America can and should manage the risks of AI without slowing the race against China. Xi Jinping’s visit will test this prospect.

When Chinese President Xi Jinping lands in Washington on September 24, “guardrails” on artificial intelligence (AI) will be on the agenda—the same word US President Donald Trump used after meeting President Xi in Beijing in May. It arrives at an odd moment in public discussions about artificial intelligence. Days earlier, two of the most important figures in American AI staged a public argument over whether the country should slow down at all.

Anthropic’s Dario Amodei published an essay, “We Must Pace the Frontier,” arguing that AI models are now improving so quickly and “recursively” (ie independently) that developers should deliberately slow the rate of capability gains to let safety catch up. David Sacks, the co-chair of the President’s Council of Advisors on Science and Technology, answered in effect: “Go ahead. You are the frontier; if your unreleased models frighten you, slow them down yourselves. But stop pretending you need Washington’s permission, an antitrust waiver to coordinate with rivals, or a friendly third party to certify everyone else.”

Read as a domestic regulatory spat, the exchange is inconclusive. Read through China, it resolves—and it should shape what the United States says to Xi.

Both men are half right. Amodei is correct that the risks are real and that verification makes safety credible. His one unilateral commitment—letting “embedded evaluators” inside AI companies to supervise development—is the strongest part of his plan. Sacks is correct that a market-leading duopoly of OpenAI and Anthropic asking the government to suspend antitrust so competitors can jointly restrain output looks exactly like what antitrust exists to prevent, whatever the motive. He is also right about the hard fact both essays circle: China will not sign a real pause, and a slowdown it does not match is not caution. It is unilateral disarmament.

The way out is a distinction neither man draws cleanly. The United States should pace the risk, not the race.

Pace the AI Risk, Not the Race

Pacing risk means hardening how frontier models are built and tested—embedded evaluators, interpretability, alignment audits, incident reporting. It is largely internal, it costs little strategic ground, and much of it can be done unilaterally and now, as Amodei proposes. Pacing the race means slowing capability itself in the hope that Beijing reciprocates. That is the move the country cannot afford. America’s lead is roughly eight months by most estimates—a margin China openly believes it can close, and one it attributes, in its own words, mainly to US export controls. This year’s wave of capable Chinese open-weight models—DeepSeek, Qwen, Kimi, Zhipu—and Huawei’s maturing Ascend hardware stack are the evidence that the gap is narrowing, not widening.

That distinction should be the American posture on September 24—because Xi will arrive selling the opposite.

Trump and Xi’s AI Guardrails

The template was set in May. The two leaders talked “guardrails”; within days, Washington cleared sales of Nvidia’s H200 chips to Chinese firms. Expect the same choreography: warm language about “working together” on AI safety, offered as the wrapper for concessions on chips and manufacturing equipment. The United States should refuse this trade. Verifiable controls that constrain China’s compute are the single most effective safety measure America has. Trading them for unverifiable promises of restraint is not guardrails; it is giving away the guardrail.

Genuine cooperation is still worth pursuing—but only where it is narrow, mutually self-interested, and verifiable. A bar on using frontier models to synthesize biological weapons serves Beijing as much as Washington. So does an agreement, of the kind reached at the working level in 2024, that humans and not machines control the use of nuclear weapons. These are floors, not speed limits, and they are the realistic ceiling of what the September track can deliver. Anything grander—a jointly monitored cap on the rate of self-improvement—founders on the problem Amodei himself concedes and Sacks names bluntly: neither side can verify the other’s secret models, and the incentive to defect is enormous.

America’s AI Lead Depends on Allied Hardware

A deeper reason to hold the line remains. America’s lead is not really America’s alone. It rests on an allied hardware base—Taiwanese fabrication, Dutch lithography, Japanese and Korean materials and tools—that no rival can yet replicate. That is the country’s most durable advantage and its least fungible one. It should be defended as infrastructure: controls enforced, model weights secured, distillation of frontier models by adversaries treated as the theft it is. It should not be spent on the atmospherics of a summit.

China Is Competing Through Open-Weight AI

But denial is only half a strategy, and the more brittle half on its own. Export controls limit the compute China can buy; they do nothing about the models China gives away. Chinese labs—DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi, Zhipu—are releasing capable open-weight models into the world for free, and developers from Jakarta to Nairobi are building on them because they are good enough and available today. At July’s World AI Conference in Shanghai, Beijing lined up AI partnerships with some 28 countries. Each is a wager that the next layer of the world’s software will sit on Chinese foundations.

That is a contest export controls cannot win, because it is not fought in chips. It is fought in the model layer, where whoever the global developer base standardizes on inherits the defaults, the dependencies, and the leverage—the lock-in that made American operating systems and cloud strategic assets for a generation. 

America Needs Both AI Denial and Diffusion

So the strategy has to run on two prongs, not one. The first is denial: guard the closed frontier, secure its weights, treat distillation of it as the theft it is. The second is diffusion: field American and allied open-weight models good enough to become the world’s default, and put real weight behind their adoption. These do not contradict each other—protect the crown jewels at the top, and compete for the mass market below them. Denial keeps China a step behind; diffusion keeps the world building on us.

So who sets the pace? Not, in the end, a lab in San Francisco. The pace of the frontier will be set by whether Washington can do two things at once: harden safety at home without slowing the race abroad, and write verifiable floors with an adversary without mistaking them for a truce. Amodei is right that the country needs more time. Sacks is right that it cannot ask China for it. The time comes from the lead—and the lead is kept or lost in decisions like the one on the table on September 24.

The question for that meeting is not whether America will pace the frontier. It is whether it will give away the reason it still sets it.

About the Author: Jason Hsu

Jason Hsu is a senior fellow at the Hudson Institute, where he focuses on the United States’ technological cooperation with allies and partners. From 2016 to 2020, Mr. Hsu served as a legislator-at-large in Taiwan’s Legislative Yuan (the national parliament), where he focused on defense, technology, trade, and foreign policy.

The post Why the US Should Pace the Risk but Not the Speed of the AI Race with China appeared first on The National Interest.