The United States and Japan are tightening oversight of semiconductor supply chains. Beijing continues to make strides in AI, but at a slower pace 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 restrictions. But it now faces tighter and more technical oversight of its supply chains. The United States has tightened controls on advanced chips, packaging, foundries, re-export routes, and end users. Japan, for its part, maintains a significant lock on critical manufacturing equipment and has incorporated semiconductors into its economic security strategy. The result is not a sudden halt to Chinese AI. It is a targeted slowdown. Beijing can still deploy AI in industry, video, services, and defense. However, training the most advanced models, building clusters comparable to the best U.S. standards, and securing regular volumes of cutting-edge chips is becoming more difficult, more expensive, and riskier. This is precisely the objective sought by Washington and, to a lesser extent, by Tokyo.
Increased oversight targets technological power rather than ordinary trade
The crux of the matter is simple. The United States is no longer seeking merely to prevent the direct sale of certain chips to China. It seeks to monitor the entire chain that allows these chips to be designed, manufactured, assembled, shipped, re-exported, and then used in advanced data centers. This involves licensing, end-user verification, due diligence rules for foundries, investigations into circumvention routes, and, in some cases, physical tracking of shipments deemed risky. Reuters has revealed that Washington has traced smuggling networks passing through Malaysia, Singapore, and the United Arab Emirates, and then placed tracking devices on certain shipments to detect diversions to China.
This surveillance approach is driven by a very specific U.S. fear: that advanced chips or related equipment would end up not only in the hands of leading Chinese civilian companies, but also in defense universities, laboratories linked to the People’s Liberation Army, simulation programs, autonomous systems, and dual-use computing capabilities. Reuters reported in December 2025 that entities affiliated with the military, elite universities, and Chinese data centers were already seeking to obtain Nvidia H200 chips through gray markets. This point is crucial. In Washington’s view, the line between commercial AI and strategic AI is too porous to remain passive.
The United States has shifted control from the product to the entire supply chain
The U.S. crackdown is not limited to a straightforward 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 certain diversions. The idea is therefore to require industry players to verify more thoroughly who is ordering, for what use, with what technical specifications, and through which intermediaries.
The framework then shifted further in 2025 and 2026. In January 2025, Washington also published a comprehensive framework on the global dissemination of AI and the flow of advanced chips. This framework was subsequently rescinded in May 2025, then replaced in early 2026 by a different approach—more flexible in certain respects 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 the Nvidia H200 and AMD MI325X, subject to strict conditions regarding security, commercial availability, and guarantees. This does not signify a general relaxation of restrictions. Rather, it means that Washington wants to maintain control on a case-by-case basis, rather than letting the market run wild.
Japan is restricting manufacturing tools rather than end uses
Japan’s role is different, but just as important. Tokyo is not at the center of the global AI GPU market like Nvidia. However, it remains a key supplier of equipment and components essential for semiconductor manufacturing. The Japanese Ministry of Economy, Trade and Industry notes that in July 2023, Japan expanded its controls to cover 23 categories of semiconductor manufacturing equipment. Documents from METI and its White Paper clearly indicate that these rules affect advanced tools related to lithography, deposition, etching, inspection, and other critical stages of the process.
Why does this matter so much to China? Because an advanced AI ecosystem does not depend solely on the finished chip. It depends on the ability to consistently produce high-performance chips, with stable industrial yields, sophisticated packaging, suitable memory, and reliable interconnects. By restricting access to certain equipment, Japan is not preventing all Chinese production. It is making the rapid ascent to the most advanced nodes and processes more difficult. And this restriction is now part of a broader Japanese strategy for economic security, in which semiconductors are considered a strategic industry. Reuters noted again this week that Tokyo is aiming for 40 trillion yen in annual domestic semiconductor sales by 2040, up from about 8 trillion today.
The feared risks concern military AI, autonomy, and industrial superiority
American and Japanese fears are not abstract. They stem from several specific applications. First, the training of very large foundation models capable of powering advanced systems for generation, planning, analysis, and automation. Second, the integration of AI into military systems: intelligence processing, sensor fusion, drones, electronic warfare, cyberdefense, decision support, and simulation. Finally, industrial power in the broadest sense: smart factories, robotics, logistics, healthcare, finance, and connected cities.
When Beijing announces its intention to deploy AI across its entire economy in its new five-year plan, Washington understands that the issue extends far beyond consumer applications alone.
The problem for Western authorities is that the same computing clusters serve multiple purposes. A cluster designed to train a large language model can also be used for military computer vision, sensitive industrial optimization, or dual-use research. This is why the debate focuses as much on advanced semiconductors as on end users. Chips are no longer viewed as mere products. They are treated as power multipliers.
China’s slowdown is real, but it is far from a complete halt
We must avoid an overly binary interpretation. China hasn’t stopped. It continues to advance in AI. Reuters reported that Huawei is developing and deploying its Ascend 910C chips as a domestic alternative, and that companies like Alibaba and Baidu are also increasingly using their own chips for certain tasks. Beijing is clearly seeking to reduce its dependence. But there is a difference between continuing to make progress and keeping pace with global leaders. It is this gap that supply chain monitoring aims to widen.
The Huawei example is telling. In June 2025, a senior U.S. export control official estimated that Huawei would be able to produce no more than 200,000 advanced AI chips in 2025, a volume deemed insufficient to meet Chinese demand. The same source, however, emphasized an important point: China is catching up quickly. This sums up the situation well. Controls slow things down; they do not neutralize them. They force China to make do with lower volume, lower efficiency, and often higher power consumption or greater system complexity to achieve the same result.
The AI capabilities most affected are heavy training and very large clusters
The restrictions do not affect all layers of AI equally. The most affected are tasks that require a lot of parallel computing, a lot of high-bandwidth memory, and fast interconnections between a very large number of chips. This primarily concerns the training of the most advanced models. Reuters reported that Chinese demand for the H200 stemmed precisely from the fact that it was the most powerful processor Chinese companies could hope to access, and that its training capabilities remained unmatched among domestic alternatives, which are better suited for inference.
Inference—that is, the execution of pre-trained models—can tolerate more compromises. This is why China can still deploy AI across its digital services, industry, and government agencies. However, building the infrastructure needed to train next-generation models or run high-end superclusters is becoming more complicated. Reuters reported in September 2025 that Huawei was aiming for an Atlas 950 SuperPod with 8,192 chips in 2026. This type of architecture shows that Beijing is not giving up the race. But it also reveals the immense effort required to compensate for reduced access to the best foreign chips.
The economic effects are less dramatic than a ban, but more profound
The first effect is rising costs. When a company must rely on less direct supply chains, downgraded chips, redesigned products, or less efficient domestic alternatives, its computing costs increase. The second effect is supply volatility.
Reuters reported as early as March 2025 that H3C was warning of a possible H20 shortage in China, amid geopolitical tensions and increased demand following the breakthrough of models like those from DeepSeek. The third effect is technical fragmentation: Chinese players must optimize their software for multiple architectures, often less homogeneous than the standard Nvidia ecosystem.
The fourth effect is strategic. As supply chains come under greater scrutiny, some companies are moving part of their training operations out of China or seeking workarounds. In late 2025, Reuters reported that major Chinese conglomerates were training certain models abroad to gain access to Nvidia chips. This is a clear sign: advanced computing is becoming a geopolitical issue, not just a choice of cloud architecture.
The slowdown sought by Washington and Tokyo is a race against time
Ultimately, the United States and Japan are not trying to prove that China will never achieve advanced AI. They are trying to ensure that it gets there later, at a higher cost, with greater efficiency losses, and under constant pressure on its supply chains. It is a bet on time. Every quarter gained counts for American, Taiwanese, Korean, and Japanese players. Every technological milestone delayed in China, every tool blocked, every smuggling route disrupted pushes back the deadline for China to fully catch up.
But this gamble comes with a trade-off. The longer the pressure persists, the more it pushes Beijing to invest heavily in self-reliance. Reuters noted in December 2025 that China had launched a sort of “Manhattan Project” to compete with the West on AI chips and lithography equipment. This is the central paradox of this technology war: controls are slowing the integration of advanced AI capabilities in China today, while potentially accelerating the drive to build a less dependent ecosystem tomorrow.
Sources
Reuters, January 31, 2025, on 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 of 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 circuits.
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, overview of Japanese controls on 23 categories of semiconductor manufacturing equipment.
METI, English-language FEFTA guide, Japan’s export control framework.
Reuters, March 10, 2026, Japan’s target of 40 trillion yen in domestic semiconductor sales by 2040.
Reuters, April 21, 2025, on shipments of the Huawei 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 via the gray market and its value for model training.
Reuters, September 18, 2025, on the Huawei Atlas 950 SuperPod roadmap.
Reuters, March 5, 2026, on China’s plan to accelerate technological self-reliance and the adoption of AI.
War Wings Daily is an independant magazine.