Introduction
On September 4, 2026, Bloomberg reported that DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at a data center it is building in Inner Mongolia. If the order is filled, it would become one of the largest known clusters of Chinese-designed AI chips anywhere.
The detail that matters most is what the chips will not do. Huawei positions the 950DT for model training as well as the decode stage of inference. DeepSeek intends to take only the second half of that: the cluster will serve its models via inference, while training stays on Nvidia hardware.
What Bloomberg Reported
- At least 160,000 Ascend 950DT units destined for a new Inner Mongolia site.
- The chips will primarily run DeepSeek's models, including the DeepSeek V4 family.
- The broader site is being built at gigawatt scale — enough power at full utilization for roughly 750,000 homes.
- DeepSeek is in talks to raise billions to fund the buildout.
- DeepSeek wants more of Huawei's top silicon than Huawei can currently produce.
How the Ascend 950DT Compares
Huawei's published specifications put the 950DT within reach of Nvidia's Hopper generation on memory:
| Huawei Ascend 950DT | Nvidia H200 | |
|---|---|---|
| HBM capacity | 144 GB (HiZQ 2.0) | 141 GB (HBM3e) |
| Memory bandwidth | 4 TB/s | 4.8 TB/s |
Huawei's own figures for the part are 1 PFLOPS at FP8, 2 PFLOPS at MXFP4, and 2 TB/s of interconnect bandwidth. Worth noting that at Huawei Connect 2025 the company scheduled the 950DT for the fourth quarter of 2026 — it is a chip that has only just begun reaching customers, not one with a deployment record.
Memory is the right axis to look at for inference — high bandwidth memory capacity sets how large a model fits on a device, and bandwidth largely sets how fast tokens come out of it. But close on paper is not the same as close in production. Bloomberg positions the 950DT at roughly the level of Nvidia's Hopper generation — the H100 and H200 — which began shipping in 2022. Nvidia has since moved two architectures past it, through Blackwell to Rubin. Delivered throughput also depends on interconnect topology at cluster scale and on a software stack: Huawei's CANN toolchain has a fraction of the maturity, tooling and kernel coverage that CUDA has accumulated since 2007.
Inference on Huawei, Training on Nvidia
The split is the most informative part of the story, and it is not new. DeepSeek has tried training on Huawei silicon before and returned to Nvidia. Training is the harder problem for a young accelerator platform: it runs for weeks across tens of thousands of chips, is unforgiving of numerical instability, and depends on distributed training frameworks tuned against a specific interconnect and collective-communication library. A single failure mode that appears once every few days can cost a run.
Inference tolerates far more. It is stateless between requests, horizontally scalable, and a node that misbehaves can be drained without losing a training checkpoint. For a company serving a mixture-of-experts model at scale, capacity and cost per token dominate — which is exactly where a domestic chip that Washington cannot embargo has an argument.
So the headline is a real milestone for Huawei, and also a bounded one: China's most prominent model lab is willing to bet its serving fleet on Ascend, and not yet its training runs.
The Supply Constraint
The order is aspirational until Huawei can build the parts. Bloomberg reports 950DT output in the low hundreds of thousands of units this year, limited by component shortages — high-end memory in particular — and Huawei has other domestic customers plus overseas shipments. Filling DeepSeek's order alone could take more than a year.
That bottleneck is the same one running through China's wider push to substitute domestic silicon, from the $295 billion state data center program to the chokepoints in the AI chip supply chain. Advanced packaging and HBM, not logic design, are where the constraint currently binds.
Conclusion
A cluster this size would be the clearest evidence yet that Chinese accelerators can carry production AI workloads at scale. It is also a plan rather than a deployment: the chips are for inference, not training; the part itself was slated for Q4 2026 availability; the specifications sit a generation or two behind Nvidia's current silicon; and Huawei may need more than a year to deliver the volume.
Worth watching next: whether DeepSeek discloses serving throughput on Ascend hardware, and whether any lab commits a frontier training run to it.
Sources
- DeepSeek Plans Big Huawei AI Chip Order to Power New Data Center — Bloomberg
- DeepSeek's Huawei Order Still Leaves Nvidia Doing the Training — Bloomberg via Yahoo Finance
- Groundbreaking SuperPoD Interconnect: Leading a New Paradigm for AI Infrastructure — Huawei's Ascend roadmap keynote, the primary source for the 950DT specifications