The tech war between the United States and China over artificial intelligence has recently entered a new chapter filled with nuances and contradictions. On one hand, Washington has opened the door for NVIDIA to sell its H200 GPU to Chinese customers, imposing an additional 25% fee on those sales. On the other hand, Beijing is moving in the opposite direction: banning foreign chips in publicly-funded data centers and strengthening local manufacturers like Huawei and Cambricon.
Amid these two moves, two Chinese giants—ByteDance and Alibaba—have already contacted NVIDIA to try to secure their own batches of H200, aware that they still need its power to train large-scale AI models.
The result presents an ironic picture: China excludes NVIDIA from its official infrastructure projects, while its leading tech companies push to buy exactly the chip that symbolizes U.S. dominance in AI.
Washington loosens restrictions on the H200… but at a cost
The White House’s decision comes after months of industry pressure and changes in export policies. President Donald Trump authorized the sale of the H200 to China under specific conditions:
- only to approved customers,
- under supervision of the Department of Commerce,
- and with a 25% cut of the revenue going to the U.S. government.
This measure explicitly excludes NVIDIA’s most advanced GPUs—the current Blackwell family and the upcoming Rubin—still banned for China. In other words, Washington allows NVIDIA to monetize part of the Chinese market with a powerful, but not cutting-edge, chip, in exchange for increasing U.S. tax revenue and maintaining control over the actual computing capacity accessible to Chinese companies.
For NVIDIA, this move is crucial: the company had gone from dominating approximately 95% of the Chinese advanced AI chip market to virtually zero percent after recent restrictions. The H200, even “capped” compared to Blackwell, could reopen a business worth tens of billions of dollars.
Beijing hardens its stance: only Chinese chips in state-funded data centers
Meanwhile, China has taken another step toward technological self-sufficiency. A recent regulation mandates that new data centers receiving public funding must exclusively use domestically produced AI chips. Projects less than 30% complete must remove any foreign chips already installed; the rest will be assessed case by case.
This measure directly affects NVIDIA, AMD, and Intel, serving as a sort of “reference supplier list” where only local players like:
- Huawei, with its Ascend AI GPUs,
- Cambricon, specializing in AI accelerators,
- and other companies like MetaX, Moore Threads, or Enflame.
The goal is clear: reserve major contracts related to strategic projects for domestic manufacturers, while reducing reliance on U.S. technology in critical infrastructure.
The challenge for Beijing is that the ecosystem of software and tools around these Chinese GPUs still lags behind NVIDIA’s. Many developers continue to prefer the CUDA + NVIDIA ecosystem, now reinforced by innovations like CUDA Tile, a new “tiling” programming model that further enhances its AI platform’s attractiveness.
ByteDance and Alibaba caught between political regulations and competitive pressure
In this context, the reactions from major Chinese tech companies are revealing. According to sources cited by Reuters, ByteDance and Alibaba have already approached NVIDIA to inquire about the availability of the H200 following Trump’s approval of exports.
Their reasons are straightforward:
- The H200 is much more powerful than the H20, the “adapted” chip that could legally be exported to China so far. Some analyses suggest performance up to six times higher in certain workloads.
- Chinese domestically-made GPUs still outperform in inference tasks compared to training large AI models, where NVIDIA maintains a clear edge.
- The software ecosystem, libraries, and frameworks optimized over more than a decade around CUDA remain a key differentiator.
However, Beijing has not yet given a clear response on whether it will permit these purchases. Authorities have held urgent meetings with giants like Alibaba, ByteDance, and Tencent to discuss potential demand for the H200 before making a final decision on how open or restricted the supply will be.
Essentially, the Chinese government needs to balance two opposing forces:
- Not losing ground in the global AI race against the U.S.
- And not harming its technological independence goals or hurting local champions, starting with Huawei and Cambricon.
NVIDIA prepares for a monitored market… even in its location
While negotiations over who can buy what continue, NVIDIA has also begun to act on controlling and tracking its chips. The company has developed a location verification technology that could identify where its GPUs are being used, leveraging confidential computing capabilities and communication latency with NVIDIA servers.
The goal is to provide governments and large clients with a tool to combat smuggling and illegal diversion of chips into restricted countries. NVIDIA emphasizes that this is not about a “kill switch”:
- the software would be optional,
- only read telemetry data would be sent,
- and remote shutdown or control of chips would not be allowed.
This type of solution aligns with growing pressure from Washington and the U.S. Congress to ensure AI chips do not end up fueling military programs or surveillance efforts in sanctioned countries.
Time is the real currency of exchange
The pattern of recent decisions reveals a clear strategy. The U.S. and China are not just fighting over hardware and market share; they are also competing over something more intangible: time.
- The U.S. sells China less advanced H200 chips, with economic extra costs and regulatory controls, while securing the latest GPUs for its own ecosystem and allies. It gains revenue, keeps Chinese companies a step behind, and strengthens manufacturers like NVIDIA, AMD, or emerging accelerators.
- China reserves its strategic projects for domestic providers, even if that means accepting lower performance and an less mature software ecosystem for now. It buys time for Huawei, Cambricon, and others to close the technological gap.
- Chinese tech giants like ByteDance and Alibaba aim not to get caught in the middle: they need competitive chips today to stay competitive in generative AI and cloud services, even if that involves navigating bans, lists, and special authorizations.
In this landscape, NVIDIA is not regaining its former dominance in China but is entering—albeit partially—again with a product that remains critical for training advanced models. It does so knowing that both Washington and Beijing, to some extent, need what it offers: computing power and a software ecosystem that’s hard to replicate in the short term.
The battle for AI infrastructure is no longer measured solely in teraflops or GPU units sold. It’s about who can make the other waste the most time in their race to reach—or maintain—leadership.