Engineers found the Chinese processors too unstable for large-scale training of the company’s new R2 model, forcing DeepSeek to rely again on Nvidia GPUs — the very technology U.S. export rules were meant to keep out of Beijing’s reach.
The move underscores a hard truth: while China can innovate at the software and algorithmic level, its AI future still hinges on semiconductors designed in California. This reliance not only exposes the fragility of China’s self-reliance narrative but also raises pressing questions about how far Washington’s export curbs can really go in controlling the global AI race.
DeepSeek’s Huawei gamble falters
When DeepSeek announced it would train its next-generation R2 model on Huawei’s Ascend GPUs, the move was hailed in Beijing as proof that China could shed its reliance on American semiconductors. The plan didn’t last long.
By June 2025, engineers inside DeepSeek privately acknowledged that Ascend chips failed to deliver the consistency required for massive-scale training. Sources familiar with the project told the Financial Times (July 2025) that the Ascend processors suffered from unstable performance, weaker interconnect bandwidth, and a lack of mature software tools — all critical weaknesses in an era when model training can consume tens of thousands of GPUs simultaneously.
The result: Huawei’s silicon is still being used, but only for inference workloads (running the trained model), while training has quietly shifted back to Nvidia hardware, the very dependency China’s AI sector was under political pressure to escape.
Why Nvidia still dominates China’s AI race
Despite U.S. export controls, Nvidia’s grip remains unshaken. Even China’s most sophisticated AI companies struggle to replicate the CUDA software ecosystem that Nvidia has spent nearly two decades refining. Training a model like DeepSeek’s R2 — rumored to involve over 700 billion parameters — is less about raw chip speed and more about orchestration, driver support, and optimization libraries.
In practice, Chinese engineers describe Huawei’s platform as “running a marathon in sandals while Nvidia wears carbon-fiber spikes.” That blunt analogy, shared by one engineer who worked on both setups, captures the gap. Hardware is only half the battle; software maturity is the other half, and Nvidia still leads.
How DeepSeek got its Nvidia chips
Here’s the thorny part. U.S. rules technically bar Nvidia from selling its most advanced chips, such as the A100 and H100, directly to China. Yet congressional investigators revealed in April 2025 that DeepSeek somehow amassed tens of thousands of Nvidia GPUs through shell distributors in Singapore and the Middle East. Lawmakers on Capitol Hill now accuse DeepSeek of sidestepping export rules, with several Republicans calling for tighter scrutiny.
Nvidia, for its part, points to its H20 line of “downgraded” GPUs — designed to meet U.S. restrictions — which it can legally export. But multiple industry insiders note that DeepSeek’s scale suggests it also tapped into backchannels to acquire restricted units.
This dual reality highlights Washington’s dilemma: export bans slow China, but they don’t stop it.
What this means for China’s AI ambitions
DeepSeek’s return to Nvidia exposes the contradiction at the heart of Beijing’s technology push. On one hand, China is pouring billions into domestic chip design and production. On the other, its most visible AI breakthrough still leans on American silicon.
For policymakers in Washington, this is both reassurance and alarm. Reassurance, because it confirms that U.S. technology remains indispensable. Alarm, because despite layers of export controls, Chinese companies are still finding ways to secure the hardware.
For Chinese AI startups, the lesson is more pragmatic: innovation at the algorithmic level — as DeepSeek demonstrated with its mixture-of-experts architecture that slashed training costs — can stretch limited resources, but it cannot fully replace cutting-edge chips.
The bigger picture: an AI arms race with supply chain chokepoints
The DeepSeek saga isn’t just about one company’s hardware choices. It underscores a larger geopolitical reality: the global AI race hinges not only on data and talent but on who controls the chip supply chain.
- China’s vulnerability: Without access to advanced lithography tools (still dominated by ASML in the Netherlands), domestic fabs cannot produce chips on par with Nvidia’s.
- U.S. leverage: Washington’s chip restrictions remain the single most effective lever in slowing Chinese AI ambitions.
- Market implications: Nvidia stock surged past a $3 trillion market cap in June 2025, driven in part by relentless demand from both Western hyperscalers and Chinese firms willing to pay premiums through gray channels.
As one Beijing-based venture capitalist told: “Every AI startup pitch deck begins with the same line — how many Nvidia GPUs they can get. Nothing else matters until that question is answered.” China’s DeepSeek may represent cutting-edge algorithmic ingenuity, but its reliance on American hardware reveals the fragile foundation of the country’s AI push. For now, the future of Chinese AI still runs, quite literally, on Nvidia.
The critical question going forward: can Beijing close the gap before Washington tightens controls further? Or will the world’s most ambitious AI firms continue to operate in a paradox — building revolutionary software on hardware they cannot officially buy?
Either way, the story of DeepSeek’s pivot back to U.S. chips is a reminder that in the AI arms race, semiconductors remain the real battlefield.
FAQs:
Q1. Why is China’s DeepSeek using American Nvidia hardware instead of Huawei chips?
Because Huawei’s Ascend chips were unstable for large-scale AI training, forcing DeepSeek to return to Nvidia GPUs.
Q2. What does DeepSeek’s reliance on Nvidia mean for China’s AI future?
It shows China’s AI breakthroughs still depend heavily on U.S. semiconductors despite domestic innovation efforts.
