China is experiencing an unexpected situation in the midst of the race for artificial intelligence: the country’s big technology companies want access to the chip NVIDIA H200but this time it is not Washington that sets the pace, but Beijing. The American government has opened the door to its export under clear conditions, although the final permit now depends on China, that has been tightening its policy for months on foreign semiconductors. Alibaba and ByteDance move in this delicate balance, aware that their ability to advance in AI in the immediate future will depend on what their own regulator decides.
Two giants with enormous needs: Alibaba and ByteDance are not simple technology companies, but two of the companies with the greatest demand for computing capacity in China. Alibaba maintains an e-commerce network and cloud services that centralizes a good part of the purchases and sales that go through Taobao, Tmall or AliExpress, both in China and abroad. ByteDance operates TikTok and its Chinese version, Douyin, in addition to maintaining Doubao, its own AI chatbot. This combination of platforms with massive loads turns each jump in power into more than just a technical improvement: it conditions their ability to keep up with the pace of the sector.
The change of course in Washington: On December 8, President Donald Trump announced that the United States would allow H200 to be exported to “approved customers” in China, a move that marked a turn from previous restrictions. The agreement contemplates that the US government receives 25% of the income from these sales, above the 15% applied to H20. The White House presented the decision as a formula to strengthen domestic manufacturing and sustain high-skilled jobs, while maintaining direct control over the flow of chips to China.
Where the H200 fits into the NVIDIA lineup: The H200 belongs to the Hopper architecture, presented in 2022, and occupies an intermediate position between the generations already established in the market and the new Blackwell line, which is NVIDIA’s priority today. Blackwell-based servers can achieve tenfold performance gains on certain models compared to systems using H200, according to recent company data. Still, the H200 remains a relevant product for advanced training, especially in markets where access to newer hardware is restricted by export controls or limited supply capacity.
NVIDIA H200
Why the H200 makes such a difference: The distance between the H200 and the H20 is still notable. According to the Institute for Progressthe H200 achieves a total throughput of 15,840 TPP, almost six times more than the 2,368 TPP of the H20. Compared to the most advanced domestic chips, the gap continues. He Huawei Ascend 910C It reaches 12,032 TPP and offers a memory bandwidth of 3.2 TB/s, while the H200 reaches 4.8 TB/s. That combination of power and speed explains why this chip is so coveted for training large-scale models.
Alibaba and ByteDance have conveyed to NVIDIA their willingness to acquire large batches of the H200 if they receive approval from Beijing, according to information shared with Reuters by several sources. Chip availability is reduced because some manufacturing capacity is geared toward newer generations, increasing pressure on the purchasing window. In this scenario, both companies are trying to anticipate whether the Chinese regulator will allow a processor of this level to be incorporated into their training systems without additional restrictions.
Access conditioned by the Chinese strategy: Authorization to purchase H200 depends not only on company demand, but on how it fits into the self-sufficiency goals set by Beijing. According to sources cited by the aforementioned agency, regulators are likely to demand precise details about the purpose of each order. In all this, it is no secret that China tries to accelerate the development of its own products through manufacturers such as Huawei and Cambricon, and any import of advanced hardware is examined in light of that strategic horizon.
The situation leaves a market in which the rules seem inverted: chips like A100 and H100 They remain under export control, while the H200, more powerful and recent, could arrive in China under an exceptional framework. This asymmetry conditions the advancement of the country’s most ambitious models, which need competitive hardware to continue evolving. The outcome will depend on what Beijing decides in the coming days.
Images | NVIDIA | Arthur Wang |
GIPHY App Key not set. Please check settings