But this story, like so many spoon-fed to us by the captured press, is a carefully crafted illusion. Dig beneath the glittering surface of the press release, and you find a tale of coercion, scarcity, and state-enforced obedience. Huawei’s chips are not the world-beaters they are hyped to be. In fact, they can’t even train a cutting-edge AI model. The real story here is not about superior Chinese engineering. It is about Beijing’s iron-fisted quota system, which is strangling its own tech giants into submission and forcing its most promising AI lab onto second-tier hardware, not because it’s better, but because the Communist Party demands absolute loyalty.
Key points:
- DeepSeek ordered 160,000 Huawei Ascend 950DT chips for a massive new data center in Ulanqab, China.
- The chips are being used strictly for inference, not training, after Huawei hardware failed to train DeepSeek's R2 model last year.
- Severe shortages of Huawei's homegrown high-bandwidth memory are bottlenecking production to the low hundreds of thousands of units annually.
- A senior Trump administration official alleges DeepSeek is training on smuggled Nvidia Blackwell chips, though Bloomberg has not independently verified this.
- Beijing has approved only a fraction of the Nvidia H200 chips the U.S. cleared for export, revealing that China's own government, not American restrictions, is the primary force pushing domestic labs toward Huawei.
The Two Lives of a Chip: Training vs. Inference
To understand this charade, you first need to grasp the fundamental difference between the two phases of an AI model’s life cycle. Think of “training” as the brutal, months-long boot camp where an AI is forged. You take tens of thousands of powerful processors, link them together in perfect harmony, and feed them an ocean of data. Every book, every website, every scrap of human language goes in. It’s an exhausting, energy-sucking ordeal that requires immense computational power and near-flawless hardware. “Inference” is what happens after boot camp. It’s the everyday job of answering your question, drafting an email, or summarizing a document. It’s the easy part, the part that makes money for the companies that own the trained model.
Huawei’s executives made a big show last September, promising that their new Ascend 950DT chip could handle both training and inference with equal prowess. Reality, however, has a way of humbling the boastful. Bloomberg’s sources have confirmed that DeepSeek has absolutely no intention of using these 160,000 chips for training. Not a single model.
Why? Because they tried. Under immense pressure from Beijing to showcase domestic hardware, DeepSeek spent months last year trying to train its R2 model on older Huawei Ascend chips. Huawei even sent engineers to sit on-site and troubleshoot.
According to the Financial Times, the result was a complete and utter failure. Not a single successful training run was produced. DeepSeek was forced to quietly crawl back to Nvidia for the heavy lifting, relegating Huawei to the sidelines. A full year and a new hardware generation later, nothing has changed. Huawei’s silicon is being used for the easy job - inference - and not the crown jewel of technological superiority: training.
The Bottleneck Nobody Wants to Discuss
Here is the part that shatters the myth. Huawei cannot mass-produce these chips because they can’t get their hands on enough high-bandwidth memory. This is the specialized component that determines how fast an AI accelerator can actually function. When Washington cut off China’s access to advanced memory from South Korea’s SK Hynix, Samsung, and America’s Micron in December 2024, Huawei was forced to build its own memory from scratch. This marks the first Ascend generation to rely on homegrown memory, and it is crippling output to the low hundreds of thousands of units per year. That is a pittance for a nation racing toward AI dominance. DeepSeek has already had to petition Beijing directly just to get a larger allocation, while begging the government for the scarce chips they need to build their showcase data center.
Meanwhile, Alibaba and Tencent, the giants of Chinese tech, did exactly what you would expect the instant American restrictions eased slightly in December. They snapped up Nvidia H200 hardware the moment the Trump administration cleared exports to vetted buyers. But here's the irony: Beijing only approved a sliver of the volume Washington was prepared to authorize. The Chinese government itself (not American export controls) is the primary force pushing its domestic labs toward Huawei’s inferior chips. This detail unravels the entire narrative. Washington wants you to believe its controls are strangling China’s ambitions. Beijing wants you to believe domestic silicon is winning on merit. Neither story is relaying the truth.
Huawei is not winning because its chips outperform Nvidia’s. It’s winning because the Chinese government is manufacturing scarcity and mandating loyalty. They are transforming a hardware competition into a political instrument. Jensen Huang, Nvidia’s CEO, warned for years that export controls would eventually forge a self-sufficient Chinese chip industry. That prophecy is unfolding now, not through Chinese ingenuity alone, but through the iron hand of a state that is willing to ration its own crown jewel and starve its top AI lab into compliance.
Sources include:
Please contact us for more information.
