DeepSeek is building a 160,000-chip Huawei cluster. The NVIDIA exodus just got real

DeepSeek is building a 160,000-chip Huawei cluster. The NVIDIA exodus just got real

If you've been wondering how serious China is about replacing NVIDIA in its AI infrastructure, here's your answer. DeepSeek — the company that shocked the world with its low-cost open-source models — is planning to build a data center in Inner Mongolia powered by at least 160,000 Huawei Ascend 950DT AI accelerators, according to reports emerging on September 4, 2026 (UTC+8). If completed, it would be the largest known Huawei Ascend cluster in existence.

That's not a pilot project. That's a declaration of independence.

What's actually being built

The scale here is hard to overstate. 160,000 chips is roughly three times the entire Hopper GPU fleet that DeepSeek's parent company High-Flyer is estimated to currently operate. For context, OpenAI's most ambitious training clusters run on tens of thousands of NVIDIA GPUs — DeepSeek is planning a single facility that dwarfs most of them in raw chip count.

The choice of Huawei's Ascend 950DT is significant. The 950DT is Huawei's data center and training-focused chip, part of the Ascend 950 family that also includes the 950PR for inference prefill and recommendation workloads. It's scheduled for volume shipment in Q4 2026 (UTC+8), which means DeepSeek is essentially pre-ordering silicon that hasn't even hit mass production yet.

Huawei's supporting infrastructure is built for scale. The Atlas 950 SuperPoD — Huawei's answer to NVIDIA's NVL systems — supports up to 8,192 NPUs in a single interconnected cluster with a UnifiedBus interconnect bandwidth of 16.3 PB/s. That's designed to compete directly with NVIDIA's NVL144 and GB200 superchips. To reach 160,000 chips, DeepSeek would need roughly 20 of these SuperPoD systems networked together.

Why Inner Mongolia?

This isn't DeepSeek's first dance with the region. Back on July 31, 2026 (UTC+8), Bloomberg reported that DeepSeek was planning a large AI data center in Inner Mongolia with roughly 1 gigawatt of compute capacity. Inner Mongolia has become China's go-to location for AI data centers because of its cheap land, abundant renewable energy (wind and solar), and cool climate that reduces cooling costs. The region already hosts facilities for ByteDance, Baidu, and JD.com.

The 160,000-chip report appears to be the hardware specifics behind that earlier 1GW plan. A gigawatt-scale facility running Ascend 950DT chips would need roughly that many accelerators to hit its power target.

The bigger shift: from NVIDIA to Ascend

What makes this story important isn't just the chip count — it's what it represents. DeepSeek built its reputation on NVIDIA hardware. Its V3 model was famously trained on just 2,048 H800 GPUs for about $5.6 million. The company's parent was estimated to hold around 50,000 Hopper GPUs total.

Now it's making a massive bet on Huawei. The groundwork was laid in April 2026, when DeepSeek released V4 and became the first trillion-parameter model to fully adapt to Huawei Ascend chips — "releasing the engine mid-flight," as one Chinese outlet put it. Since then, eight domestic AI chip makers (Huawei, Cambricon, Hygon, Moore Threads, MetaX, Baidu Kunlun, Alibaba T-Head, and Iluvatar CoreX) have all adapted DeepSeek-V4.

Sources: Huawei, DeepSeek V4 technical report, Bloomberg, industry estimates (September 2026)

The software side matters. CUDA — NVIDIA's programming platform — has been the moat that keeps customers locked in. Huawei's CANN ecosystem has been playing catch-up for years, but DeepSeek's full V4 adaptation proved that a trillion-parameter model can run efficiently on Ascend without CUDA. That was the proof of concept. The 160,000-chip cluster is the production deployment.

What this means for the AI balance of power

U.S. export controls have steadily tightened access to NVIDIA's advanced chips for Chinese companies. The H800 — a cut-down version of the H100 designed to comply with export rules — was already slower than its Western counterpart. Further restrictions have made even those chips harder to obtain at scale.

DeepSeek's move signals that the workaround isn't just "make do with slower NVIDIA chips." It's "build an entirely domestic stack at scale." If 160,000 Ascend 950DT chips can train and run models at competitive cost and performance, the entire calculus of AI hardware dependency shifts.

For other Chinese AI companies — ByteDance, Alibaba, Tencent, Baidu — DeepSeek's cluster serves as a reference architecture. If the most cost-efficient AI lab in China is going all-in on Huawei, others will likely follow.

What to watch

• Q4 2026 (UTC+8) — Ascend 950DT volume shipment begins; that's when the real build-out can start
• Inner Mongolia facility timeline — when construction begins and when the first chips go live
• Model performance on Ascend — whether DeepSeek's next model trained primarily on Huawei chips matches NVIDIA-trained benchmarks
• Competitor response — whether ByteDance, Alibaba, or Tencent announce similar large-scale Ascend deployments

The era of Chinese AI running on NVIDIA isn't over. But the era of it only running on NVIDIA just ended, and DeepSeek's 160,000-chip bet is the loudest signal yet.