DeepSeek Huawei Ascend 950DT 2026: Ulanqab 160k Chip Cluster
4 Sep Bloomberg: DeepSeek wants 160k Ascend 950DT cards in Ulanqab for inference, not training. 1 GW, timeline hangs on Huawei.
To read the 4 September 2026 chips story, split two facts: DeepSeek plans at least 160,000 Huawei Ascend 950DT accelerators at a roughly 1 GW site in Ulanqab, Inner Mongolia; and those cards are slated to run models, not to train them.
Bloomberg, citing people familiar with the matter, says the Hangzhou lab wants a first slice live by late 2027 or early 2028. Installation hangs on Huawei’s output. If it lands as described, it would be one of the largest known Huawei AI chip clusters.
What is actually on the order
Ascend 950DT
Huawei’s next accelerator, announced September 2025, with HiZQ 2.0 HBM: 144 GB and 4 TB/s. Huawei had written a ship window in the fourth quarter of 2026.
The site is assembled
About 1 GW at full buildout, some owned, some leased. Yahoo Finance, on the same reporting, notes Alibaba, VNET and Kuaishou already sit in Ulanqab, and more than 90% of local compute is AI.
Not a training swap
Bloomberg writes DeepSeek does not currently plan to train on the 950DT, even though Huawei sold it for that job. The lab tried Huawei for training and went back to Nvidia.
Line up the dates already on paper
Dates and constraints are calmer than “China has already replaced Nvidia.”
-
1
4 September: a 160,000-unit intent
Bloomberg: at least 160,000 Ascend 950DT cards, for running models. No dollar total, financing structure or power-procurement partner was named.
-
2
2026: output in the low hundreds of thousands
Follow-ups say yearly 950DT production sits in that band, DeepSeek’s allocation is tight, filling the order may take more than a year, and the lab has asked Beijing to push Huawei on priority.
-
3
Late 2027 to early 2028: a first slice
The company wants at least part of the capacity live then. That is a plan window, not a lit hall.
Do not write a purchase plan as “Nvidia is already out”
| Claim | Public line | Still missing |
|---|---|---|
| 160k cluster | Bloomberg: a plan for at least 160,000 950DT cards | Not energized, accepted racks |
| Training swap | Explicitly: run models first | Training went back to Nvidia |
| 1 GW | Full-buildout target, owned plus leased | No named power contract |
Why buyers and policy desks care
Moving the inference stack onto domestic accelerators is a parallel file to Commerce treating remote access to controlled chips as an export, and to reviews of data-center power gear. Do not flatten them into one headline. See the remote-access chip rule and the data-center power stack.
Aligning cluster size and delivery lists does not need a full meeting suite: in figmeet a room ID is enough for voice plus a board; see the guide and features.
# Public markers (Bloomberg 2026-09-04; follow-ups 2026-09-04/07)
Buyer: DeepSeek (Hangzhou)
Chip: Huawei Ascend 950DT (HiZQ 2.0 HBM, 144GB, 4TB/s) — announced Sep 2025
Count: at least 160,000 accelerators
Site: Ulanqab, Inner Mongolia; ~1 GW at full buildout
Mix: company-owned + leased capacity
Use: operate / infer models — not training (Bloomberg)
First slice: late 2027 or early 2028, if Huawei can ship
2026 output: low hundreds of thousands of 950DT units; fulfillment >1 year
Neighbors: Alibaba, VNET, Kuaishou already in Ulanqab; >90% local compute is AI
Prior: Huawei post-trained DeepSeek-V4-Pro on Ascend 910C (Jun 2026 reports)
Training: DeepSeek previously tried Huawei, reverted to Nvidia for that jobBloomberg’s direction is that these processors are meant to run models, even though Huawei designed and marketed them for the heavier training job. Do not rewrite that as “DeepSeek already dropped Nvidia.”
- Ulanqab
- An Inner Mongolia compute corridor. Coverage says hyperscalers and short-video firms are already there; cold weather and nearby wind and solar suit high-draw halls.
- Ascend 950DT
- Huawei’s September 2025 accelerator. The Atlas 950 SuperPoD can hold up to 8,192 of them; 64 SuperPoDs make a SuperCluster. Those are Huawei product specs, not DeepSeek’s accepted rack list.
- Inference vs training
- Inference serves a trained model. Training is the distributed job that punishes weak interconnect and software. That split is the point of this note.
Questions worth checking
Are the 160,000 cards already in racks?
No public acceptance figures. Bloomberg described a plan and a purchase intent. Lead time hangs on Huawei output; follow-ups estimate more than a year.
Will DeepSeek also train on Huawei?
Bloomberg says it does not currently plan to train on the 950DT. It tried Huawei for training and went back to Nvidia when the stack and interconnect were not ready for large distributed jobs.
Is this the same as the Commerce remote-access rule?
No. This note is a Chinese lab stacking domestic accelerators onshore. The remote-access rule is about who can log into controlled chips in third-country halls.
Did Bloomberg print a dollar total?
No. Public copy names no dollar total, financing structure or power partner. Do not multiply a street unit price by 160,000 and call it a contract.