From May 13 to 15, 2026, former U.S. President Donald Trump visited China for the first time in nine years, accompanied by more than a dozen senior executives from major American companies. The delegation included Apple CEO Tim Cook, Tesla CEO Elon Musk, NVIDIA founder and CEO Jensen Huang, as well as leaders from Qualcomm, Micron, Boeing, Citigroup, Goldman Sachs, Blackstone, Cargill, and others, covering key sectors such as technology, finance, aviation, and agriculture. The highly anticipated visit was not only an important test of the current “fragile trade truce” between the U.S. and China, but also stirred significant waves in the semiconductor industry. Just before the visit, Trump announced the relaxation of export controls on NVIDIA’s H200 chips to China, allowing Jensen Huang to join the delegation — a move that became one of the most telling signals in the U.S. economic and trade strategy toward China.
The H200 is an AI computing chip developed by NVIDIA based on the Hopper architecture, manufactured using TSMC’s 4N advanced process, with a GH100 GPU core and 80 billion transistors. It is the first to feature 141GB of HBM3e high-bandwidth memory, delivering 4.8 TB/s of bandwidth — nearly twice the capacity of the H100 and a 1.4x increase in bandwidth. Its FP8 computing power reaches 3,958 TFLOPS, and it is three times more efficient than the A100 when training trillion-parameter large language models. In NVIDIA’s product lineup, the H200 is currently the second-tier product, just below the Blackwell series, and its performance far exceeds the previously approved “castrated” H20 for China — the latter offers only 148 TFLOPS of FP16 computing power and 296 TFLOPS of FP8 computing power. Compared to the H200, the H20 is a generational downgrade.
However, the export of such a “near-top-tier” chip to China has been far from smooth. During the Biden administration, the H200 was placed on the export ban list for China. On December 8, 2025, Trump announced on Truth Social that, while “ensuring national security,” he would allow NVIDIA to export the H200 to approved customers in China, with the U.S. government taking a 25% cut of sales as a “transaction” condition. The Trump administration introduced a new security review process, designed a routing path through U.S. territory to circumvent legal restrictions on export taxes, and may also attach strict “Verified End User” (VEU) compliance requirements, including quasi-military-grade physical facility upgrades, real‑time telemetry monitoring, and location verification.
It is worth noting that after Trump’s announcement, foreign media reported that Chinese regulators were studying possible restrictions on H200 imports, and that Beijing remains reserved about domestic companies using American technology, especially NVIDIA chips. This has cast significant uncertainty over whether the H200 will truly enter the Chinese market.
Facing the potential entry of the H200 into China, China’s response strategy exhibits a clear dual logic: while unwaveringly prioritizing GPU localization, it would selectively accept some compliant chips to achieve a delicate balance between short‑term computing power supplementation and long‑term industrial independence.
Progress in localization has already provided ample confidence. In 2024, the market share of domestically developed AI chips in China reached 30%, and IDC projects it will exceed 50% in 2025. In 2025, the shipment of domestic AI accelerators in China was about 1.65 million units, accounting for about 41% of the market share. The launch of Huawei’s Ascend 950PR has been particularly symbolic — its performance is nearly three times that of the NVIDIA H20, with FP8 computing power reaching 1 PFLOPS, integrated self‑developed HBM memory, and overall computing power approaching that of the H200. Even more critical is the breakthrough in ecosystem collaboration. When the DeepSeek V4 large language model was released, eight domestic chip manufacturers — including Huawei Ascend, Cambricon, Hygon, Moore Threads, Kunlunxin, and T‑Head — completed “Day 0” simultaneous adaptation, meaning that domestic chip vendors are now capable of co‑iterating with cutting‑edge large models without relying on NVIDIA’s CUDA ecosystem debugging cycle.
From an economic perspective, even if the H200 is approved for entry into China, its total cost of deployment would be much higher than that of domestic alternatives. The H200’s unit price is about 3,250) and VEU compliance costs (about 26,500 per chip — more than three times that of a domestic chip. Moreover, state‑owned operators like China Telecom have already stipulated in their 2024 procurement tenders that the share of domestic chips must be no less than 40%, sending a clear policy signal.
Therefore, even if the H200 were allowed to enter the Chinese market in a compliant manner, its market space would be severely limited. Possible response measures include: restricting H200-approved customers to a few specific domains (e.g., medical, research, and other non‑core applications); imposing mandatory localization quotas and security review mechanisms; or conditionally allowing entry while strictly capping volumes and use cases. Such a strategy would provide a valuable supplement of computing power for the development of China’s indigenous AI industry without undermining the growth of domestic chip makers through massive imports of H200s.
If the H200 eventually gains approval to enter the Chinese market, its impact on China’s computing power development will be profound and complex, requiring examination from multiple dimensions.
On the positive side, the arrival of the H200 would directly ease the computing power shortage during the ramp‑up period of domestic advanced chip production, providing an effective supplement of computing power for domestic large‑model training and AI application deployment. Jensen Huang has admitted that NVIDIA’s exit from the Chinese market costs the company tens of billions of dollars in potential annual revenue, and the Chinese market once contributed over $17 billion in annual revenue to the company. If the H200 re‑enters the Chinese market, it could lower the marginal cost of accessing advanced computing power for domestic enterprises, accelerating the overall pace of AI industry development.
However, the hidden risks are equally significant. The Trump administration’s decision to allow the H200 into China is far from goodwill — it is a highly calculated strategic design. By exporting a “relatively advanced but not the most cutting‑edge” chip (Blackwell and Rubin remain on the banned list), the U.S. seeks to maintain China’s dependence on the American AI technology stack while extracting market returns through a 25% levy. Even more controversial is the potential technical monitoring clause hidden in the VEU agreement. According to reports, the agreement requires cooperation with “location verification” technology — meaning the U.S. government could monitor chip cluster distribution and operational status in real time, and high‑precision telemetry could even reverse‑engineer model parameter sizes, KV cache distributions, and sparse attention communication patterns. This means that what China would be buying is not just computing power, but also potentially handing over the core secrets of its algorithms.
From a long‑term strategic perspective, China’s attitude toward the H200 — whether to actively procure or strategically reject it — will send a critical signal. For the domestic computing power industry, the entry of the H200 represents both competitive pressure and a driving force for improvement. Huawei’s Ascend 950PR has already begun mass supply, and leading internet companies like ByteDance and Alibaba are planning large‑scale purchases. The “window of substitution” for domestic computing power is rapidly narrowing. As an expert from the Ministry of Industry and Information Technology put it: “Core technology has never been something you can just ‘buy.’ Only by firmly grasping it in your own hands can you fundamentally break free from the risk of being choked.”
The signal of H200 lifting the ban conveyed during Trump’s China visit is essentially a carefully crafted game of “granting limited access in exchange for technological dominance.” In the short term, the U.S. policy shift secures fiscal gains and industrial benefits for the U.S., while providing a temporary boost of computing power for China’s AI development. In the long run, however, the underlying reality of U.S.-China technology rivalry has not changed. China’s computing power industry now stands at a critical crossroads — should it embrace the influx of high‑end computing power to accelerate catch‑up, or resist external technological traps and persist in independent breakthroughs?
The answer is perhaps not an either‑or choice. By steadfastly maintaining strategic resolve on localization while pragmatically and prudently selecting compliant chip imports — using imports to fill gaps and domestic production to secure the long term — China can adopt the optimal game strategy in the current landscape. The deeper contest behind the H200’s potential entry into China — the battle over “algorithmic sovereignty” — is the true core of this computing power game.