Nvidia’s last remaining exportable AI chip to China—the H20—is now under threat from two directions: newly enforced domestic energy-efficiency rules and an expanding web of U.S. export restrictions.
The chipmaker’s stock saw a decline throughout this week, closing at $111.43 on Thursday, after reports suggested that its H20 processor may soon be locked out of China’s fast-growing data center sector due to non-compliance with government efficiency standards.
On March 27, Reuters reported that Chinese server manufacturer H3C had warned clients about an impending shortage of H20 chips. The alert pointed to soaring demand and ongoing uncertainty around supply—part of a growing trend following months of hardware stockpiling in China.
With all other Nvidia AI chips—A100, H100, A800, and H800—already blacklisted, the H20 has become the company’s last legal foothold in the Chinese AI market.
Green Standards Upend Market Access
China’s National Development and Reform Commission (NDRC) has added a new layer of complexity with the rollout of energy-efficiency benchmarks for AI data centers. These rules focus on “energy per computing power” and power usage effectiveness (PUE)—a metric that evaluates how much energy a data center uses for actual computing versus overhead like cooling or infrastructure.
Nvidia’s H20 underperforms local competitors like Huawei’s Ascend 910B and Biren’s AI chips on these fronts. These alternatives offer “double the efficiency at lower energy consumption”, putting Nvidia at a disadvantage as Beijing prioritizes low-carbon AI infrastructure.
The new standards are not short-term guidelines. China’s State Council outlined a 2024 policy calling for average PUE below 1.5 by 2025, and further efficiency goals through 2030.
Nvidia is reportedly considering technical modifications to the H20 to comply, according to earlier reporting. However, altering the chip could reduce its performance and affect its competitiveness against already optimized local alternatives.
H20 Stockpiling Driven by DeepSeek AI
The recent spike in H20 demand wasn’t spontaneous. It was largely driven by the rise of DeepSeek AI, which quickly gained traction this winter as a Chinese rival to OpenAI.
As DeepSeek’s usage increased across enterprise platforms, companies scrambled to secure the compute power needed to run inference workloads and model training. That scramble turned into a stockpiling spree, as businesses rushed to acquire as many H20 chips as possible before further export restrictions hit.
Washington first banned Nvidia’s A100 and H100 GPUs in October 2022, citing national security concerns. Nvidia responded by launching the A800 and H800—downgraded chips that later faced bans of their own in late 2023. The H20 emerged as a final workaround. But even that window may be closing.
On March 26, one day before the H3C alert, the U.S. Commerce Department expanded its AI export restrictions again, adding dozens of Chinese firms to its entity list. Among them were six subsidiaries of the Inspur Group, a major Nvidia customer. This move, alongside tightened licensing for AI hardware, raises the likelihood that even H20 chips will soon require case-by-case export approvals—if they’re permitted at all.
Nvidia itself has warned that if the H20 is added to the U.S. export blacklist, it would eliminate Nvidia’s last remaining AI chip sales in China.
Tencent and Others Shift AI Strategies
Even Nvidia’s top customers are rethinking their dependency on its hardware. Tencent, one of China’s largest tech conglomerates, disclosed during its Q4 2024 earnings call that it had begun optimizing its AI infrastructure by adopting DeepSeek’s more efficient models. These new architectures are designed to reduce GPU reliance while maintaining performance for tasks like training and inference.
According to the company’s chief strategy officer: “We’re getting much higher productivity on a large language model training from existing GPUs without needing to add additional GPUs at the pace previously expected.” This statement reflects a broader industry shift toward efficiency-focused compute strategies.
While Tencent did place substantial H20 orders to support DeepSeek model integration in platforms like WeChat, its longer-term direction favors smaller models, architectural specialization, and local sourcing.
Even so, infrastructure constraints persist. In early February, DeepSeek temporarily paused API refills due to overwhelming traffic, underscoring the real strain on China’s compute backbone as previously reported.
The Compliance-Competitiveness Trade-Off
The H20 is a chip built out of compromise. After earlier bans on its flagship AI GPUs, Nvidia engineered the H20 to fall below U.S. export thresholds while still serving key customer needs. But in doing so, it left itself open to a different kind of regulatory squeeze—one imposed from within China’s borders.
With Beijing aiming to cut emissions from data centers and shift procurement toward energy-optimized hardware, Nvidia finds itself selling a product that may no longer be politically or commercially viable. At the same time, Washington’s tightening grip on export compliance could stifle any effort to revise the chip for better performance or efficiency.
Chinese policymakers are also clamping down on indirect chip acquisition routes. Reports from multiple sources confirm that the U.S. is pressuring allied countries like the Netherlands and Japan to limit service and support for chipmaking tools such as ASML’s DUV lithography machines, which are still widely used in China.
Nvidia continues to generate revenue from Chinese AI customers for now. But that position is being steadily eroded by geopolitics, policy mandates, and a growing preference for homegrown alternatives. What once looked like a clever regulatory workaround is now colliding with a policy landscape that is increasingly intolerant of compromise.
