The engineers discovered it was too unstable for large-scale training of China's new R2 model, forcing DeepSeek to rely on Nvidia GPUs again.
This move highlights a difficult truth. China can innovate at the software and algorithm level, but its AI future still depends on semiconductors designed in California. This trust not only reveals the vulnerability of China's independent narrative, but also raises pressing questions about how far Washington's export curbs can actually go in controlling the global AI race.
Deepseek's Huawei Gamble is spinning around
When Deepseek announces that it will train the next generation R2 model Above Huawei's Ascend GPUthe move was welcomed in Beijing as evidence that China could rely on American semiconductors. The plan didn't last long.
By June 2025, engineers within Deepseek personally acknowledged that Ascend chips were unable to provide the consistency needed for large-scale training. A source familiar with the project said Financial Times (July 2025) Ascend processors suffered from unstable performance, weak interconnect bandwidth, and lack of mature software tools. This is all important weakness in an era where model training can consume tens of thousands of GPUs simultaneously.
The result: Huawei's silicon is still in use, but only in inference workloads (running trained models), but training quietly returned to Nvidia hardware, but the Chinese AI sector was under political pressure to escape.
Why Nvidia still controls the AI race in China
Despite US export controls, Nvidia's grip remains unshakable. Even China's most sophisticated AI companies have a hard time replicating CUDA Software Ecosystem Nvidia has been refined for nearly 20 years. Training models like Deepseek's R2 – rumored to be involved Over 700 billion parameters – It's not about raw chip speeds, it's about orchestration, driver support, and optimization libraries.
In fact, Chinese engineers describe Huawei's platform as “Nvidia will run a marathon in sandals while wearing carbon fiber spikes.” The dull analogy shared by one engineer who worked on both setups captures the gap. Hardware is only half of the battle. The software is in the remaining half of its maturity, and Nvidia is still leading.
How Deepseek got Nvidia chips
This is the thorny part. The US Rules technically prohibit Nvidia from selling the most advanced chips, such as the A100 and H100, directly to China. However, Congressional investigators revealed in April 2025 that Deepseyk had gathered in some way. Tens of thousands of nvidia gpus Through shell distributors in Singapore and the Middle East. Capitol Hill lawmakers are now denounced Deepseek from Sidestepsing Export Rules, with some Republicans calling for more scrutiny.
Nvidia refers to a part of it “Downgrade” GPU H20 Line – Designed to meet our restrictions – it can be legally exported. However, several industry insiders point out that Deepseek's size suggests that it utilized back channels to obtain restricted units.
This double reality highlights the Washington dilemma: The export ban slowly bans China, but they don't stop it.
What does this mean for China's AI ambitions?
Deepseek's return to Nvidia exposes a contradiction at the heart of Beijing's technology push. On the one hand, China is pouring billions of dollars into the design and production of domestic chips. Meanwhile, its most visible AI breakthrough is still leaning American silicon.
For Washington policymakers, this is both a sense of security and vigilance. It gives us peace of mind as we make sure that US technology remains essential. The alarm is because Chinese companies are still finding ways to protect their hardware despite the layer of export control.
For Chinese AI startups, lessons are more practical: innovation at the algorithm level – as Deepsec demonstrated in that Mixture architecture While significantly reducing training costs can increase limited resources, it cannot completely replace cutting-edge chips.
Big Picture: AI Arm Race with Supply Chain Choke Points
Deepseek Saga is more than just a choice of hardware for one company. It highlights a larger geopolitical reality: the global AI race is not only about data and talent, but also about the world. People who control the chip supply chain.
- China's Vulnerability: Without access to advanced lithography tools (which are still dominated by Dutch ASML), domestic fabs cannot produce chips comparable to Nvidia.
- US Leverage: Washington's tip limit remains the most effective lever in slowing down China's AI ambitions.
- Market meaning: nvidia stock passed a Market capitalization of $3 trillion in June 2025It is driven in part by unfortunate demand from both Western hyperscalers and Chinese companies willing to pay premiums through grey channels.
As one of the Beijing-based venture capitalists said, “All AI startup pitch decks start on the same line. You can get the number of nvidia GPUs. Until you answer that question, nothing else is important.” China's deep seek could represent the creativity of cutting-edge algorithms, but America's reliance on hardware reveals the vulnerable foundations of the country's AI push. For now, the future of China's AI is still literally running on Nvidia.
A key question for the future: Can Beijing close the gap before Washington tightens control even further? Or will the world's most ambitious AI companies continue to operate in paradoxes? Building innovative software on hardware that cannot be officially purchased?
Either way, the US Chips story back to the pivot of Deepseek reminds us that in AI Arms Race, semiconductors continue to be the real battlefield.
FAQ:
Q1. Why is Chinese Deepseek using American Nvidia hardware instead of Huawei chips?
Huawei's ascend chips are unstable for large-scale AI training, so DeepSeek forced it back to the Nvidia GPU.
Q2. What does Deepseek's dependence on Nvidia mean for China's AI future?
Despite domestic innovation efforts, China's AI breakthroughs still show a strong reliance on US semiconductors.
