The US has been trying to slow down China’s artificial intelligence progress for three years without being obvious. Export controls were imposed on the tool, not on talent or code, but rather silicon.
It all boils down to one product: NVIDIA’s H200. One chip is the cause of a major tug-of war between China and the US.
This week’s events show how export controls have slowed China’s progress but also the leverage Beijing has when rules start to be relaxed.
What’s the deal with an older chip?
It is part of Nvidia Hopper Generation, which will be released in 2024. The H200 is below the Blackwell, which continues to be blocked by the US.
The data doesn’t support this assumption.
By the end of 2025, many of the largest AI clusters in use today still depend on Hopper processors. According to public disclosures, 18 of the world’s 20 largest clusters are based primarily on either H100 or H200 system.
The chips are still capable of running and training AI models until 2026.
The gap between Hopper’s and Blackwell’s performance at the system-level continues to narrow. Performance is more important than headline specs when chips are clustered into large groups.
According to independent analyses, H200-based systems are comparable in performance with Blackwell training systems but cost approximately 50% more.
The difference is even smaller for inference workloads that are limited more by memory bandwidth than processing power.
Cost is not a constraint for China. Access.
China’s bottleneck isn’t demand
China is not short of AI experts or eager companies to invest. It lacks the capacity to mass produce high-end chips.
China’s AI production is estimated to be only 1 – 4% of US output in 2025, with a further decline projected for 2026.
Limitations are a result of limited access to high-bandwidth memories, advanced manufacturing equipment and packaging.
The US chip market is more important than it appears. The H200 doesn’t replace Chinese production, it just adds to them.
Even optimistic roadmaps show that a chip comparable to the H200 will not be produced in-house before 2027.
Each Hopper chip imported by China increases the total computing pool in a period when computation is still the main limiting factor.
The scale of the issue has also alarmed policymakers. According to Reuters Chinese companies have ordered more than 2,000,000 H200 chips priced at around $27,000 per chip.
This volume is greater than Nvidia’s inventory, and according to ex-US national security officials it would be roughly equivalent to the computing footprint of an average US frontier AI firm.
Under scrutiny, the argument of dependency collapses
Export supporters often claim that the sale of advanced chips makes China more dependent on American technology. This is not the case.
Chinese companies buy both domestic and Nvidia chips at the same. The procurement mandates guarantee that local suppliers will continue to be in demand, regardless of the outcome with regards to imports.
In China, self-sufficiency is not the result of a competitive market. This is a policy directive.
Jensen Huangs claim that decoupling was unrealistic is a contradiction to reality. Decoupling is not always possible. Leverage is not the same as leverage through sales.
The sale of H200 chips will not cause China to delay its long-term plan. They overlap with it.
Chinese companies use Nvidia to upgrade models today, but continue to invest in alternative hardware that will replace them later.
The dependency will end when the alternatives become good enough. However, the ability to do so does not.
The export controls work unevenly
The export controls never intended to bring down the Chinese semiconductor industry. Export controls were never intended to collapse China’s semiconductor industry.
They have proven effective in that regard. China, despite repeated claims of technological breakthroughs has been unable to move beyond the leading nodes that have acceptable yields.
In recent years, a large number of semiconductor companies have left the market. The promised jump to advanced processing has failed to happen on time.
AI results clearly reflect these constraints. The primary reason US models are better than Chinese today is that they have access to more computing.
The US’s decision this week adds another wrinkle. Trump’s administration approved exports of H200 under complex conditions. The third party technical review of shipments is required.
China sales must not exceed 50% of the total volume sold in US. NVIDIA has to certify that there is a sufficient supply of chips in the country. Chinese purchasers must certify that chips won’t be used in military applications.
The US Government will receive a 25% charge.
Some analysts have already expressed concern about the enforceability of this framework once chip routing is done through cloud service providers. They describe it as an imperfect compromise which may prove difficult to monitor in reality.
Approval does not equate to access
The next step is just as important.
Reuters reports that within days after the US ruling, Chinese customs officials informed agents of their refusal to allow H200 chips to enter China.
The domestic technology companies were ordered to refrain from purchasing the chips except in extreme cases.
Officials haven’t clarified if the action is selective or temporary. The use of harsh language was compared to a ban in fact.
According to reports, some exemptions will be discussed in the near future. This is especially true for projects that are linked with universities.
This is similar to China’s previous handling of Nvidia H20, a weaker chip that the US approved, but Beijing blocked effectively, driving Nvidia AI’s market share in China down to zero.
This episode exposes a much deeper truth. The approval of exports in Washington is not a guarantee for access to China.
Beijing has its own power of gatekeeping and is willing to exercise it to either protect its domestic champions, or strengthen its position ahead of high-level negotiation.
When is an asset traded?
Investors are not concerned with ideology, but rather compound advantage.
AI advances are strongly influenced by scale. Faster training cycles and more experiments are possible with more compute. These effects are cumulative.
The impact of a one- or two year acceleration is far reaching.
The sale of H200 chips reduces the time that China’s AI industry is constrained due to hardware.
It is not worth blocking them. Both sides seem to understand it, despite the fact that they frame their debates in commercial terms.
Commercial incentives and strategic interest are in conflict. NVIDIA is looking to enter a market with the potential of generating tens or even hundreds of billions dollars in revenue.
China is interested in the chips, as they are still among the most desirable. Time is the best tool for US.
As new information becomes available, this post US approves Nvidia chips sales but China hesitates could be updated.
This site is for entertainment only. Click here to read more
