DeepSeek, the Chinese AI lab that wiped $589 billion off Nvidia's market cap with its R1 model, is now making its biggest hardware bet yet. According to Bloomberg, the company plans to deploy at least 160,000 of Huawei's Ascend 950DT chips in a new gigawatt-scale data center in Ulanqab, Inner Mongolia. If realized, this would be the largest known cluster of Chinese-made AI chips — and a symbolic milestone in China's quest for AI sovereignty.
The timing is no accident. Four years of US export controls have progressively narrowed DeepSeek's options, from A100s to H100s to H800s to H20s, each restriction forcing a retreat to increasingly compromised hardware. The latest ban on domestic data centers using foreign chips pushed Beijing to mandate Chinese silicon — and DeepSeek is responding with the infrastructure to prove it can scale.
The Inference Split
Here's the critical nuance: DeepSeek isn't abandoning Nvidia entirely. The company is splitting its compute strategy along the training-inference divide. Training — the expensive, months-long process of creating models — remains on whatever Nvidia hardware DeepSeek can secure. Inference — running those models to answer user questions — will migrate to the new Huawei cluster.
This distinction matters enormously. Training requires the absolute peak performance that Nvidia's H100 and H200 chips still deliver, even in their China-specific variants. Inference, by contrast, is more forgiving. It doesn't need the most powerful single chip; it needs mass at scale, efficient power usage, and cost-effective deployment. Huawei's Ascend 950DT, designed with in-house high-bandwidth memory, aims to deliver exactly that.
"Training is like a student studying a vast library of books for years, requiring immense, specialised brainpower. Inference is that same student taking an exam, answering questions efficiently based on what they learned," explained one industry analyst familiar with the strategy. DeepSeek is keeping its best teachers on Nvidia and moving its exam-takers to Huawei.
The Ulanqab Facility
The data center sits 350 kilometers northwest of Beijing in Ulanqab, a region chosen for two reasons: cheap renewable energy and naturally cold air. Wind and solar power keep electricity costs low, while an average temperature of 4.3°C enables free cooling for most of the year. At full capacity, the facility will consume one gigawatt — enough power for 750,000 homes — and house DeepSeek's growing inference workload.
This represents a fundamental shift for DeepSeek. Historically, the company rented computing from cloud providers, a strategy that enabled its rapid rise but left it vulnerable to supply chain disruptions. Now, backed by over $7 billion raised in June 2026, DeepSeek is building its own physical infrastructure — a move that signals both confidence in its business model and a strategic hedge against geopolitical risk.
The Supply Chain Reality
There's a catch: Huawei probably can't deliver the full order any time soon. The Ascend 950DT, launching in Q4 2026, faces production constraints that will stretch the timeline. High-bandwidth memory (HBM) shortages are the primary bottleneck. While China's CXMT has achieved small-batch HBM3E production, yields sit at just 25% — far behind Samsung and SK Hynix, which are already mass-producing HBM4.
Huawei expects AI chip revenue to jump from $7.5 billion in 2025 to approximately $12 billion in 2026, but fulfilling DeepSeek's 160,000-chip order could take over a year. The company is building toward self-sufficiency, but the path is crowded with semiconductor supply chain challenges that no amount of national determination can instantly solve.
The Geopolitical Timeline
DeepSeek's hardware pivot maps directly onto four years of US-China tech war escalation:
- Oct 2022-Oct 2023: US mandates licenses for Nvidia A100 and H100 sales to China; H800s blocked
- Jan 2025: DeepSeek launches R1, trained on 2,048 rented H800s for $5.6 million, wiping $589B off Nvidia's market cap
- Apr-Aug 2025: US blocks Nvidia H20; Huawei launches CloudMatrix 384; DeepSeek fails to train R2 on Ascend, returns to Nvidia
- Sep-Nov 2025: Beijing bans ByteDance and Alibaba from Nvidia testing; Nvidia's China share drops from 95% to 0%; state data centers mandated to use Chinese chips
- 2026: Conditional H200 sales allowed; DeepSeek commits to Huawei for inference at scale
The pattern is clear: each US restriction has accelerated Chinese alternatives. DeepSeek's failure with Ascend for training (R2 couldn't be built on it) pushed it back to Nvidia, but for inference — the workhorse of commercial AI services — Huawei is proving viable.
What This Means
DeepSeek's commitment to Huawei infrastructure marks a transition from adaptation to substitution. For years, Chinese AI labs operated under the assumption that US sanctions would ease — that economic interdependence would prevail over geopolitical tension. That assumption has shattered.
The 160,000-chip cluster won't make DeepSeek independent of Nvidia tomorrow. Training remains dependent on American hardware, and the global frontier still belongs to companies with access to the most powerful chips. But for the billions of daily inference requests that power real products — chatbots, coding assistants, content generation — DeepSeek is betting that Chinese silicon is "good enough" at scale.
The irony is layered. Nvidia CEO Jensen Huang once claimed China had gone from 95% to 0% market share. DeepSeek's new facility suggests that 0% may not mean zero forever — just zero of Nvidia's product, replaced by Huawei's.
As the US continues tightening export controls and China accelerates domestic chip production, the AI world is fracturing along hardware lines. DeepSeek's 160,000-chip bet is a flag planted in that emerging territory: a declaration that Chinese AI will run on Chinese silicon, even if it takes years to get there.
🔥 Hot Takes
1. DeepSeek proved you can train frontier models on a fraction of the hardware — now it's proving you can run them at scale without Nvidia. The R1 model showed the world that Chinese AI labs could match OpenAI's outputs for 1/100th the cost. But training is the glamorous part; inference is the business. By committing 160,000 Huawei chips to inference, DeepSeek is betting that the margin between "good enough" and "best in class" shrinks as Chinese silicon improves — and that's where the real money is made.
2. The training-inference split is the new normal for AI sovereignty. Every country building domestic AI capability will face the same calculation: training needs the best chips (which might require diplomatic exceptions), but inference needs scale (which domestic chips can provide). China isn't trying to beat Nvidia at its own game; it's changing the game to one where scale matters more than peak performance. That's a smarter strategy than most realized.
3. Huawei's Ascend 950DT won't beat Nvidia on benchmarks — and that's exactly why it wins commercially. The Chinese chip will lag in raw performance, but inference workloads are increasingly commoditized. When a user asks ChatGPT a question, they don't care if the answer came from the best chip or a good-enough chip. They care about latency, cost, and availability. Huawei's cluster delivers all three at Chinese prices, and that's the real disruption.