Nvidia’s once-uncontested dominance of the Chinese AI chip market is unraveling faster than almost anyone in the industry expected. According to multiple reports from analysts and cloud customers in Beijing, sales of Nvidia’s H20 chip, the most advanced processor the company is currently allowed to sell into China under U.S. export controls, have effectively stalled over the past two quarters. Local competitors, led by Huawei’s Ascend line, are now absorbing the bulk of new orders from Chinese hyperscalers, state-owned cloud providers, and a growing list of sovereign AI initiatives.
The shift is more than a routine story about foreign competition. It represents the most concrete evidence yet that the U.S. strategy of restricting advanced AI hardware exports to China is producing the opposite of its intended effect. Instead of slowing Chinese AI development, the restrictions are accelerating the maturation of a domestic chip ecosystem that may, within a generation, no longer need American silicon at all.
What Happened to the H20
When Nvidia unveiled the H20 in late 2023, the chip was widely viewed as a clever compromise. Throttled to comply with U.S. export rules but still powerful enough to run most mainstream AI training workloads, it became the de facto standard for Chinese buyers who needed sanctioned hardware that could actually deliver results. At its peak, the H20 was generating billions of dollars in quarterly revenue for Nvidia and accounted for a meaningful share of the company’s data center business.
Over the past nine months, however, Chinese customers have begun quietly reallocating budgets toward domestic alternatives. Huawei’s Ascend 910C and 910B chips, produced at scale by SMIC, are now widely deployed inside the country’s largest cloud platforms. Cambricon, Hygon, and several smaller players have all reported sharp increases in shipments, and Chinese government procurement documents show a clear preference for local suppliers whenever performance parity can be demonstrated.
Why Chinese Customers Are Switching
- Huawei Ascend chips now match H20 performance on most mainstream training workloads
- Domestic supply is not subject to sudden U.S. export rule changes
- Local vendors offer bundled software stacks tuned for Chinese AI frameworks
- State-owned cloud buyers are under political pressure to prefer domestic silicon
- Long-term service and support are more reliable with homegrown suppliers
The Strategic Backstory
The original logic of U.S. export controls was straightforward. By denying Chinese companies access to the world’s most advanced AI accelerators, Washington hoped to preserve a meaningful performance gap that would translate into a sustained military and economic advantage. The restrictions were calibrated carefully, allowing older generations of Nvidia chips to flow while blocking the cutting-edge products that power the largest AI training runs.
The strategy has run into three problems that nobody fully anticipated. First, it gave Chinese chip designers an enormous, captive market, allowing them to scale production and iterate on designs far faster than they could have in a competitive global market. Second, it created a clear, government-backed demand signal that made it rational for Chinese cloud providers to absorb the integration costs of switching away from Nvidia’s well-optimized CUDA software stack. Third, it gave Chinese vendors a powerful narrative. Every time a new restriction was announced, it became free advertising for domestic alternatives.
The export controls didn’t slow China down. They built the Chinese chip industry a customer base it would never have had otherwise.
Recent reporting suggests that even Nvidia’s older chips, which remain legal to export, are losing ground. Customers who previously split workloads between H20 and A800 silicon are consolidating around Huawei-only deployments, citing simpler procurement and fewer compliance headaches. The trend is not universal; some specialized workloads still run best on Nvidia hardware. But the center of gravity in the Chinese AI infrastructure market has unmistakably shifted.
Industry analysts are now openly debating whether Washington should reconsider its export strategy. Some argue that the restrictions should be tightened further to preserve whatever advantage remains. Others contend that the damage is already done and that the policy focus should shift toward subsidizing domestic chip research and rebuilding U.S. foundry capacity. A third camp wants the rules relaxed in hopes that flood of Nvidia hardware might at least slow Huawei’s momentum, even if it cannot reverse it. Each option carries significant political and economic tradeoffs that will shape the next decade of semiconductor competition.
What This Means Going Forward
The most immediate consequence is financial. Nvidia’s data center revenue from China has fallen sharply and is unlikely to recover as long as export restrictions remain in place. The longer-term consequence is more troubling. Each quarter that Huawei ships at scale is a quarter in which its software stack matures and its performance improves. Within two to three years, the technical case for choosing Nvidia in China is likely to disappear entirely, even if export rules were relaxed tomorrow. The H200 chip, which Nvidia has signaled it may eventually be allowed to sell into China, is unlikely to reverse the trend unless accompanied by a major relaxation of the broader restrictions. The primary keyword Nvidia China chip stall captures the heart of a tectonic shift in the global AI hardware market, and it is one that policymakers in Washington will be studying for years to come.