DeepSeek bets $2.5bn on Huawei to run its models at scale

answer: DeepSeek plans 160,000 Huawei Ascend chips for a 1GW Inner Mongolia inference cluster.

takeaways:

DeepSeek is putting its money behind Huawei for the job of serving its AI models to users, planning to deploy at least 160,000 Ascend 950DT chips at a data centre it is building in Ulanqab, Inner Mongolia, Bloomberg reported on 4 September, citing people familiar with the matter. Neither company commented. The order would rank among the largest known clusters of Huawei AI chips anywhere, and it lands as Nvidia's access to the Chinese market keeps shrinking under US export controls.

The site is designed for roughly 1 gigawatt of capacity, built partly by DeepSeek itself and partly leased, with the 160,000 chips covering only part of that footprint. A July report indicated part of the facility would come online by late 2027 or early 2028. At Huawei's roughly $16,000-a-chip pricing, the order implies hardware spending of around $2.56 billion, according to an outside estimate, and amounts to nearly 20 of Huawei's 8,192-chip Atlas 950 SuperPoDs.

The deployment is for inference — running trained models to answer queries — rather than training them. DeepSeek continues to train its models on Nvidia hardware, though its V4 model was tuned to run on Ascend 950. The company is also reportedly developing its own inference chip with SMIC, a second front in reducing reliance on foreign silicon. The Ascend 950DT itself carries 144 GB of HBM memory and 4.0 TB/s of memory bandwidth, specifications aimed squarely at high-throughput inference workloads.

Delivery is not guaranteed on DeepSeek's timeline. Filling an order this size depends on how fast Huawei can manufacture chips under US export restrictions, and people familiar with the plan say it could take more than a year to complete. The backdrop is stark: one forecast cited by tech-ish has Nvidia's share of China's AI chip market falling from 40% to 8% in 2026.

The split matters: DeepSeek is voting with its wallet for Huawei on inference, the workload that scales with usage and revenue, while keeping Nvidia for training, the workload export controls have made hardest to replace. That is a pragmatic hedge, not a full break, and it tells rivals and investors that Huawei's Ascend line is now credible enough to carry production traffic at gigawatt scale, even if supply remains the binding constraint.

For Nvidia, the order is another marker of share loss in China that policy, not product quality, is driving. For Huawei, landing a cluster of this size is a proof point it can use with other Chinese AI firms facing the same export squeeze — provided it can actually manufacture and ship 160,000 chips on a workable schedule, as detailed by XenoSpectrum.