What did Bloomberg report about DeepSeek and Huawei chips?

On September 4, 2026, Bloomberg reported that DeepSeek plans to deploy at least 160,000 of Huawei's top AI accelerators, the Ascend 950DT, at a massive data center it is building in Ulanqab (乌兰察布), Inner Mongolia, China. Secondary reporting describes the project as a roughly 1GW-scale AI cluster.

Where is the data center located?

Ulanqab, Inner Mongolia — an autonomous region of northern China, not the country of Mongolia. The location is a common point of confusion in English coverage. Inner Mongolia is a established data-center corridor thanks to cheap power and cool climate.

What is the Ascend 950DT?

Huawei's flagship AI accelerator in the Ascend 950 series (codename 'David' in Huawei CANN source code, per Pandaily reporting). It carries HiZQ 2.0 memory with 144GB capacity at 4TB/s bandwidth, and earlier reporting suggests the 950DT line was co-designed with DeepSeek's participation.

How does this relate to DeepSeek's models?

Industry reporting through 2026 holds that DeepSeek's V4 generation was optimized for domestic Chinese silicon — the Ascend 950 series — under US export restrictions. A 160K-unit cluster is the compute backbone for serving V4-family inference (V4-Pro, V4-Flash, Vision-Exp) at scale, which is what the API price stability and capacity headroom ultimately ride on.

Does this change DeepSeek API pricing?

No immediate change. Current rate cards stand (V4-Flash peak $0.44/$1.32, V4-Pro peak $1.74/$3.48, with off-peak and weekend discounts). The build-out is a capacity and sovereignty story: domestic-silicon scale reduces exposure to export-control shocks and supports the aggressive peak/off-peak pricing DeepSeek introduced in August.

Why does a chip order matter to API builders?

Serving capacity is the hidden variable behind API reliability and price trajectories. A 1GW-class domestic cluster signals DeepSeek is provisioning for sustained V4-family demand growth — relevant context for anyone routing production traffic to DeepSeek endpoints directly or through OpenAI-compatible relays.

News / Model Watch · DeepSeek infrastructure

Infrastructure 160K+ Ascend 950DT ~1GW cluster Ulanqab, Inner Mongolia

2026-09-05 · Model Watch · Bloomberg exclusive (Sep 4) + secondary confirmation

DeepSeek to Deploy 160K+ Huawei Ascend 950DT Accelerators at a 1GW Inner Mongolia Cluster

Bloomberg's September 4 exclusive: DeepSeek plans at least 160,000 of Huawei's top AI accelerators — the Ascend 950DT — at a massive data center under construction in Ulanqab, Inner Mongolia, with secondary reporting framing it as a ~1GW-scale AI cluster. It is the lab's biggest domestic-silicon build-out yet, and the clearest physical answer to the question behind every DeepSeek API price card: where does the serving capacity come from?

Direct answer

DeepSeek is provisioning a 160,000+ unit Huawei Ascend 950DT cluster in Ulanqab (Inner Mongolia, China — not the country of Mongolia), per Bloomberg (Sep 4). The 950DT is Huawei's flagship accelerator — 144GB HiZQ 2.0 memory at 4TB/s, reportedly co-designed with DeepSeek input — and the V4 generation was optimized for Ascend silicon under US export restrictions. No API or pricing change today; this is the capacity-sovereignty story underwriting V4-family serving scale (V4-Pro, V4-Flash, Vision-Exp) and the aggressive peak/off-peak rate card.

What's confirmed, and by whom

Bloomberg (Sep 4, exclusive): DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at a data center it's building; Bloomberg's own video desk frames it as ramping competition with Silicon Valley.
Location: Ulanqab (乌兰察布), Inner Mongolia — an autonomous region of China. English headlines saying "Mongolia data center" mean this Chinese region, a classic data-center corridor (cheap power, cool climate).
Scale: secondary wire coverage (KuCoin news flash, Lookonchain feed) describes a 1GW-scale AI cluster project in Ulanqab.
The chip: Ascend 950DT — codename "David" in Huawei CANN source per Pandaily; HiZQ 2.0 memory (144GB, 4TB/s); earlier reporting says the line was co-designed with DeepSeek participation.
Software alignment: industry analysis (Capacity Media, dev.to stress-tests) holds that DeepSeek V4 was optimized for the Ascend 950 series — the cluster is the serving backbone that thesis was pointing at.

Why API builders should care

Capacity headroom is price stability. DeepSeek's August rate card (peak/off-peak tiers, weekend off-peak billing) presumes serving scale that can absorb demand shifts. A 1GW domestic cluster is the physical version of that promise.
Sovereign silicon de-risks the roadmap. Under US export restrictions, NVIDIA supply is a policy variable. 160K domestic accelerators move DeepSeek's dependency profile toward hardware it can actually procure at scale.
The co-design flywheel. If V4 was tuned for Ascend and the fleet is Ascend, future models (and their pricing) increasingly reflect domestic-silicon economics — a structural cost curve distinct from NVIDIA-renting competitors.
Caveats: "plans to deploy" is forward-looking; build-out timelines at 1GW scale run in years, not weeks. No rate-card or SLA change follows from this news alone.

Primary sources

FAQ (2026)

What was reported?

Bloomberg Sep 4: DeepSeek plans ≥160,000 Huawei Ascend 950DT accelerators at a new Ulanqab, Inner Mongolia data center (~1GW per secondary wires).

Inner Mongolia or Mongolia?

Ulanqab is in Inner Mongolia, an autonomous region of China — not the country. A recurring mix-up in EN headlines.

What's the 950DT?

Huawei's flagship accelerator: 144GB HiZQ 2.0 memory, 4TB/s bandwidth, codename "David"; reportedly co-designed with DeepSeek input.

Does it change API prices?

No — rate cards stand. It's the capacity/sovereignty layer under the V4 price structure, not a new price event.

Why care as an API user?

Serving scale = reliability and price-stability headroom. Domestic silicon at 1GW scale de-risks the V4 roadmap under export controls.

Timeline?

"Plans to deploy" — 1GW build-outs run in years. Treat as directional capacity signal, not near-term capacity.

DeepSeek API guide

Vision-Exp open weights · V4-Pro official release · All news

Join waitlist for OpenAI-compatible access