DeepSeek is preparing to deploy at least 160,000 Huawei Ascend 950DT chips at a new data center in Ulanqab, Inner Mongolia, scheduled to partially come online by late 2027 or early 2028. This facility represents one of the largest known clusters of Chinese-made AI chips and reflects a dramatic shift in how China's AI companies are responding to four years of US export restrictions on Nvidia hardware.

The move underscores a critical reality: as Washington tightens access to Nvidia's flagship Blackwell chips and Beijing throttles imports of permitted alternatives, China has developed distinct workarounds that are reshaping the global AI landscape. DeepSeek's strategy reveals how geopolitical pressure is forcing a fundamental reorganization of AI computing infrastructure.

Why Is DeepSeek Splitting Training and Inference?

To understand DeepSeek's approach, it helps to think about AI model development in two phases. Training is like a student studying a vast library of books for years, requiring immense, specialized computing power. Inference is that same student taking an exam, answering questions efficiently based on what they learned.

Because of ongoing trade restrictions, DeepSeek is splitting these tasks geographically and by hardware. The heavy lifting of training new models remains on whatever Nvidia hardware the company can secure. Inference, running the models for everyday users, will move to the new Huawei 950DT chips. Notably, Huawei designed the Ascend 950DT for training, but DeepSeek is not using it for that purpose; instead, the company is repurposing the chips for inference workloads.

This split strategy allows DeepSeek to maximize the utility of restricted hardware while building redundancy into its operations. It also signals confidence that inference on Huawei chips can deliver acceptable performance for production use.

What Makes the Ulanqab Facility Strategically Important?

The Ulanqab data center is not just large; it is strategically positioned for efficiency. Located 350 kilometers northwest of Beijing, the region offers cheap wind and solar energy coupled with an average temperature of 4.3 degrees Celsius for natural cooling. At full load, the one-gigawatt site will consume enough power for approximately 750,000 homes.

DeepSeek raised over $7 billion in June 2026 to fund this physical infrastructure, marking a significant shift from the company's historical reliance on rented computing resources. This investment signals that Chinese AI companies are moving beyond short-term workarounds and building permanent, owned infrastructure to reduce dependence on US-controlled supply chains.

How to Understand China's Three-Pronged AI Strategy

• Architectural Efficiency: DeepSeek V3 has 671 billion parameters, but only 37 billion activate per word, and its attention design cuts memory use. Together they squeeze maximum utility from cut-down chips, keeping V3's final training run under $6 million, whereas OpenAI's GPT-4 cost over $100 million.
• Brute-Force Scale: Huawei's CloudMatrix 384 achieves 1.7 times the compute of Nvidia's GB200 NVL72 by grouping five times as many weaker chips and using four times the power. The Ulanqab gigawatt site is this concept scaled to the extreme.
• Open-Weight Distribution: By giving models away, Chinese open-weight models took 41 percent of Hugging Face downloads in spring 2026 and 61 percent of tokens on OpenRouter in May. They drastically undercut Western pricing; DeepSeek V4 Pro costs $1.74 per million input tokens compared to $5 for GPT-5.5.

These three strategies work together to create a competitive alternative to the US-dominated AI ecosystem. Architectural efficiency means doing more with less hardware. Brute-force scale compensates for hardware limitations by deploying more chips. Open-weight distribution builds market share through aggressive pricing and accessibility.

What Is the Timeline for Huawei's Chip Launch?

Huawei plans to launch the Ascend 950DT in the fourth quarter of 2026 with its own high-bandwidth memory. However, shortages of that memory will hold output to the low hundreds of thousands this year, meaning fulfilling DeepSeek's order of 160,000 chips could take over a year.

Despite these supply constraints, Huawei still expects its AI chip revenue to jump from $7.5 billion in 2025 to approximately $12 billion in 2026. This growth reflects broader Chinese investment in domestic semiconductor alternatives, even as individual product launches face delays.

The thing to watch is whether Huawei's Q4 2026 950DT launch stays on schedule. Any delay will directly stall DeepSeek's 160,000-chip deployment plan and could ripple across China's broader AI infrastructure buildout.

How Has US-China Trade Policy Reshaped AI Hardware Access?

The geopolitical timeline reveals how rapidly the rules have changed. From October 2022 to October 2023, the US mandated licenses for Nvidia A100 and H100 sales to China and later blocked the slower, China-specific H800s. In January 2025, DeepSeek launched R1, wiping $589 billion off Nvidia's market cap, following V3's training on 2,048 rented H800s for $5.6 million.

Between April and August 2025, the US blocked Nvidia's H20 chip, and Nvidia wrote off $4.5 billion. Huawei responded with CloudMatrix 384. The US later allowed the H20 back for a 15 percent cut of revenue. DeepSeek attempted to train R2 on Ascend but failed, returning to Nvidia hardware.

By September to November 2025, Beijing told ByteDance and Alibaba to halt Nvidia testing. Nvidia's CEO Jensen Huang stated that Nvidia went from 95 percent market share to zero percent in China. State-funded data centers were ordered to use Chinese chips instead.

From January to August 2026, the US conditionally allowed H200 sales, but Beijing bottlenecked imports. A US official claimed DeepSeek trained on smuggled Blackwells. DeepSeek V4, with 1.6 trillion parameters, arrived during this period. ByteDance and Tencent each secured 10,000 H200s, and Nvidia planned for zero China revenue.

What Does This Mean for the Global AI Industry?

The global AI industry has effectively split in two. The US frontier, represented by Anthropic's Fable 5.1 and OpenAI's GPT-6 Astra, is built on closed weights and, at OpenAI, more than 100,000 Nvidia GPUs, driving Nvidia to a $96.2 billion quarter. China's frontier, represented by Moonshot's Kimi K3, Alibaba's Qwen 3.8-Max, and Zhipu's GLM-5.3, runs on a hybrid model: Nvidia for training where possible, Huawei for inference, and open weights for global distribution.

This bifurcation is not temporary. DeepSeek's $7 billion infrastructure investment and Huawei's aggressive chip development suggest that China is building a self-sufficient AI ecosystem designed to operate independently of US supply chains. The success of this strategy will depend on whether Huawei can deliver chips at scale and whether Chinese companies can continue to innovate despite hardware constraints.

For Nvidia, the implications are stark. The company's dominance in AI training remains unchallenged in the West, but its near-total exclusion from China represents a permanent loss of market share in one of the world's largest AI markets. For the rest of the world, the emergence of a competitive Chinese AI ecosystem may accelerate innovation by forcing US companies to move faster and may create new options for countries seeking to reduce dependence on any single supplier.