Over the past 14 months, the Trump administration has presided over a significant loosening of export controls on AI chips to China – including the relicensing of H20 and H200 chips, and the establishment of a revenue-sharing agreement between NVIDIA and the U.S. government.
At each step, proponents of relaxation have insisted these chips are “deprecated,” a generation or more behind the cutting edge, and maintained that denying them accomplishes nothing while China builds its own alternatives. It is better, they claim, to keep Chinese developers hooked on NVIDIA’s software ecosystem — such as CUDA and related libraries and optimization tools — than to force Chinese AI developers into using Huawei’s Ascend chips and prompt China to build its own technology stack.
A growing chorus of industry-aligned analysts and consultants has amplified this logic into a broader indictment of the United States’ broader project to control technology exports to China. The “small yard, high fence” doctrine, they argue, has expanded beyond its initial remit —provoking retaliatory Chinese restrictions on rare earths, damaging U.S. chipmakers’ revenue and R&D capacity, and accelerating the very Chinese self-sufficiency it was designed to prevent. The Asia Society’s Paul Triolo, for example, has advocated lifting controls altogether on memory chips, redefining “advanced” upward to exclude everything but the absolute frontier, and effectively returning to the pre-2022 status quo.
This diagnosis gets the problem exactly backward.
China’s “Acceleration” Is Not What It Appears
The most commonly held argument for loosening export controls is that they have backfired — either by accelerating China’s domestic chip industry by forcing investments in self-sufficiency that Beijing might otherwise have deferred; or by forcing Chinese companies to work begrudgingly with Huawei when they otherwise would not have.
This narrative overstates both the pace and the quality of Chinese substitution in the AI industry. China has pursued semiconductor self-sufficiency since 2014, spending hundreds of billions of dollars and squandering billions more in waste, fraud, and abuse. It is important to understand that U.S. export controls did not create the Chinese government’s interest in indigenization — they forced Chinese firms to adopt inferior alternatives during a critical window of the industry’s development, with material impact to China’s hardware and software industry. SMIC’s 5nm-class chip fabrication process is estimated to be 40 to 50 percent more expensive than TSMC’s equivalent, with yields reportedly around one-third of TSMC’s on the same node.
Chinese developers are already beginning to feel the effects of this widening compute moat. Successfully forcing Chinese AI companies to turn to inferior hardware or foreign rental during a period of explosive AI growth is not a failure of export controls; it is direct evidence of the strategic friction the controls were meant to impose. Directly selling H200—or better—chips to China directly undercuts the intended frictions that previous controls imposed.
Moreover, China’s AI chip fabs face component shortages far beyond lithography equipment. They compete against NVIDIA and AMD for access to the very same TSMC fabrication and packaging lines that produce the chips Washington is trying to deny. As things stand, Chinese chipmakers are not capable of supplying enough high-bandwidth memory or advanced packaging to build large numbers of AI accelerators on pace with the United States. This means that selling completed NVIDIA modules to Chinese AI labs hands them chips they could never have fabricated themselves, assembled from materials they cannot yet produce at scale, on production lines where their own designs are queued behind American orders.
Providing Chinese AI labs with more compute only strengthens their capacity to train and run inference for globally competitive AI services, while also freeing up Huawei’s scarce compute for potential export along a “Digital Silk Road.” Selling scarce chips directly to Chinese developers strengthens their ability to compete against U.S. companies in global markets.
There Has Never Been a More Important Time to Restrict China’s Access to Compute
Critics of U.S. technology restrictions are correct to point out that, despite being made to rely on a smaller quantity of less-performant hardware, Chinese AI labs have succeeded in building globally competitive models with a fraction of the compute available to their American counterparts. DeepSeek’s R1 rivaled frontier American systems at a fraction of the training cost; and Chinese open-source models — led by Qwen and DeepSeek — now account for nearly thirty percent of global open-source AI usage.
But there is a world of difference between building a model and serving it to the planet. The latter is the crux of the diffusion race U.S. and Chinese AI labs today find themselves locked into. Even though H200 exports to China will primarily serve domestic AI workloads, the absolute increase in compute will allow Chinese AI companies to more readily serve their products to global publics — undercutting the Trump administration’s aspiration to export the American AI stack.
Chinese labs continue to face critical compute shortages. Zhipu AI, which released its open-weight GLM 4.7 model in December, was forced to cap new sign-ups for its coding product to 20 percent of prior daily registrations after demand overwhelmed its servers. DeepSeek logs a steady drumbeat of service degradation events — the telltale signs of an inference fleet straining under load.
Proponents of relaxing U.S. export controls have consistently described the Trump administration’s decision to resume licenses for H20 and H200 as harmless because they trail the cutting edge by a generation. This misunderstands what the chips are for. Chinese AI companies do not necessarily need Blackwell-class GPUs to serve inference at global scale; they need a prodigious volume of decent compute to keep their APIs responsive and their prices low as they expand into markets from the Gulf to Southeast Asia to Latin America. Chinese AI executives themselves identify computational power as the one bottleneck constraining their global competitiveness. The correct response to that admission is to tighten the pressure, not ease it.
A Higher Fence Prevents the Yard from Expanding Further
If there is a legitimate criticism of the current U.S. technology control framework, it is not that the yard is too large but that the fence is too short. The $2.5 billion Supermicro indictment exposed an uncomfortable truth long known to U.S. export control enforcement officers: The current U.S. compliance system, which runs on paper declarations, end-use certificates, and the good faith of intermediaries stretching across dozens of jurisdictions, can be defeated through a method as simple as relabeling shipments in a rented warehouse.
The Chip Security Act, which the House Foreign Affairs Committee advanced on a bipartisan basis this week, could address this problem. Location verification for exported chips does not expand the yard, but raises the fence for protected technologies. Such an approach is fundamentally pro-diffusion: if the United States can confirm that chips shipped to Abu Dhabi or Singapore have not in fact been rerouted to Shenzhen, the case for loosening license requirements for other global markets is made materially stronger.
America’s technology siege is working as intended. The task now is to maintain it – and reinforce it where the fence has gaps — long enough for American compute, software, and APIs to embed in global markets, before Chinese chipmakers become capable of contesting the global AI market on their own terms.