Chinese AI: The major brake of the United States on semiconductors

The United States and Japan tighten surveillance of semiconductor supply chains. Beijing is still advancing in AI, but more slowly and at a higher cost.

In Summary

China is not cut off from the semiconductor world. It remains a major industrial player, importing, producing, adapting, and circumventing. However, it now faces stricter and more technical surveillance of its supply chains. The United States has strengthened controls over advanced chips, packaging, foundries, re-export routes, and end users. Japan, for its part, maintains a significant hold on critical manufacturing equipment and incorporates semiconductors into its economic security strategy. The outcome is not a sudden halt to Chinese AI. Instead, it is a targeted slowdown. Beijing can still deploy AI in industry, video, services, and defense. Conversely, training the most advanced models, building clusters comparable to the best American standards, and securing regular volumes of cutting-edge chips have become more difficult, expensive, and risky. This is precisely the goal sought by Washington and, to a lesser extent, Tokyo.

Increased Surveillance Targets Less Ordinary Trade Than Technological Power

The crux of the matter is simple. The United States is not just trying to block the direct sale of certain chips to China. It aims to monitor the entire chain that allows these chips to be designed, manufactured, assembled, shipped, re-exported, and ultimately used in advanced computing centers. This involves licenses, end-user verification, due diligence rules for foundries, investigations into circumvention routes, and in some cases, physical tracking techniques for shipments deemed risky. Reuters revealed that Washington has traced smuggling networks passing through Malaysia, Singapore, and the United Arab Emirates, and placed tracking devices on certain shipments to detect diversions to China.

This surveillance logic is based on a very specific American fear: the prospect of advanced chips or associated equipment fueling not only China’s civilian champions but also defense-linked universities, laboratories connected to the People’s Liberation Army, simulation programs, autonomous systems, and dual-use computing capabilities. Reuters noted in December 2025 that entities affiliated with the military, elite universities, and Chinese data centers were already seeking Nvidia H200 chips through gray markets. This point is crucial. In Washington’s view, the boundary between commercial AI and strategic AI is far too porous to remain passive.

The United States Has Shifted Control from Products to the Entire Supply Chain

The American ramp-up cannot be reduced to a linear embargo. In January 2025, the Bureau of Industry and Security tightened restrictions on advanced computing semiconductors and added due diligence requirements for foundries to better prevent diversions to China. The federal text explicitly states that previous controls had not been sufficient to prevent some diversions. The idea is therefore to require industry players to check more carefully who is ordering, for what use, with which technical characteristics, and through what intermediaries.

The regulatory framework then shifted again in 2025 and 2026. In January 2025, Washington also published a global device regarding the dissemination of AI and flows of advanced chips worldwide. This framework was rescinded in May 2025 and replaced in early 2026 with a different, more flexible approach in certain areas but still focused on securing sensitive exports to China. On January 13, 2026, the BIS announced a case-by-case review policy for certain chips such as Nvidia H200 and AMD MI325X, under strict conditions of security, commercial availability, and guarantees. This does not indicate a general relaxation. Rather, it means that Washington wants to maintain control, case by case, instead of letting the market run its course.

Japan Locks Down Manufacturing Tools Rather Than End Uses

Japan’s role is different but equally important. Tokyo is not at the center of the global GPU market for AI like Nvidia. However, it remains a key supplier of essential equipment and components for semiconductor manufacturing. The Japanese Ministry of Economy, Trade and Industry (METI) reminds that in July 2023, Japan extended its controls to 23 categories of semiconductor manufacturing equipment. METI documents and its White Paper clearly indicate that these rules affect advanced tools related to lithography, deposition, etching, inspection, and other critical process stages.

Why does this matter so much for China? Because an advanced AI ecosystem does not depend solely on the finished chip. It relies on the ability to regularly produce high-performance chips, with stable industrial yields, sophisticated packaging, suitable memories, and reliable interconnections. By limiting access to certain equipment, Japan does not completely hinder Chinese production. It makes it more difficult for China to quickly ascend to the most advanced nodes and processes. Moreover, this lockdown now fits into a broader Japanese strategy of economic security, where semiconductors are considered a strategic industry. Reuters noted earlier this week that Tokyo aims for 40 trillion yen in annual domestic semiconductor sales by 2040, compared to about 8 trillion today.

Fears Surround Military AI, Autonomy, and Industrial Superiority

The American and Japanese concerns are not abstract. They relate to several specific uses. First, training heavy foundation models capable of powering advanced systems for generation, planning, analysis, and automation. Next, integrating AI into military chains: intelligence processing, sensor fusion, drones, electronic warfare, cyber defense, decision support, and simulation. Finally, there is the broader industrial power: smart factories, robotics, logistics, healthcare, finance, and connected cities. When Beijing announces its intention to diffuse AI throughout its economy in its new five-year plan, Washington understands that this issue goes far beyond mere consumer applications.

The problem for Western authorities is that the same computing clusters serve multiple purposes. A cluster intended for training a large language model can also be used for military computer vision, sensitive industrial optimization, or dual-use research. This is why the debate centers so much on advanced semiconductors as well as on end users. Chips are no longer seen as mere products. They are treated as force multipliers.

Chinese Slowdown Is Real, but Not a Total Halt

It is essential to avoid an overly binary reading. China is not stopped. It continues to advance in AI. Reuters has shown that Huawei is preparing and deploying its Ascend 910C chips as a domestic alternative, and groups like Alibaba or Baidu are also increasingly using their own chips for certain tasks. Beijing is clearly seeking to reduce its dependence. However, there is a gap between continuing to make progress and progressing at the same pace as global leaders. This is the gap that supply chain surveillance aims to widen.

The Huawei example is telling. In June 2025, a senior U.S. export control official estimated that Huawei would be unable to produce more than 200,000 advanced AI chips in 2025, a volume deemed insufficient for Chinese demand. That same source stressed, however, that China is quickly catching up. This sums up the situation well. The controls slow things down; they do not neutralize them. They force China to make do with lower volume, reduced yield, and often higher electrical consumption or system complexity to achieve similar results.

Most Affected AI Capabilities Are Heavy Training and Very Large Clusters

The restrictions do not hit all layers of AI equally. Those most affected are tasks that require extensive parallel computing, a lot of high-bandwidth memory, and rapid interconnections between numerous chips. This primarily concerns the training of the most advanced models. Reuters reported that Chinese demand for the H200 came precisely from the fact that it was the most powerful processor accessible to Chinese companies, and its training capabilities remained unmatched among domestic alternatives, which are more suited for inference.

Inference, which refers to executing already trained models, can more easily accommodate compromises. That is why China can still roll out AI in its digital services, industry, or administrations. In contrast, building the infrastructure necessary to train next-generation models or run very high-level superclusters becomes more complicated. Reuters reported in September 2025 that Huawei aimed for an Atlas 950 SuperPod of 8,192 chips in 2026. This kind of architecture shows that Beijing is not abandoning the race. However, it also reveals the brutal effort required to compensate for reduced access to the best foreign chips.

Economic Effects Are Less Spectacular Than A Ban, But Deeper

The first effect is rising costs. When a company must go through less direct channels, use crippled chips, redesign products, or rely on less efficient domestic alternatives, its computing costs increase. The second effect is supply volatility. Reuters reported as early as March 2025 that H3C warned of a potential shortage of H20 in China amid geopolitical tensions and rising demand following breakthroughs by models like those from DeepSeek. The third effect is technical fragmentation: Chinese actors must optimize their software for multiple architectures, often less homogeneous than the standard Nvidia ecosystem.

The fourth effect is strategic. As supply chains become more scrutinized, some companies are moving part of their training outside of China or seeking alternative routes. Reuters reported at the end of 2025 that large Chinese groups were training certain models abroad to access Nvidia chips. This is a clear symptom: advanced computing is becoming a geopolitical issue, not just a cloud architecture choice.

The Slowdown Desired by Washington and Tokyo Is a Bet on Time

Ultimately, the United States and Japan are not trying to prove that China will never achieve advanced AI. They seek to ensure that it happens later, at higher costs, with greater losses in yield and under constant pressure on its supplies. This is a bet on time. Every quarter gained counts for American, Taiwanese, Korean, and Japanese players. Every delayed technological node in China, every blocked tool, every destabilized smuggling route pushes back the deadline for complete catching up.

However, this bet comes with a downside. The longer the pressure lasts, the more it drives Beijing to invest heavily in autonomy. Reuters reminded in December 2025 that China had launched a kind of “Manhattan Project” to compete with the West on AI chips and lithography equipment. This is the central paradox of this technological war: controls slow down the integration of advanced AI capabilities in China today, while potentially accelerating the desire to build a less dependent ecosystem tomorrow.

Sources

Reuters, January 31, 2025, on the smuggling routes for AI chips to China via Malaysia, Singapore, and the United Arab Emirates.
Reuters, August 13, 2025, on the use of tracking devices by U.S. authorities to monitor certain chip shipments.
BIS, January 15, 2025, tightening restrictions on advanced semiconductors and due diligence measures for foundries.
Federal Register, January 16, 2025, regulatory text on additional due diligence measures for advanced computing chips.
BIS, January 13, 2026, revision of the license review policy for certain chips exported to China.
Reuters, March 13, 2026, on the evolution of U.S. restrictions in early 2026.
METI, White Paper 2024, reminder of Japanese controls on 23 categories of semiconductor manufacturing equipment.
METI, English guide FEFTA, Japanese framework for export controls.
Reuters, March 10, 2026, Japan’s goal of 40 trillion yen in domestic semiconductor sales by 2040.
Reuters, April 21, 2025, on the deliveries of Huawei’s Ascend 910C.
Reuters, June 12, 2025, on the U.S. estimate of 200,000 advanced Huawei chips in 2025.
Reuters, December 10, 2025, on the use of the H200 in China through the gray market and its importance for training models.
Reuters, September 18, 2025, on Huawei’s Atlas 950 SuperPod roadmap.
Reuters, March 5, 2026, on China’s plan for accelerated technological autonomy and AI diffusion.

Scroll to Top