Photo: Wong Yu Liang/GETTY IMAGES
Current analysis of U.S. semiconductor export controls on China often misses the forest for the trees. Analysts fixate on the flicker of every new development without first asking, “What are the real stakes of the competition?”
The underlying reason for this confusion is that export controls are not primarily about semiconductors themselves. They are about AI, and specifically, the pursuit of advanced AI systems that require access to exponential computing power (i.e., compute). Training ChatGPT-4 in 2023 required approximately 25,000 graphics processing units (GPUs); by 2030, experts estimate this could rise to millions of chips for a single frontier AI model.
This perspective refocuses how we should evaluate the effectiveness of export controls. They are a means to an end, and crucially, just one half of a paired strategy. Export controls strangle China’s compute pipeline while U.S. domestic industrial policy and international partnerships seek to widen the absolute compute gap. Together, these measures aim to push China back as far as possible and catapult the United States toward artificial general intelligence (AGI). Regardless of whether one thinks AGI is a misnomer, for strategic planners, the thinking goes that the first country to secure the AGI laurel will usher in the hundred-year dynasty.
In this respect, there is broad agreement that China’s access to and production of advanced semiconductor technologies must be curbed. However, consensus on the starting point is not enough. Without well-defined success metrics, directionality risks becoming drift. For one, it encourages a form of legibility theater, where analysts measure what is easily quantifiable, not what matters. Second, it often leads to analytical whiplash, with export controls alternately hailed one moment and vilified the next.
This paper argues that compute power should be the principal unit of measurement in U.S.-China technological competition and estimates that by the end of 2025, the United States will have 9.5 million more AI accelerators than China (14.3 million vs. 4.6 million)—a three-fold advantage that translates into an even larger compute gap given that the United States has more performant chips. But this aggregate metric misidentifies the decisive factor. The real question is whether China can centralize sufficient compute resources to enable at least one domestic AI lab to match U.S. capabilities. National compute totals are not what matter; concentrated capabilities are.
This reframing has significant implications for policy effectiveness. Much like nuclear deterrence, China does not need to achieve parity. Current evidence suggests that China has access to enough GPUs to match the scale of leading U.S. training runs (approximately 100,000 GPUs by xAI in Memphis) and possibly keep pace as U.S. labs scale toward 300,000–500,000 GPU clusters and beyond in the future; in other words, China can achieve critical mass even while operating at a significant overall disadvantage.
However, in the long run, and if current scaling trends continue, training a single frontier AI model by 2030 will require millions of GPUs. As compute requirements grow exponentially, China will need ever-larger resources just to have a seat at the table, let alone compete effectively.
This paper further aims to challenge two analytical pitfalls in U.S.-China technology competition. The first is short-termism. Given that advanced AI systems have an uncertain and indeterminate timeline, policymakers must guard against the comforting instinct to celebrate premature victories in a contest that may be in its opening chapters. To this end, it is necessary to understand how the Chinese government is girding itself for long-term technological containment.
The second pitfall is a bias that tends to frame the United States as a dynamic protagonist, while casting China as a static and predictable foil. This reduces the United States’ ability to anticipate counterfactuals that policymakers should be considering.
As such, this section takes the vantage point of a Chinese planner. Taking a first-principles approach, it begins by asking: Given constrained inputs and uncertain breakthroughs, how might China tactically reshape the terms of engagement and redefine the competitive landscape to its advantage? This paper identifies four broad elements:
Streamlining: Beijing realized that the 2018 expansion of the Ministry of Science and Technology (MOST) was a classic centralization overreach. MOST’s sprawling mandate, absorbed from 15 other state organizations in 2018, drowned it in administrative tedium. As such, in 2023, significant responsibilities were reallocated away from MOST, refining its role into a leaner agency focused on long-term strategic planning.
Consolidation: Streamlining alone does not confer decisive advantage. To this end, China established a new party body in 2023—the Central Science and Technology Commission (CSTC)—at the Politburo Standing Committee (PBSC) level. In fact, the entirety of the trimmed-down MOST was designated as CSTC’s executive office. In June 2024, state media revealed that Vice Premier Ding Xuexiang leads the CSTC. Ding is the first-ranked vice premier of China and the sixth-ranked member of the Politburo Standing Committee, as well as its only trained engineer. That President Xi Jinping delegated CSTC leadership to Ding rather than assuming it himself reflects a pragmatic recognition that technical expertise must guide semiconductor policy.
Importantly, Ding holds dual reins for the CSTC and the semiconductor-focused “Leading Small Group”, which approves mergers and acquisitions and channels R&D funding to the private sector for “bottleneck technologies.” This dual authority prevents the common pitfall of leading groups devolving into mere coordinators of competing agencies. It also ensures technical breakthroughs are rapidly operationalized through targeted investments, industrial consolidation, or resource reallocation.
Launched in 2022, the National Unified Computing Power Network (NUCPN) is China’s most ambitious effort at centralizing compute resources, pooling power across the country much like an electrical grid. It aims to deliver over 300 exaflops of computing power by 2025, with 60 percent concentrated in “national hub node areas.”
For comparison, by the end of 2024, based on shipment data and installed base figures from SemiAnalysis, major U.S. AI labs (OpenAI, Google, Meta, Anthropic and xAI) are projected to command 2.21 million NVIDIA H100 GPUs, each delivering 4 petaflops, for roughly 8,840 exaflops—roughly 30 times China’s target. This gap widens dramatically when including all other AI accelerators and access to commercial data centers and neoclouds. By the end of 2025, U.S. labs will have access to 14.31 million AI accelerators.
A more precise comparison involves examining China’s projected compute resources by the end of 2025. According to SemiAnalysis data, China will have roughly 4.6 million AI accelerators. This figure excludes smuggled chips due to quantification challenges and assumes projected deliveries in the face of ever-changing export controls. While China has domestic cloud providers, their GPU deployments are already captured in our direct count. Lastly, analysis of Chinese access to foreign clouds through intermediaries is beyond the scope of this paper.
U.S. Compute Resources (End of 2025): 14.31 Million Total
China Compute Resources (End of 2025): 4.8 Million Total
However, absolute compute advantage is not necessarily decisive. China merely needs to amass sufficient concentrated compute power for breakthrough AI development. To this end, China can likely already build GPU clusters matching the current largest U.S. training runs (100,000 GPUs) and scale these to match anticipated 300,000–500,000 to 1 million U.S. GPU training clusters in the future.
In any asymmetric war, the weaker power must dominate the information game. This will likely take place in two key ways. First, the weaker power will obscure its true capabilities. Manufacturing capacity and yield rates will remain opaque, making accurate U.S. assessments difficult.
Opacity will be just as much about concealing strengths as it will be about hiding weaknesses. China may mask efforts to build large GPU clusters by repurposing existing industrial infrastructure. Energy-intensive sites like aluminum smelters or industrial parks near major hydroelectric facilities such as the Three Gorges Dam could serve as cover—AI computing’s massive power consumption could be hidden within normal industrial usage patterns, maintaining plausible deniability. Moreover, significant breakthroughs are likely to remain hidden until China has fully developed the capability, avoiding a premature clash with the full U.S. cavalry during vulnerable stages. This suggests that the United States may only discover major Chinese technological advances one to two years after the fact.
Second, China will employ strategic misdirection. Targets like the NUCPN 300 exaflops goal may be a calculated feint—ambitious enough to signal intent, but not so threatening as to provoke maximum U.S. response. Tactically, China will aim to keep U.S. attention on “weaker” visible metrics while advancing along alternative technological paths.
Much of current analysis assumes that China must follow the same technological trajectory as the United States and its allies—progressing from DUV to EUV lithography. While China will pursue conventional technologies, analysts should avoid over-indexing on this path as the sole way forward. China is equally likely to invest in alternative approaches, exploring unconventional pathways to bypass chokepoints.
First, it is not a given that EUV will remain the frontier. The history of semiconductor manufacturing is replete with unexpected disruptions of presumed trajectories. In the 1980s, optical lithography was thought to be nearing its limits (1 micron, or 1000 nanometers [nm]), leading IBM and Bell Labs to invest heavily in X-ray lithography. DUV faced skepticism, but ASML overcame barriers and made breakthroughs in optics, resists, and processes. Similarly, EUV faced technical and economic doubts—less than 1 percent of energy from an EUV light source reaches the wafer, requiring enormous power and vacuum systems. With initial costs exceeding $100 million, many, including Intel, doubted its viability. China, facing constraints, has every incentive to explore less charted and exotic routes.
Second, China has a history of sidestepping legacy transitions altogether and leapfrogging the competition. It dominated electric vehicles without mastering internal combustion engines, with companies like BYD leading global EV production. Similarly, Huawei became a telecommunications giant by focusing on Voice over IP (VoIP) rather than traditional telephone switching, bypassing incumbents like Lucent and Nortel.
The Conventional Path
Huawei’s so-called Tashan (Battle) Plan aims to systematically “design-out” U.S. components across the semiconductor supply chain, including electronic design automation (EDA) software. EDA tools serve as the bridge between chip design and manufacturing, translating circuit designs into detailed instructions by simulating and optimizing trade-offs between performance, power consumption, and manufacturability.
While Huawei’s previous success with HarmonyOS after the 2019 U.S. restrictions on Android is instructive, the challenges in EDA are more severe. HarmonyOS was built on open-source foundations and widely available software tools. By contrast, EDA development requires starting almost from scratch—designing proprietary algorithms for circuit layouts, running complex simulations for verification, and creating process models based on decades of industrial know-how.
On semiconductor manufacturing, China will likely continue to employ engineering work-arounds with older DUV tools. In 2022, for example, SMIC achieved 7 nm production, well beyond what experts thought possible, by using multi-patterning techniques. This process compensated for DUV’s limited precision by breaking complex chip patterns into multiple simpler exposures, such as creating a detailed image by overlaying several less detailed stencils.
On the EUV front, Huawei is allegedly pursuing two paths. Its primary effort involves laser-produced plasma (LPP), which, similar to ASML’s approach, uses lasers to strike tin droplets, creating intense EUV light for etching chip features. Huawei reportedly has a prototype under testing, aiming for 5 nm production by 2026 and volume manufacturing by 2027. The secondary path explores steady-state microbunching (SSMB), which uses electron beams to generate continuous EUV light rather than discrete pulses. While potentially more efficient, SSMB demands breakthroughs across multiple technologies.
Alternative and Experimental Pathways
China is likely to invest in areas where U.S. export controls are weaker or nonexistent. Neuromorphic computing represents one such avenue—it mimics the brain’s neural architecture, potentially offering dramatic efficiency gains over conventional processors.
Beyond alternative architectures, China might pursue fundamental materials breakthroughs in graphene and other two-dimensional materials, bypassing silicon entirely.
Moreover, the assumption that AI progress depends on concentrating GPU clusters in a single data center may not hold true. China is reportedly already training AI models in distributed data centers. This approach would require breakthroughs in asynchronous gradient descent but demonstrates that current hardware-software setups represent just one possible equilibrium.
Crucially, China will likely attempt to shift the battlefield to areas where hardware dependencies are reduced. Current AI systems’ massive compute requirements are not inevitable—they are a direct result of how transformer architectures emerged to leverage available GPU hardware. As such, the Chinese asymmetric move would be to focus on making fundamental breakthroughs in compute-efficient AI algorithms that render current U.S. controls on semiconductors less relevant.
The effectiveness of U.S. export controls must be evaluated against two time horizons. In the near term, China can play Tetris with smuggled GPUs, modified chips for the Chinese market, and domestic manufacturing, stacking resources to assemble GPU clusters matching current U.S. capabilities. China’s challenges in scaling beyond 7/5 nm are rendered less decisive by this strategy of resource centralization.
However, this hack will hit a wall in the long run. A scrappy patchwork approach will not be tenable when the minimum ante is measured in millions of GPUs. The goal of export controls then is not to create a perfect embargo, but to raise the barriers to entry and make the floor of competition so high that cobbled-together strategies collapse under their own weight. This will be a contest of industrial might. Scale is not just an advantage; it is the game itself.
But the United States must guard against complacency. The real risk is not that China will stumble into predictable breakthroughs—that is baked into the long arc of technological progress—it is that it will bypass the expected game board entirely and leapfrog the United States through a blindside maneuver it does not see coming. Of course, the flipside is that China makes concentrated bets in the wrong places. For now, the United States holds the stronger hand, but this remains an open contest.
Barath Harithas is a senior fellow in the Economics Program and Scholl Chair in International Business at the Center for Strategic and International Studies in Washington, D.C. The author is grateful to Andreas Schumacher, former visiting fellow, for his valuable suggestions for this paper.
