Breaking

DeepSeek Seeks 160,000 Huawei Ascend 950DT Accelerators for a Gigawatt Inner Mongolia Site

September 4, 2026 at 15:52 EDT

Infra & Chips
Policy & Regulation

DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at a gigawatt-scale data center in Inner Mongolia, a build that would rank among the largest publicly known clusters of Huawei AI chips. The processors are intended to run the Hangzhou lab’s models; the company does not currently plan to use them for training, which has so far stayed on Nvidia hardware.

September 4, 2026 · DeepSeek / Huawei

160,000 Ascend chips for a 1 GW campus

DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators in Inner Mongolia to run its models. Training is still expected to stay on Nvidia, and delivery could take more than a year.

Key figures

160,000

Ascend 950DT chips planned

1 GW

Ulanqab campus power scale

750,000

homes' energy equivalent at full use

>1 yr

possible wait to fill the order

Cluster scale, drawn to size

Each block = 10,000 chips. DeepSeek's planned 160,000-chip cluster is more than 3× Meituan's 50,000-chip domestic training run.

50,000

Meituan
June disclosure

160,000

DeepSeek
planned 950DT cluster

How the stack would split

1 · TRAIN

Nvidia accelerators

Still used for core model training

→

2 · SERVE

Ascend 950DT

160,000 chips to run DeepSeek models

→

3 · BOTTLENECK

Memory supply

HBM shortages may delay delivery past a year

What supporters see

A Huawei cluster large enough to dwarf other publicly known Ascend installations, moving inference onto domestic silicon at gigawatt scale.

What skeptics note

Training stays on Nvidia, Huawei cannot yet fill the order, and DeepSeek has not confirmed the plan.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.