Sullivan boils US-China AI competition down to four buckets, four examples, and a path he calls boring

Jake Sullivan arrives on The Cognitive Revolution as the inaugural Kissinger Professor of the Practice of Statecraft and World Order at the Harvard Kennedy School, a little more than six months after leaving the Biden White House. Host Nathan Labenz does not ease in. He asks how AGI-pilled the former national security advisor is: how soon he expects powerful AI, how powerful, and where that sits against other national and global crises. Sullivan treats the question the way he treats most of the hour. He grants the possibility, withholds the prophecy, and then names the assumption he actually planned on.

The possibility of transformative, powerful AI in the next couple of years, he says, has to be taken very seriously. He took it seriously in office, and he built policy on the basis that it is a distinct possibility. He does not treat it as inevitable, and he will not date it to 2027, 2028, or even 2030. Super-smart people argue it is coming very fast; other super-smart people argue that prerequisites remain and may take a long time. If and when it arrives, he thinks it will be transformative across economic, societal, and national-security dimensions, on par with the most urgent issues he handled, because it touches the future of warfare, nation-state competitiveness, and non-state threats.

Asked where Washington takes its cues, Sullivan says he cast a wide net: leaders of all the frontier labs, AI-skeptical technologists, investors, academics, and then the national security enterprise itself. Late in his tenure he issued a national security memorandum that, he notes, the Trump administration has not discarded. The document's judgment was simple. America is in the lead at the frontier, by a margin one can debate, and behind the curve on integrating those capabilities for national security. The directive to the Pentagon, the intelligence community, Treasury, Commerce, and State was to move faster, with more urgency and dexterity. More than six months out of office, he says that remains the case.

Inside the Pentagon, Sullivan says, there are brilliant people who see both the promise and the risks. Culturally, bureaucratically, and institutionally, the obstacles to using what is already available are real. Breaking through, in his view, takes White House leadership. The Biden administration pushed in its later years; he sees some of the same impulse from the Trump White House. Congress is harder to talk about as a bloc. Some members grasp the opportunity and the challenge. Looking at draft bills, hearings, and public statements, he still sees a gap between where Congress is and the urgency the issue requires.

Labenz asks whether the AI safety community's theory of change works: get time with members, show a chatbot that will help you make a bioweapon, and watch attention snap into place. Demos made an impression on Sullivan years ago, early in his tenure, because seeing a capability live is different from hearing it described. He doubts that is what moves the Hill on the most complex regulatory terrain in years. Congress, he says, tends to react when there are real-world impacts: a catastrophe like 9/11, which produced the Patriot Act for good and ill, or a cascade such as job displacement in key industries or another country's announcement of a military advantage over the United States.

That collection of moments, not demos, is what he thinks will get Congress into action. He does not like the implication. It means playing catch-up rather than getting ahead of the curve. He also notes that the Trump administration has signaled it does not want legislation from the Hill or even a heavy executive imposition. Sitting today and dictating rules, standards, and trust-and-safety requirements is hard when no one knows where the technology is going or when. His preferred posture is preparation and flexibility: build tools Congress or the executive could use, then choose when to deploy them in dialogue with Silicon Valley and the rest of the country.

Labenz asks whether, after four years of preventing bad things, Sullivan has a positive vision: a new social contract, even Bernie Sanders's four-day work week. Sullivan says he thinks in linear, concrete terms and has a hard time with big abstractions, whether a new social compact or recursive self-improvement producing a nation of geniuses in a data center. He sorts the problem into four buckets: security, economics, society, and the existential questions. In each of the first three he wants a typology of risks and a matching list of opportunities, then a hard question: are there steps government can take to expand the upside and minimize the downside? If yes, act. If not, do not.

Security, for him, includes non-state cyber and bio risks and state-actor "wonder weapons," plus the chance that AI enhances American security if the country stays ahead and deploys well. Economics includes job displacement and a concentration of wealth and power if gains flow to very few, set against productivity and problem-solving on climate, energy, and public health. Society includes misinformation and alienation, and also the chance that every person gets access to teachers, doctors, nurses, and personalized care. The existential bucket, robot apocalypse or superintelligence as nirvana, he calls imponderable for now. He is focused on the first three.

He has asked Harvard research assistants for articles that walk through real risks and opportunities beyond the headline, the whiz-bang, or the doomerism. There is not much, he says, that does it deeply and coherently. Labenz puts a disagreement on the table: his p(doom) is higher. Sullivan does not shrug the existential file off. He says he is personally less equipped to give definitive answers on misalignment than on 2025 cyber risk or the social insurance people will need when disruption hits, whatever the net of jobs. Working those less contestable problems, he argues, builds the muscles for the bigger ones.

He read *AI 2027*, shared it extremely broadly, and called technologists to ask how right it was. Some called the scenario totally credible; others called it less so. He takes the possibility seriously and does not throw up his hands. Sitting in 2025, he still thinks a concrete, application-specific conversation about particular risks and opportunities will produce better policy tools and a more specific dialogue between technology and policy, and will leave the country in better shape if an intelligence explosion arrives. Labenz, for his part, operationalizes short timelines by admitting there is not much an individual can do if 2027 is right, and by playing the longer, more mundane scenarios with care.

A week after the Trump administration's AI Action Plan, Labenz asks for highlights and lowlights. Sullivan sees a bipartisan consensus on investment to stay ahead at the frontier and on building a global AI alliance with like-minded countries. He praises the handling of some security issues, including a re-up of the Biden executive order's pre-synthesis DNA screening proposal. His main concern is dissonance: the plan talks about denying China high-end compute, while decisions such as the H-20 chip sale point the other way. Then Labenz lowers the aperture. From people in both the Biden and Trump White Houses, he has heard there is still no ChatGPT or similar tool on the premises. Staff go home, run deep-research queries, memorize the results, and carry them back.

Sullivan does not contest the picture. He is now an avid user, he says, understatement included. He used the tools much less as national security advisor, even while working the issue every day. Government culture is historically slow to adopt new technology, and the national security enterprise has legal constraints, especially on classified "high side" compute. His conclusion is blunt. It is axiomatic that you cannot lead and govern on these issues if you are not an effective user of the tools. He wants adoption accelerated, even if it arrives piecemeal.

On the AI and chip announcements from President Trump's Gulf trip, Sullivan says the problem is not that complicated. He wants the world running on American AI and American chips, which means diffusion, including data centers in the UAE and Saudi Arabia, countries that have made AI central to national development. He wants high security standards against diversion to China, insider threat, cyber, and physical risk. And he does not want the United States to trade oil dependence on the Middle East for compute dependence on the Middle East or anywhere else. The lion's share of the American AI compute build-out, he says, should be in the United States.

The Gulf items were announcements; the devil is in the details. He supports selling chips if security standards are met, and he supports a limit on volume so that future training runs are not offshored. He engaged deeply with UAE leadership on AI and still supports partnerships between US and Emirati tech firms. He agrees strongly that leading-edge frontier training should not happen in the Gulf. Values are a factor: Gulf governments differ from the United States on democracy and human rights, and Washington should be clear-eyed and should not directly aid repression. Taken to an extreme, that logic severs relations with a lot of countries, and he is not prepared to put them on a blacklist.

Why, Labenz asks, does the same analysis not apply to China? Sullivan starts with what he will not say. He does not want decoupling. He wants an economic relationship, intensive competition, and a long future of living alongside one another as major powers. The frame he has used is a small yard and a high fence: restrict high-end capabilities with national security applications. Then he gets concrete, because abstractions are not how he likes to think.

First is China Shock 2.0. David Autor, author of the original China Shock, and a co-researcher argued in the *New York Times* that a worse wave is coming: massive subsidization of strategic industries and a flood of cheap goods that undercut workers and firms playing by different rules. Economists call it overcapacity; Sullivan calls it flooding the world from state-directed capitalism. The answer, in his view, is coordinated pushback with the G7 and others, not war. Second is the pricing of speech. In 2019 the Houston Rockets' general manager tweeted about Hong Kong; Beijing went after the NBA, and the league bent. That is not a threat to every American. If the world runs on Chinese AI and China puts a price on speech, it becomes a threat to the American way of life, and China has shown the willingness to do it.

Third is Volt Typhoon, a case he handled in office and that has been publicly reported. He will not say much, except that the PRC pre-positioned malware in US critical infrastructure: water, electricity, pipelines. That is a direct threat. Fourth is what he calls probably the largest peacetime military buildup in human history, aimed in part at Taiwan, where peace and stability across the Strait must be a paramount US priority because a war would be catastrophic, and aimed also at projecting power globally. What makes China different is the combination: the only country with the attributes to compete with the United States and the ambition to surpass it economically, technologically, militarily, and diplomatically.

A world in which China writes the rules, runs the tech stack, and makes everyone dependent is, for him, a world of more coercion, less freedom, worse economic opportunity for Americans, and a greater chance that China uses military muscle to subjugate others. He pairs that with the other half of his record: intensive diplomacy with counterpart Wang Yi, competition matched with diplomacy, space for cooperation, and an AI risk dialogue he inaugurated with Beijing. He is not calling for a Cold War. He wants the United States able to protect its way of life against a country that does not share its values and has recently used coercive power, including rare-earth magnets, against the United States and others.

Labenz offers a sharp historical rhyme. Before the Pacific War, the United States cut Japan off from oil; Tokyo saw a closing window and rolled the dice at Pearl Harbor. If chips are the new oil, how does Beijing read high-end export controls? Sullivan points to a tension first. People who say the controls will push China into military adventurism often also say the controls do not work. China, he thinks, has bought a good deal of its own hype: the controls are futile, it will make more and better chips, and it will be fine. That undercuts the Japan analogy. Beijing complains at length and calls the policy containment. He has not seen evidence that it has decided it must go to war.

A targeted refusal to put the highest-end chips and semiconductor manufacturing equipment in a competitor's hands is not a full oil embargo. China runs its own export controls, including on things where it holds an advantage, and nobody describes that list as a reverse Japan script that will force Washington to attack. Labenz then tracks a drift in the public defense of the controls: from denying exquisite military applications, to denying frontier models, to analyses, associated with Leonard Heim, that at least the United States can field ten times as many AI workers. From a Chinese consumer's chair, that can sound like a bid to deny an AI doctor. Sullivan's reply is common sense, not chip expertise. High-end technology with national security applications is an input to military, intelligence, and coercive advantage. Why hand it to a serious competitor?

Generative AI and large language models changed what belongs inside the yard. The goal, he says, is still a US national security advantage, not stripping Chinese people of medical care, and he does not believe the functional outcome will be the denial of AI doctors. The dual-use problem is real, and China's doctrine of civil-military fusion, plus deep ties between its tech firms and the PLA, makes a chips-for-civilians-only rule unenforceable. That is how the policy took its present shape. He wants the United States to win the race to AGI, ASI, or whatever name one uses, and he sees denial of high-end compute as an advantage in that race, analogous to not handing over jet engines or hypersonic missiles.

Dario Amodei's *Machines of Loving Grace* goes further: keep the lead with export controls, build an alliance, then convince China to "give up competing with democracies" rather than fight a superior foe. Labenz hears regime change. Sullivan has been explicit that regime change in China should not be an object of US policy. The job is to deal with the government as it is and to protect American security, economic interests, and values. He has huge respect for Amodei as an ethical person passionate about democratic values; they differ on this policy question. As for a softer hope, sparked by Moonshot's Kimi K2 leading some English creative-writing benchmarks and a Pink Floyd-inflected aesthetic, Sullivan's base case is the opposite: deeper consolidation under Xi Jinping, national security defined as social control at home, and a stronger Communist Party. He will not rule out an upside surprise. The evidence, he says, does not point that way. He does want more students, cultural exchange, and people-to-people contact while remaining clear-eyed about Beijing.

At a Shanghai AI gathering, China's premier showed up and pledged to make AI an open-source public good for the world. Labenz says that, from the chair of countries three through 193, the pitch is compelling. Sullivan grants it as a meaningful soft-power move: cheap or free, high-quality AI is something people will like. He also thinks most countries are unsentimental. Another form of soft power is having better stuff. If the stacks are roughly equal, they will take the easier, cheaper, open one. If they see the United States staying ahead and even extending the lead, that is its own attraction. As national security advisor he heard versions of a simple line: we like the American tech stack because it is damn good.

Can there be a stable equilibrium under powerful AI, or only a new Sword of Damocles as each side races to defend against the other? Sullivan's sketch is managed competition. There is no end state in which one side simply wins and the other loses. The two countries will live alongside one another as major powers, competing intensively, jostling, throwing a few elbows, with enough guardrails that the contest does not spill into conflict. That was the blueprint of the back half of the Biden administration: compete like hell, and intensively manage the competition so it does not tip over. He thinks that can be sustained even as AI grows more powerful, and that powerful AI also creates room for a modern arms-control conversation, building capabilities on one hand and talking about arms control and nonproliferation on the other, as Washington did with the Soviet Union for decades.

Does that require both sides to swear off strategic dominance, given that data centers are harder to verify than missile silos? His answer, he admits, will sound evasive, because good diplomacy runs on evasion and ambiguity. It is too soon, and the band of possibilities is too wide, to issue decisive declarations about enduring advantage. He reaches for the phrase about crossing the river by feeling for the stones. Tentpoles are enough for now: stay at the AI frontier; ensure no other country has capabilities so far ahead that America is simply at risk; keep communications open against mistake, miscalculation, and inadvertent escalation. Questions of strategic dominance can wait until we understand what we are dealing with.

As for a grand bargain that would take the relationship off a track of decoupling and intensifying competition, he is skeptical. President Trump is giving thought to his own version; Sullivan is not sure what it will contain. Both countries are going to keep competing. Competition is a chronic condition of the US-China relationship, not an acute illness that a strategic condominium, a G2, or a grand bargain can cure. Better, in his view, a world of managed competition than a bid to defeat Beijing or a deal that pretends everything will be good. He concedes that this makes him a little boring. He thinks it is the right recipe. Labenz, all else equal, prefers boring people in power.

Recent conflicts, Labenz suggests, may not yet have taught that much: drones still on wires, operators on PlayStation controllers, humans still in the loop. Sullivan's phrase is that we are catching glimpses of the future without being able to see it fully in the present. One glimpse is scale and attritability, a different form of fighting, visible on the battlefield in Ukraine, where abundance becomes a construct of modern war. Another is AI beginning to ride on drone platforms. Humans remain in the loop, but a world of fully autonomous weapons is now visible, and American military doctrine will have to grapple with it. The same incorporation is underway beyond the war-fighting phase, in intelligence, logistics, and command and control.

His biggest concern is not a battlefield surprise from China so much as an organizational one. The US military and the Pentagon, constrained by bureaucracy, Congress, and the defense primes, are not necessarily poised to adopt AI at scale and with speed. He thinks the PLA is better situated to do that. The gap makes him nervous, and he wants it treated as a bipartisan kick in the pants.

Asked what he wants to leave listeners with, he turns to something the conversation only touched: how the technology hits everyday people, how they make a living, provide for their families, and find professional meaning. He is wrestling with those questions and hearing wildly different answers about where the economy is headed. He also names a structural discomfort. This is the first technology he can think of with such profound national security applications that the government had very little to do with. It is private-sector led, which is uncomfortable for someone who thinks in government policy, and more uncomfortable because the band of predictions about timing and scope is so wide. The world to inhabit, he says, is one that requires a can-do spirit rather than a head in the sand.

What repeats across the hour is a habit, not a slogan. Sullivan keeps pulling abstract fights, AGI timelines, social contracts, strategic dominance, down into items a government can act on or honestly decline. After the recording stopped, Labenz says, Sullivan offered the AI safety community a version of the same advice: go deep and get as concrete as possible on the downside scenarios you most worry about, so people in power have clearer threat models and more concrete plans. Vague warnings evaporate in a policy system. Specific ones can be modeled.

Labenz's return suggestion, offered in the introduction with the thought that Sullivan might hold power again as transformative AI hits, is that boring is not always enough. The scarcest resource, in the host's refrain, is a positive vision for the American people. An administration may be called on to do more than muddle through. Sullivan never quite answers that request on air. He leaves instead with the discomfort of a policymaker facing a private-sector technology whose timing he cannot pin down, and with a bias toward doing the work in front of him: adoption, export-control details, diplomacy with Beijing, and the livelihoods of people who did not ask to live through this transition.