Conceptual art: small amber training chip grid on the left, large cyan inference chip constellation on the right, power bar hinting ~1GW
Conceptual cover: split train/infer stacks—not a product photo, Programmatic cover art, not a news photo

Alternate headlines

  1. 160,000 Ascend chips: DeepSeek’s 1GW Ulanqab bet splits inference from NVIDIA training
  2. DeepSeek plans at least 160,000 Huawei Ascend 950DT accelerators for a ~1GW data center
  3. NVIDIA keeps training; Huawei carries inference: DeepSeek’s Inner Mongolia compute wager

Lead

According to Bloomberg (as relayed in secondary reporting), DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at a data center in Ulanqab, Inner Mongolia, targeting roughly 1 GW of compute capacity (some owned, some leased). If completed, it could rank among the largest known Huawei AI chip clusters.

The story is not simply “replace NVIDIA.” Reporting says the cluster is intended for operating/inference, while training remains on NVIDIA where obtainable. Huawei has marketed the 950DT for training, but DeepSeek’s earlier Ascend training attempts for R2 failed and it returned to NVIDIA. Neither company has publicly commented, per tech-ish.

What happened

From Bloomberg and tech-ish:

  • Scale: at least 160,000 Ascend 950DT chips; site in Ulanqab (~350 km northwest of Beijing), with cheap wind/solar and an average temperature of about 4.3°C that aids cooling; full load is likened to power for ~750,000 homes.
  • Timeline: partial operations targeted for late 2027 or early 2028, subject to Huawei production. Ascend 950DT launch is planned for Q4 2026; HBM shortages constrain output, and fulfillment may take more than a year.
  • Use case: inference/operations, not training; training stays on NVIDIA where possible.
  • Funding: DeepSeek raised more than $7B in June 2026 for physical infrastructure.
  • Context: Bernstein (via AP, cited secondarily) forecasts NVIDIA’s China AI share falling from roughly ~40% in 2025 to ~8% in 2026; Huawei AI chip revenue is projected from about $7.5B in 2025 to ~$12B in 2026.

Public reporting did not name a dollar amount for this order; this article invents none.

Why it matters

This is post–export-control infrastructure news: how a leading lab re-splits scarce training GPUs from exploding inference demand. A 160k-class Ascend build at ~1GW would test three things at once—Huawei’s advanced-accelerator production, the power/cooling economics of mega inference campuses inside China, and whether “NVIDIA for training + domestic silicon for inference” becomes the default engineering template. For apps and clouds, inference cost curves may hitch more tightly to Ascend supply than to a single GPU roadmap.

Outlook

Over the next 12–18 months, watch whether 950DT hits the Q4 2026 volume cadence, whether Ulanqab power and racks actually land, whether DeepSeek inference traffic meaningfully moves onto Ascend, and whether NVIDIA availability still sets the training schedule. If fulfillment slips beyond a year, the 1GW narrative shifts from cluster news to capacity news—editors should treat delivery milestones, not intended scale, as the next verification points.