behind the order is the export-control map. Nvidia's AI market share in China is forecast to slide from 40 percent to single digits in 2026, analyst reporting said, with DeepSeek effectively leading the counter-trend toward domestic silicon.

Ecosystem check: the order sizes beyond what prior public reporting lists for almost any other lab on Huawei silicon, which makes DeepSeek the customer forcing Huawei's supply line to industrialize.

Huawei will struggle on three stages: wafer supply for Ascend, power and cooling at Ulanqab, and the software stack for inference at scale. All three had to work at once.

Supply chain reality: an inference-only gigawatt-scale deployment on domestic chips removes most of the training-driven exposed cases. If it works, every Chinese inference market follows.

The Nvidia exposure question

Precedent says export controls push demand sideways rather than away, and this is that move in hardware.

DeepSeek's open models already anchor the cheapest inference APIs in the market. Serving them on domestic chips, rather than renting Nvidia compute offshore like rivals reported in October coverage, shortens the loop between cost and control.

The unresolved edge sits in the chips themselves. Ascend 950DT throughput and software compatibility remain company claims until operators publish independent numbers, the same caveat our launch coverage applied to every vendor.

I get the sense the practical read lands on inference pricing rather than sovereignty. If the cluster works, DeepSeek's API price floor drops again.

Source