H200 exports mark reset in US-China chip war

On January 13, 2026, the Trump administration formally approved limited sales of Nvidia’s H200 AI chips to China, subject to the approval of non-military use cases and the review of chips by a third-party testing laboratory.

This decision is being hailed as a significant development in US-China technology policy because, since 2022, Washington has restricted Beijing’s access to advanced AI chips with computing power beyond the A100. The H200 is Nvidia’s second-best chip and, until now, Chinese companies could only procure the H20, a degraded version of the H200 with nearly seven times less total processing power (TPP).

Lifting export controls on the H200 effectively ends this policy of containment, but it also appears to be a carefully orchestrated shift driven by the fact that Huawei is fast emerging as a potential competitor to Nvidia in AI chips.

For example, Pakistan has been in talks with Huawei’s vice president regarding the development of GPU-based AI infrastructure in Pakistan under the government’s cloud-first policy. Other developing nations are increasingly seeking alternatives to American cloud providers, driven by cost considerations as well as geopolitical alignment preferences. At the Belt and Road Forums for International Cooperation, China has repeatedly emphasised its Global AI Governance Initiative. At the same time, China’s open-source model, DeepSeek-R1, has been described as “China’s Sputnik moment”.

However, if we benchmark the performance of Nvidia’s chips against Huawei’s, the US government’s assumption that Huawei is becoming a strong competitor appears weak. Huawei has diverse product lines and does not have a specialised niche in GPUs comparable to Nvidia’s. According to DeepSeek researchers, Huawei’s top-of-the-line chip, the Ascend 910C, performs at only about 60% of Nvidia’s H100, the predecessor of the H200. Similarly, Huawei’s annual chip production rate is estimated at under one million units, which cannot meet the growing demand of Chinese companies for high-performance chips.

As per Nvidia’s roadmap, the memory bandwidth of its top-of-the-line chips is expected to increase from 8 TB per second in 2026 to 20 TB per second in 2027, while Huawei’s roadmap suggests its chips will reach a bandwidth of around 9 TB per second by 2028. By contrast, Nvidia’s Rubin Ultra is expected to exceed 50 TB per second by 2028.

To take full advantage of these loosened export controls, Chinese technology companies have reportedly already placed orders for more than two million H200 chips, priced at around $27,000 each and paid for upfront, bringing the total export bill to approximately $54 billion. With Nvidia’s inventory reportedly running low at around 700,000 units, the company has already ramped up production to fulfil new orders, leading to a shortage of high-performance memory chips as a consequence.

This policy reversal, from containment to conditional opening, has provided a much-needed reset for Chinese companies. China already leads in several other aspects of the AI value chain, including data availability, algorithms, human resources and electricity generation. Access to high-end GPU chips had been the primary bottleneck over the past three years. While the United States continues to hold commanding advantages in semiconductor design and manufacturing partnerships with TSMC and other leading foundries, Beijing now has an opportunity to learn, adapt and compete on more equal footing.

That said, Beijing remains cautious about fully reopening its market to Nvidia. Authorities have reportedly informed technology companies that purchases will be approved only under special circumstances, such as for research purposes. Local firms are also being discouraged from stockpiling US chips without approval, in an effort to reduce dependence on foreign technology where possible.

In a nutshell, by allowing H200 exports, the US has effectively pressed the “reset” button on the chip wars. For Nvidia, it represents a $54 billion validation of its technological supremacy. For China, it removes the primary bottleneck constraining its AI ambitions. While the US currently retains the lead in raw processing power and memory bandwidth, Beijing’s cautious approach to stockpiling suggests it is unwilling to trade long-term self-reliance for short-term hardware gains.

In this high-stakes game of silicon diplomacy, the US is betting on the speed of its innovation, while China is betting on its ability to integrate these chips into a superior, state-led AI ecosystem.

THE WRITER IS A CAMBRIDGE GRADUATE AND is working AS A STRATEGY CONSULTANT