On September 4, 2026, Bloomberg reported that DeepSeek, one of China‘s most prominent AI labs, had placed an order for 160,000 units of Huawei’s next-generation Ascend 950DT AI chips. The order is valued at approximately $2.56 billion, based on a street price of roughly 111,000 yuan (about $16,000) per chip. The chips will populate a 1-gigawatt data center in Inner Mongolia.
This is not a pilot program. This is not a token gesture. This is an industrial-scale deployment of domestic AI silicon—and it represents the single most significant breach in the wall of U.S. export controls since the restrictions began.
The Order That Changes Everything
The 160,000-chip order is staggering in scale. To put it in perspective: the total order value of $2.56 billion is roughly equivalent to the market capitalization of many publicly traded AI companies. And the 160,000 chips “would comprise only one chunk” of the data center‘s total capacity. The project is 1-gigawatt in scale—comparable to the largest AI infrastructure projects anywhere in the world.
The chips are Ascend 950DT, Huawei’s most advanced AI accelerator. They are built on Huawei‘s self-developed third-generation DaVinci v3 architecture and manufactured using SMIC’s N+3 process technology—an advanced 7-nanometer-class process. Each chip comes equipped with 144 GB of specialized High Bandwidth Memory (HiZQ 2.0 HBM), achieving 4.0 TB/s memory bandwidth and 2.0 TB/s chip-to-chip interconnect speed. The chips natively support low-precision formats such as FP8, FP4, and Huawei‘s proprietary HiF8.
DeepSeek is not buying these chips for training. The company currently relies on Nvidia chips for model training. The 950DT chips are for inference—running already-trained models at scale. This is a crucial distinction. Training requires the absolute cutting edge. Inference requires scale, efficiency, and cost control. DeepSeek is betting that Huawei can deliver the latter, even if it can’t yet match Nvidia on the former.
The Partnership That Bypasses the Ban
The order is not just a purchase. It is a co-design partnership.
DeepSeek and Huawei are jointly defining the “Ascend supernode” architecture, working together on long-context inference optimization—reducing latency, improving throughput, and controlling costs while scaling toward ten-thousand-card cluster capabilities.
The partnership is built around Huawei‘s Atlas 950 SuperPoD infrastructure, which links up to 8,192 Ascend 950DT chips to deliver 8 EFLOPS of peak AI compute power. The supernode technology, powered by Huawei’s proprietary “Lingqu” interconnect, makes multiple physical machines behave like a single computer—handling learning, reasoning, and inference as one unified system.
The Atlas 950 supernode is scheduled for release in the fourth quarter of 2026. The chip itself is also slated for Q4 2026. DeepSeek is not waiting for the technology to be proven. It is placing its order now, betting that Huawei‘s production capacity—currently constrained to “only a few hundred thousand units per year”—will scale in time.
The delivery timeline is telling. Due to production constraints, fulfilling the order could take at least 18 months. Huawei needs to balance DeepSeek’s massive order with other customers and limited overseas exports. DeepSeek is willing to wait. It is playing the long game.
What Nvidia Is Losing
The order represents a seismic shift in the Chinese AI hardware market.
DeepSeek has reportedly eschewed Nvidia‘s China-specific H200 GPUs in favor of Huawei’s Ascend 950DT chips. This is not a decision driven by performance—it‘s driven by availability and strategic necessity. Nvidia’s China-specific chips are subject to U.S. export controls. Huawei‘s chips are not.
Nvidia is reportedly “giving up on China as a lost cause” amid Huawei‘s unassailable lead in the domestic market. In July 2026, DeepSeek’s CEO Liang Wenfeng made a bold statement: “all tasks [NVIDIA] GB300 can do, the Huawei supernode can do, latency and all the same. The only cost: 4”. The implication: Huawei‘s hardware may require more chips to match Nvidia’s performance, but it can match it—and at a price point that makes the trade-off acceptable.
The gap is closing. Nvidia‘s H20 GPU sells for roughly $15,000 to $25,000. Huawei’s 950DT sells for about $16,000. The price difference is marginal. The performance gap is narrowing. And the political risk of relying on Nvidia is growing by the day.
The Export Control Failure
The DeepSeek-Huawei deal exposes a fundamental flaw in the U.S. export control strategy.
The U.S. has spent years restricting Chinese access to advanced chips—first targeting specific companies, then expanding to entire categories of semiconductors, and most recently attempting to restrict model weights themselves. The strategy was based on a simple premise: if China cannot access advanced chips, it cannot build advanced AI.
The premise is collapsing.
Chinese companies are not waiting for the U.S. to lift restrictions. They are building alternatives. They are placing massive orders for domestic chips. They are designing their own architectures and scaling their own supply chains. The 160,000-chip order is not an isolated data point—it is part of a broader trend. In March 2026, Shenzhen launched a 11,000-petaflop AI computing cluster. In June 2026, Meituan completed full-process training of a trillion-parameter model on a 50,000-card domestic chip cluster. In July 2026, Zhipu built a 1-gigawatt data center with domestic chips.
The U.S. is trying to restrict Chinese access to advanced chips. China is building its own advanced chips and deploying them at scale. The export controls are not stopping Chinese AI development. They are accelerating the development of a domestic alternative.
What This Means for OpenAI and Nvidia
For OpenAI, the implications are indirect but significant. DeepSeek‘s V4 Flash is already a cost-competitive alternative to GPT-5. If DeepSeek can scale its inference infrastructure on domestic chips, it can reduce its cost structure even further. The price gap between DeepSeek and OpenAI—already measured in multiples—could widen.
For Nvidia, the implications are direct. China is the company’s largest potential growth market. As of 2026, it is becoming its largest lost market. The 160,000-chip order is not just a lost sale—it is a signal that Chinese customers are building alternatives that do not require Nvidia.
The U.S. export control strategy was designed to protect American technological superiority. Instead, it is accelerating the creation of a parallel technological ecosystem—one that does not depend on American chips, American software, or American standards.
The wall is not holding. DeepSeek just drove a truck through it.
Sources: Bloomberg (September 4, 2026); Wccftech (September 4, 2026); iHeima (September 7, 2026); ChinaAET (September 7, 2026); PChome (September 5, 2026); Lianhe Zaobao (September 5, 2026).
Disclaimer
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations
This analysis is based on reporting and public data available as of the article date; figures may be revised as sources update.
Forecasts from third-party analysts can change with market conditions.
Cost and pricing examples are point-in-time estimates; actual rates vary.
Country and company comparisons rely on public reporting, not operational data.
This sector moves fast; timelines and deal terms may be updated later.
Company deals and regulatory rulings may evolve; verify current status.
AI infrastructure is changing quickly; claims can become outdated soon.
Sources
- Bloomberg (September 4, 2026)
- Wccftech (September 4, 2026)
- iHeima (September 7, 2026)
- ChinaAET (September 7, 2026)
- PChome (September 5, 2026)
- Lianhe Zaobao (September 5, 2026).
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations: This analysis is based on reporting and public data available as of the article date; figures may be revised as sources update.; Forecasts from third-party analysts can change with market conditions.; Cost and pricing examples are point-in-time estimates; actual rates vary.; Country and company comparisons rely on public reporting, not operational data.; This sector moves fast; timelines and deal terms may be updated later.; Company deals and regulatory rulings may evolve; verify current status.; AI infrastructure is changing quickly; claims can become outdated soon.