OPINION:
The Trump administration’s lifting of restrictions on H200-class chips marks a turning point in the debate over export controls. It reflects a sensible judgment that denying individual chips can no longer carry the strategic weight it once did.
The more urgent question for Washington is no longer whether to block a specific processor but whether a chip-centered strategy is sufficient in an artificial intelligence competition shaped by systems, infrastructure and deployment speed.
That shift matters because it exposes the limits of a chip-only approach. Policymakers who focus on chips alone risk leaving the U.S. unprepared as China’s AI capabilities evolve. This is not because export controls have failed or China has overtaken the United States technologically; it’s because the competition has changed.
Denying a class of chips can impose short-term friction, but it no longer determines long-term outcomes. China can substitute hardware, adapt its systems and continue to advance. The AI race is no longer decided by individual components but rather by entire systems.
To understand why, it helps to distinguish between two stages of AI development. The first is training, where large models are created and refined. The second is deployment: the use of AI across factories, logistics networks, government agencies and military units. This is where AI begins to reshape economies and power.
Much of the export-control debate still focuses on training benchmarks and top-tier chips. Deployment depends less on the best processor and more on scale, reliability and integration. Once models are trained, they can run on less-advanced hardware so long as it’s available in large quantities and embedded in functioning systems.
China’s AI strategy reflects this reality. Although Chinese chips still lag behind U.S. offerings at the frontier, they are increasingly designed to support deployment at scale.
In practice, restricting chips such as the H200 can delay China’s progress in the short run, but it won’t prevent China from building and deploying AI systems over time. As domestic alternatives improve, chip denial functions more like a speed bump than a roadblock.
At this stage, China’s binding constraints are no longer primarily about access to a particular chip. They are about building and operating the systems that make AI usable at scale: reliable electricity, data centers, cooling, networks and the organizational capacity to integrate AI into real operations.
These are slow, capital-intensive challenges that no single processor can meaningfully accelerate. The competition is less about who controls a chip and more about who can build, operate and sustain complex AI systems.
This helps explain why Beijing has shown little enthusiasm for acquiring H200-class chips. Chinese firms have learned from repeated rounds of export controls that access to U.S. chips can be withdrawn abruptly. Building long-term AI systems around U.S. hardware that may be cut off again creates significant risk.
As domestic alternatives reach adequate performance for deployment, reliability and supply increasingly outweigh marginal gains in efficiency. Hardware that can be sourced, scaled and supported domestically is more valuable than superior chips whose availability remains uncertain.
The policy implication is clear: Export controls must move beyond blocking individual chips and instead focus on how AI power is generated and used. That means concentrating controls where they still matter most: on advanced AI training systems, very large computing clusters capable of operating at a national scale, and sensitive military, surveillance or intelligence applications.
This shifts attention from individual components to how computing power is assembled, scaled and applied, such as the regulation of large AI data centers or the restriction of systems intended for national security purposes.
Hardware alone is no longer the only source of leverage. Chinese open-source and open-weight AI models are becoming more capable and more widely available, including to U.S. developers, creating new dependencies that chip controls cannot address.
In other words, the AI race is being shaped not just by who builds the hardware but also by how widely models are distributed and deployed. A credible strategy, therefore, pairs targeted export controls with clear standards for evaluating and using foreign AI models in sensitive settings.
Hardware controls still matter, but chip denial alone cannot deliver lasting leverage. As the AI competition shifts, export controls must evolve accordingly. The challenge for policymakers is not to stop the race but rather to shape it in ways that preserve long-term U.S. advantage.
• Jason Hsu is a senior fellow at the Hudson Institute.
Please read our comment policy before commenting.