The Right Way to Sell Chips to China
Current export rules focus on keeping chips a generation behind. They should focus on keeping America's total compute ahead.
Alasdair Phillips-Robins and Noah Tan — April 13, 2026
Last December, President Trump announced that the United States would allow Nvidia to sell its powerful H200 AI processors to customers in China. Officials in the Trump administration have long argued that the best way to win the AI race is to promote the export of US technology around the world, not to restrict it. Selling H200s, the administration claims, will boost the market share of US chip-makers while preserving the US hardware lead.
Critics warn of national security risks to selling advanced chips, but the administration appears committed to its course. Taking a pro-export framework as a given, policymakers can balance between market share and national security by managing a quantity that currently doesn’t receive enough attention: the United States’ relative compute advantage. Total AI compute for a country is calculated as its stock of AI processors weighted by the effectiveness of each processor; the US currently enjoys a roughly 10-to-1 compute advantage over China. That advantage matters because more compute can support more domestic R&D, more customers served by American AI products, and more ability for the US government to influence the development and use of AI.
Relative compute advantage is about the quantity of chips as much as the quality. Administration officials emphasize that the H200 has been superseded by Nvidia’s more powerful Blackwell generation, but many of the world’s largest AI supercomputers still use H200s. With enough of them, Chinese developers may be able to train and deploy AI models that are competitive with US models. Chinese companies have reportedly placed orders for more than 2 million H200s already.
Whereas Biden-era export controls attempted to make the US compute advantage as large as possible, an export-friendly framework could instead focus on maintaining a fixed, favorable compute advantage. To do so, policymakers should peg the quality of exported chips to the performance of China’s domestically manufactured alternatives, while capping quantities of those US chip exports.
A Return to Relative Advantage
Before the US government launched its all-out chip war with China, in 2022, American policymakers generally sought to maintain a one-to-two-generation advantage in key technologies. When the first Trump administration restricted sales to Chinese chip producers, in 2018, for example, it cut off only a limited set of the most recent semiconductor manufacturing equipment, and placed targeted restrictions on specific companies seen as national security threats.
The United States abandoned relative advantage in 2022. When the Biden administration imposed sweeping new controls in 2022, it declared that a policy of relative advantage was no longer viable. Then-National Security Advisor Jake Sullivan explained that the United States would henceforth seek to “maintain as large of a lead as possible” in certain foundational technologies, including advanced logic and memory chips. In practice, that meant setting a fixed threshold for chip performance, one that the administration progressively lowered, curtailing an expanding swath of China’s chip imports. At the same time, the Biden administration attempted to degrade China’s ability to produce rival chips by imposing increasingly draconian restrictions on semiconductor manufacturing equipment.
A revival of the relative framework. The second Trump administration has signaled a return to the logic of relative advantage. Rather than impose a general blockade on chip exports above a fixed level, the administration is reportedly planning to sell chips that are “about 18 months behind the state of the art.” Earlier last year, when the Trump administration approved exports of Nvidia’s inference-focused H20 chip, Commerce Secretary Howard Lutnick defended the policy of selling Nvidia’s “fourth-best” chip to China.
A widening hardware-performance gap. Yet critics point out that US-made chips are so far ahead of Chinese chips that even chips well behind the US frontier are better than anything China can make. Compared with China’s best chip (the Huawei 910C), the H200 has about 32% greater processing power and 50% more memory bandwidth. Huawei forecasts that its future offerings will improve more slowly than Nvidia’s, so that gap will only grow.
Current export controls don’t guard the US relative compute advantage. In January, the Trump administration took a step toward managing the volume of chip exports, releasing rules that would allow Chinese customers to buy up to half as many H200s as are sold in the United States. But the rule could still allow as many as 2 million sales. The current ratio between US and Chinese compute is likely somewhere between 9 to 1 and 12 to 1, so allowing China sales equivalent to 50% of US sales would give Chinese companies a major boost compared with where they are today.
The current approach, then, has two gaps. First, it runs the risk of significantly raising Chinese AI capabilities, by allowing sales of chips whose capability reflects US chipmaking progress rather than China’s domestic chip quality. And, second, its limitations on sales volume, while a step in the right direction, are probably too generous to China. The administration can do better on both fronts.
Selling Chips Without Selling Out
When deciding which chips to approve for sale in China, the administration should peg export approvals to the performance of Chinese chips, not the US frontier. It should allow chip sales matching the performance of China’s latest widely available domestic offerings. That way, US chip designers will be able to compete with producers like Huawei in the Chinese market, but Chinese developers will not automatically reap the benefits of faster US progress in chip design. Considering only chips that China can produce in commercially relevant quantities — not one-off high-performance prototypes — would guard against efforts by Chinese companies to game the system.
On the question of how many chips to allow, the strategic importance of compute is a strong point in favor of restraining not only China’s access to the very best chips but also its ability to amass large quantities of sub-frontier-quality hardware — chips that could power military systems, intelligence analyses, and industrial robotics. Ultimately, the US should look to relative compute advantage to decide the quality and quantity of chips that it exports.
Selecting a target compute advantage. It is a positive sign that the administration has adopted a ratio approach by capping Chinese sales of relevant chips, but its 2 to 1 ratio is probably too generous. Aiming for a relative compute advantage a bit below the current ratio — perhaps around 8 or 9 to 1 — could still allow significant chip exports without giving China such a large step up. If the 50% ratio is applied to future approvals of more powerful chips, the gap will narrow even further. Over time, the administration should tighten its rules, especially if it considers future, more powerful chips for export. That way, it may be able to keep some Chinese developers in the US chip ecosystem while avoiding a free-for-all that could turbocharge Chinese AI at the expense of US labs and startups.
A more stringent approach was recently put forward by US Rep. John Moolenaar (R-MI), who chairs the House Select Committee on the Strategic Competition Between the United States and the Chinese Communist Party. Moolenaar argues that US officials should approve exports only when total US AI data-center compute exceeds Chinese compute by at least 10 to 1. If the ratio dips below that level, the US would cut off sales; if the ratio rises above it, the chips could flow again.
Making Relative Advantage Work in Practice
Shifting export-control policy to focus on the overall compute balance will bring practical difficulties. Just as the Biden administration struggled to set the right technical thresholds for its controls on chips and semiconductor manufacturing equipment, so the Trump administration will find it tricky to agree on a given compute ratio. Any such figure will need to be informed by an assessment of China’s existing AI computing power, which is difficult to ascertain.
Analysts will have to estimate, among other things, China’s existing installed chip base, the production capacity of Chinese fabs, the volume of legal imports, the smuggling rate, and the lifespan of different kinds of chips. There is already serious public disagreement about some of these figures: the Trump administration estimated last year that Huawei could produce 200,000 of its Ascend AI chips in 2025, while the outside group SemiAnalysis put the figure at about 800,000. Estimates of the rate of smuggling are similarly uncertain.
The trouble with yearly estimates. One suggestion, advanced by Moolenaar, is for the intelligence community to produce a yearly estimate of the US compute advantage over China. But US spy agencies are not geared toward this kind of commercial intelligence. Furthermore, updating the number just once each year will magnify the effect of any errors: if the estimate is wrong in either direction, the US government will approve or block a year’s worth of orders based on faulty information. Because chips last for several years, any mistakes will have ramifications well after the estimates are corrected. DeepSeek’s 2025 models, for example, were likely trained in part on Nvidia chips imported before updated controls went into place in 2023.
Which way to err. These are thorny but not insoluble issues. The US government can rely on both intelligence and external industry sources to estimate Chinese production. Variables like chip lifespan and smuggling rates can be known with at least some degree of confidence. The real trick will be choosing how to be wrong: if the biggest risk is underestimating Chinese progress and allowing too many chips to flow, policymakers should pick a conservative ratio. If instead the true risk is hurting American companies that could be selling to China — without compromising national security — they should err on the side of allowing greater sales.
The Trump administration has committed to selling powerful US AI hardware. That approach risks exporting increasingly advanced chips in essentially unlimited numbers — a surefire way to erase one of the main US advantages in the AI competition. Relying on compute ratio is not a perfect answer, but it offers the best chance of preserving the US advantage in AI while carving out more flexibility for industry to sell to China.
See things differently? AI Frontiers welcomes expert insights, thoughtful critiques, and fresh perspectives. Send us your pitch.
Alasdair Phillips-Robins is a fellow in the Technology and International Affairs Program at the Carnegie Endowment for International Peace, where his research focuses on emerging technology and national security. From 2023 to 2025, he served as a senior policy advisor to U.S. Secretary of Commerce Gina Raimondo, where he covered AI, semiconductors, export controls, and other emerging technology and international issues.
Noah Tan is the James C. Gaither Junior Fellow for the Technology and International Affairs Program at the Carnegie Endowment for International Peace. His work centers on AI supply chains and international technology competition. Previously, he was a research affiliate at Stanford’s Center for International Security and Cooperation and the Hoover Institution, where he worked on international security and economic statecraft. He holds a B.A. In International Relations with Honors and Distinction from Stanford University and is a 2027 Schwarzman Scholar.
Spending billions on narrative intensive marketing that pretends China is even interested in western chips will not help us sell more chips to China.
This is directionally right, but the debate is still too chip-centric.
The underpriced variable is system adaptation: model efficiency, software optimization, deployment discipline, and domestic substitution. A policy designed around product generations can still lose the contest if the other side compounds faster at the system level.