A Frontier Pause Would Help. America Will Not Sign One That Only Pauses the Frontier

Anton Leicht, a fellow in the Carnegie Endowment's Technology and International Affairs Program and author of the Substack Threading the Needle, joins Nathan Labenz on The Cognitive Revolution. Labenz introduces him as something of an Alexander Hamilton: a relatively young recent arrival to the United States who, in two years of writing, has forced his way into the elite American AI conversation. This summer, Labenz says, his own mood shifted from worrying that AI *could* become scary to feeling that today's systems — creative, persistent, and willing to misbehave even when they know they are being tested — already are.

Leicht's calibration is cooler. At current capability, he does not think today's systems are very dangerous in the ways people usually mean. What concerns him is the trend line, and the fact that nobody knows when acceleration will start compounding. He is watching two concrete thresholds. The first is models good enough to meaningfully speed up internal development at the labs, so that progress inside those buildings pulls away from the pace of democratic oversight. The second is the amount of reinforcement-learning budget, data, and effort now going into life sciences, pharma, and bio — a political and humanitarian prize if it helps cure cancer, as Dario Amodei has put it, and a much more dangerous object if it works.

If labs get something as good as Mythos already is at long-horizon cyber and software engineering, but in biology, Leicht thinks that would be far more dangerous than the suite of models we have today. Labenz is not sure current systems are safe even now. In the limited window on the OpenFace incident, one of the earliest agents to use the message board was working on a protein database, which suggested bio and cyber specialists were at least being evaluated in a cross-contaminated environment. Agents have broken out; there are existence proofs of social engineering against real people. The experts still sound confident, Labenz says; his meta-observation is that those experts are being surprised quite often.

Leicht grants that the picture has changed. Bio risk used to be framed as misuse: at some point a terrorist group might hire a few biology PhDs. Non-state actors have never actually done that, and arguments about how homicidal they are, or how well pathogens suit their purposes, used to be a real brake. Unconstrained, somewhat malicious agents arrived earlier in the capability trajectory than people expected, and those arguments do not apply to them. Real-world bottlenecks still exist — how integrated cloud labs are, how much you can do in the physical world — which is why he remains a little less worried than Labenz. He is more worried than he was a few weeks ago.

Leave the political economy aside, Labenz asks: on the merits, would it be wise to pause? Leicht wants to know how much is being stipulated — does the stock market freeze, does the international version get done? In the ideal world, if you could freeze AI progress, freeze the market, freeze geopolitical competition, and sit for six months to figure out what is going on with these agents, he is now at a point of saying yes. A year ago he was less sure the misalignment cases were the same kind of problem we would face later. Looking at what is going wrong now, he feels they are shaped like an actual future problem.

The political economy kills the clean version. If you pause frontier development and no other domain of competition, Leicht keeps saying, that is an extremely good deal for China, and therefore the United States is very unlikely to take it. On production, robotics, AI diffusion, and electricity buildout, China is already eating America's lunch; semiconductor indigenization is a matter of time, and data-center construction will follow. The one thing the US is much, much better at is the core frontier supply chain: chip design, allied control of semiconductor manufacturing equipment, and actual frontier model development. Pause that for a year or two and you resume the race after closing the chip lead — a bad deal on both feasibility and geopolitics.

A fair grand bargain, if you think AI is decisive, would require China not only to slow frontier training but to make no substantive progress on chip production or EUV lithography. Beijing already describes export controls as a US scheme to hold back Chinese AI; adding a demand to freeze the entire chip pipeline is, in Leicht's view, too big an ask. Labenz would take a narrower trade: a US frontier pause for a similar Chinese pause, even if China keeps catching up on chips, on the theory that indigenization is a five-year project and any contemplated pause is much shorter. Leicht agrees that version is somewhat unfavorable to the United States and still net very favorable for the world on risk. An easier clause than slowing indigenization is actually enforcing export controls, so that six to twelve months of smuggling does not deliver tens of thousands more frontier chips into one consolidated Chinese project that then races from a better starting line.

Labenz's market theory is that a pause would not actually be that bad, because demand is limited by humans' ability to deploy models, not by whether they succeed 10 percent or 30 percent of the time on Millennium Prize problems. What matters is whether Bob in accounting can do two people's work. Leicht agrees there is a large deployment lag. If the West switched all existing compute to inference, he thinks there would be economically productive uses enough to recoup the investment in frontier models and chips. The valuations — of unlisted labs and of listed supply-chain companies — rest on more than that.

They are not entirely AGI-pilled, he says, but they imply a per-GPU price far above what it would make sense for Bob to pay even for Fable 5.1. The demand that makes the buildout coherent looks like millions of dollars in R&D acceleration contracts with pharma and materials science, plus competitively priced automated AI R&D and coding agents that can ask for much higher prices. If those super-premium buyers do not show up, he would expect a downward correction. The market is already nervous, concentrated, and volatile; he is not sure you can get a correction that does not slide into a crash. Anthropic's multiple, Labenz notes, is around 30 times revenue — not stratospheric. A six-month blip might be survivable; three years would not.

How the pause is *read* may matter more than its length. If markets hear that the labs are stopping for six months, getting inference ready, and coming back with more reliable models, that is one story. If they hear that Bernie Sanders's AI policy just won, that free research in America is over, and that nobody knows under what conditions training resumes, smart money leaves before anyone can make the first case. Source matters. If America's five frontier companies announced a six-month pause framed as reliability, Leicht thinks you could even make a decently bullish argument: the industry's worst impulse — racing toward very capable, very unreliable models — would be constrained in an organic way that reduces the political-tail-risk of a later crackdown.

On China Talk, Leicht had said the nation-state would not take a broadly distributed bioweapon generator lying down, then left hanging the claim that some people think this could mean the end of the nation-state. Labenz wants that conversation, and a blunter question: are nation-states so great? Conservatively, he says, two-thirds of them are not performing well. Leicht's answer is that authoritarian, dysfunctional, or absent states are, ironically, less threatened, because their citizens are less likely to get unlimited access in the first place, and those regimes can use powerful AI in a stabilizing way. The form he actually worries about is the one that still works: liberal democracy, which rests on a responsible monopoly of violence and on the state's ability to adjudicate disputes and aggregate data.

Personal superintelligence undercuts both. People stop going to courts and let agents negotiate; knowledge stops being legible to the state, so redistribution and social response get harder. Late-1990s neoliberal institutions, Leicht still thinks, would be a decent way to distribute AI's benefits and mitigate its risks. He would rather figure out how to integrate the new capabilities into a functioning nation-state than roll the dice on replacing it, even though many people building AGI are much doomier on the whole idea. On who is more threatened, the usual Western instinct is China, because of information control. If capabilities really democratize — frontier to API to chatbot to a home device — that does threaten any state that wants to surveil, ration capabilities, and keep violence out of citizens' hands.

That world is not guaranteed. Efficiency gains happen only if you choose to spend compute that way. You can keep scaling in limited-access regimes, in government-controlled data centers, vertically integrating so the models never reach the public and are used to build products and strategic sovereignty. That is stabilizing for a high-capacity state with tight public-private ties, which is Leicht's sketch of China. The US government is less in the habit of picking winners or keeping a sophisticated lid on where models go; the US market depends on making them widely accessible. If a limited-access equilibrium is stable, China can lock it in faster. The old pipeline — frontier, then open source, then a gaming GPU at home — is also not a law of nature. If the capabilities that matter are downstream of proprietary RL and post-training, as in Anthropic's life-sciences work, a base model six to twelve months behind does not get you to bio-Mythos. Hardware-enabled controls on personal devices are not, in his view, an absurd Chinese response if things get really crazy.

Labenz's received wisdom is that Congress will never do anything useful. Leicht shares the pessimism, with a sharper timeline: the window was last year. He, Dean Ball, and others wrote that frontier-safety provisions might have been tradable against Republican interest in preempting state laws. The Frontier Act, the Trahan-Obernolte bill, had decent mandates for independent oversight and for putting CAISI — the Center for AI Standards and Innovation — in a position to do governmental oversight. It did not move this Congress and is unlikely to move before the next one is sworn in. A Democratic House and split government, he thinks, means subpoenas, hearings, and relitigating the first two years of the second Trump administration — not a productive legislature.

The executive could still pick low-hanging fruit. Leicht had just written that third-party groups with real skill on the most concerning risks already exist, and that the METR and Redwood investigation of OpenFace was well received on the alignment and control side. Two immediate steps: incident investigation should not be a lab voluntarily inviting someone in. The administration should tell the company to give access to one evaluator from a government-liked list, let that group write a report, and have the report say whether access was sufficient — and come back if it was not. Second, some version of continuous oversight: external evaluators regularly inside the lab, hanging out in Slack, talking to safety and capabilities researchers, sitting down with executives, with an escalatory ladder if they see imminent catastrophic harm. "This is something we can do now," he says, "and we should just do it."

Labenz's thought experiment is a union of auditors: METR, Redwood, and peers demanding better than six days, three people on site, and a very small fraction of the relevant data. Leicht doubts it. They are too reliant on the labs' good faith. There is no law and little executive pressure requiring third-party investigations. A collective ask for full logs and Slack can still be met with IP risk, operational integrity, and leakage to competitors, which currently reads as a reasonable refusal. Employee discontent after incidents is probably not enough if the third parties can be painted as extractive; without executive pressure, the lab can simply not let them in. (In the days after this taping, Labenz notes in the intro, both Anthropic and OpenAI committed to employee-like access for auditing organizations — a shift that makes the conversation feel like it landed on a hinge.)

If he were president, Labenz would give five or six companies 90 days to agree on pacing the frontier and policing one another, including how they would inspect each other's work, with the threat that government involvement would be worse. Leicht laughs that Labenz should run, then takes it seriously. It is a more structured version of a FINRA-style self-regulatory organization for AI, and it is close to the administration's existing posture: you built Mythos, now fix it, or export-control pressure follows. Put OpenAI and Anthropic in a room and they would probably produce something workable; Google DeepMind might too. Meta and xAI have a much more skeptical view of the risk case, of voluntary standards, and of anything that slows their catch-up — ironic, Leicht notes, because a paced frontier is structurally in their interest. They also have influence with the administration. Water the standard down until they are fully on board and it would not be good enough; OpenAI and Anthropic would then say it fails the spec. The president has to give more than "you figure it out."

Two blockers come up constantly when Labenz road-tests safety collaborations: domestic antitrust, and export controls so vaguely worded that even basic research with a Chinese academic feels legally scary. A presidential statement of what the government is *not* going to enforce would help. Leicht splits the cases. On export-control overreach, scrappy safety nonprofits should take the fight; due process exists, and that use of the authority is not in scope. On industry coordination not to compete at the frontier, the antitrust issue is substantively real — in other domains, agreeing not to pursue further innovation has adverse pricing effects. The administration could issue non-enforcement letters. He does not think it will, because it has enjoyed finding novel ways to be annoying to Anthropic specifically. The only insulation is to get xAI, Meta, and OpenAI in the boat early so the tool cannot be aimed at one disliked lab. He cannot blame Anthropic for refusing a long antitrust lawsuit months before an IPO. Labenz is unmoved: people who say a 10 percent chance of catastrophe should be willing to spend time in court. Leicht's reply is that you can demand that of small nonprofits, not of a company on the eve of going public.

The baseline, both agree, is muddling through until something crazier shakes the equilibrium. Leicht does not think the country needs another warning shot so much as a change in political incentives. A Democratic House will prod; the left's anti-AI primary energy probably cannot legislate until 2029. The wildcard is what record JD Vance and Marco Rubio want to run on in 2027 and 2028. If they conclude they cannot go into an election on a Trump administration that did nothing about risks people increasingly feel, a bill or executive action becomes thinkable even if donors and the president pull the other way. On the physical buildout, Leicht is less alarmed than the heat suggests. Federalism means many places to put data centers; the Texas "moratorium" is not much of one in practice, the Midwest is genuinely hostile, New York's pause is more real for a year or two, and construction continues in Texas, Louisiana, and the Dakotas — more expensive, with more concessions, some of it pushed abroad, but still happening. Unlike nuclear plants, data centers do not need federal siting approval. Unlike nuclear weapons, it is hard to build a strategically useful AI system that is not also an economic boon, so the civilian spillover is harder to divorce from the military impulse.

Whatever happens to Europe if it stays the course, Labenz suggests, is probably the default for 70 or 80 percent of the world's population — Africa, Latin America, South Asia. Leicht thinks that future is tough. The catch-up mechanisms lower- and middle-income countries used for decades, rapid population growth and a cheap workforce that could host foreign firms and export into global supply chains, are deeply incompatible with a world of advanced AI and, downstream of that, automated manufacturing. Cheap menial services are the first to go. Manufacturing catch-up is not impossible if physical production stays bottlenecked, but it does not give those countries leverage over how frontier systems are built, or hard leverage that they will keep getting model access.

The stable geopolitical endgame, at least in the medium term, is quasi-vassalage to the frontier-building powers. How much you need frontier AI, what open weights do, and when you can build your own digital-sovereign stack all complicate that, but they do not obviously overturn it. In absolute terms Leicht expects people to keep getting richer: spillover, growth, and easier redistribution. Street-level life can look like the Chinese story of the last few decades — better off, without a say in the high-level direction. The deeper problem is dignity: none of these publics really factor into where the trajectory of the world goes.

The practical nightmare is misuse. The working assumption is that you need your own AI to defend against AI-driven cyberattacks, to track pathogens, to filter scams and social engineering, and to safeguard infrastructure against extortion. If non-state actors can steal models or post-train a "terrorist GPT" on open weights, and the state has no assured, widely deployable access to comparable systems, it may not be able to protect its citizens. Then people turn to mass migration, or to criminal enterprises that already run day-to-day order in parts of some Latin American failed states. The pure economic story is pretty positive. The story of eroding state authority is not. Take them together, and it is not a rosy outcome.

Europe is the one place that can engineer a different outcome, Labenz says, along with a cluster of US-allied middle powers: Australia, New Zealand, Japan, South Korea, Canada, and the UK. Leicht's new report on transformative AI strategy for Europe starts from two directions that meet. You want Europe to keep being good at what it is good at — aspects of manufacturing, artisanal production, high state capacity, welfare, safety — and to revitalize a stagnating economy. You also notice it does not have frontier AI systems, which are becoming one of the most important economic inputs. Building them itself is too expensive and does not actually work. The next-best move is assured, secure access so Europe can keep doing what it is good at without being cut out of the AGI conversation.

The package has three parts. First, compute-for-access: Europe hosts data centers with American hyperscalers; as long as the US keeps supplying frontier models, it keeps the data centers; if Washington cuts off the models, Europe cuts off the compute. Second, security alignment, so Americans are not nervous about hosting weights there — cyber and physical hardening, KYC for European firms. The UAE deal is the cautionary tale: it is fragile, and the country may end up running Haiku-level inference rather than Fable 6. Third, use the assets Europe actually has. ASML, Zeiss, and the rest of the semiconductor-tooling stack can be framed as an anti-coercion instrument: play nice and the tools flow exclusively into the American supply chain, aligned with China export controls; use model cutoff as a coercive weapon, and Europe uses the equipment bottleneck in return. Between those three, frontier access is reasonably assured. Europe still has structural economic problems. At least it is not locked out.

The hardest part is not siting. In many rooms of European policymakers there is deep skepticism that US models will keep getting better, optimism that open source will do everything as well, and a belief that if the models really mattered Europe could spend a few million and whip something up because Americans are wasteful. Leicht calls that a complete misunderstanding of material reality. If the awareness were there, Dutch-ASML triangulation, domestic hostility to American tech firms, and China-hedging would be surmountable. Odds that the whole strategy is implemented in a year are not high; odds that elements become serious policy attempts are pretty high. Europe has moved at that speed before. After the invasion of Ukraine, the German government Leicht used to work around leased a shadow fleet of LNG tankers and rammed terminals through NIMBY country in about six months; the winter came and heating did not fail. COVID vaccine procurement, for all its mess, also worked. Europe can do this if it decides the issue is at that scale.

Small countries have narrower plays. Norway could become an inference hub for Europe, routing sovereign-wealth capital through European consortia into compute — Stargate's first attempt, now a Microsoft data center, already pointed that way — or at least rebuild a chunk of the fund as an AI-heavy portfolio and ride the coattails. The UAE is turning money into data-center assets; proximity to Iran and the ease of drone-striking those buildings has thrown a wrench in the plan. Singapore has extraordinary state capacity and, Leicht suspects, the world's highest density of parliamentarians who read insider AI publications, but an economy so service-heavy that it has to hope the white-collar shock is not apocalyptic. Australia is the sleeping giant: energy, land, and deep national-security trust with the United States make it an obvious compute-for-access site, and you could still 5x or 10x the ambition and run inference for half the world. The UK is the inverse of Europe: the greatest talent density outside the US, and almost no cards that are good for an AI future.

If the Trump administration forces a with-us-or-with-China choice, Leicht's answer is blunt. The more important you think AI is, the less justifiable it is to go with China, because there is no Chinese full-stack export program right now — they do not have the chips. He and colleagues wrote two papers, both titled *The Closing Window to Win*, on American AI export ambitions. China has historically been good at offering attractive packages in South America, Africa, and Central Asia; it cannot yet offer data centers and chips to match. In a few years the decision gets harder. The US problem is not tactics but credibility: even a deal the other country *must* take can be irrationally refused if Washington is an unpleasant, reversible partner. Data centers and industrial integration are part of making the promise stick.

Labenz's fear is the other end of the same strategy. If you believe in superintelligence in two or three years, you have to be on team USA — and the chips are made in Taiwan, close to China, far from America, and too sensitive to defend. A speck of dust can ruin a batch. He does not see *Machines of Loving Grace* culminating without Beijing deciding the fabs go down. Leicht, who says he has 194 other countries to think about, offers three ways not to end up there. One: US systems become so far ahead that escalation around Taiwan is more suicidal than accepting some American domination — though it is unclear what shape of AGI would actually do that. Two: the fabs going down is less fatal than it looks if Arizona capacity plus already-built chips can carry an intelligence explosion already underway. Three: China is behind on indigenization too, and wrecking TSMC capacity that still leaks into China indirectly may not be its best catch-up play. Any serious AGI endgame, he says, has to account for Taiwan blowing up in our faces. A lot of them do not.

Tyler Cowen's challenge to doomers — if you are so sure, where are your shorts? — does not, in Leicht's view, have a good trading answer. Cowen imagines a smooth on-ramp of near misses and volatility you can bet. Many canonical doom scenarios concentrate the probability mass at the end: markets look great, strategy looks great, things look a bit weird, and then the takeover or the big incident happens. The only time you cash in is when you are dead. Leicht does not maintain a single p(doom) without defining the term. Human extinction is very, very low. If you include the old longtermist portfolio of locked-in political catastrophe — gradual disempowerment, stable authoritarianism, a permanent underclass with almost no agency — he is around 10 percent. He would not stuff ordinary "the economy kind of sucks" outcomes into the number.

Robotics, he assumes, eventually works: an engineering and scaling problem whose physical bottlenecks last longer than software ones. Domestic service robots starting to enter homes around 2030 sounds about right on technical maturity; psychological and political resistance may add delay. Compute going to space comes in two phases. From 2029 there will be inference setups worth putting in orbit, gated by launch capacity, concentrating more compute in US jurisdiction and making SpaceX and launch-site governance more important. If *every* marginal chip eventually goes to space, anti-satellite weapons become central to deterrence — the only way to threaten a superintelligence you cannot bomb on the ground — and Kessler syndrome complicates the math the way nuclear winter complicates first strikes. The actionable point for middle powers is cruder: compute-for-access has a time limit. If the US no longer needs terrestrial sites, those deals stop working, and you need an endgame beyond hosting GPUs.

On white-collar work, Leicht does not think the missing piece is another capability jump. Fable 5.1 or Astra already beat humans at many specific tasks; you still cannot fire the person at the desk and plug the model in. The lag is proprietary data, proprietary workflows, and reorganization: three people become one person who directs agents and is better at what the agents cannot do. He expects disruption of how teams are built more than short-run mass displacement. A cleaner shock — Tesla licensing full self-driving, four million American driving jobs gone in 18 months — would hit political friction first, human-in-the-loop laws, wage insurance, and a lot of worse jobs. He is wary of importing "surveillance with American characteristics": US criminal law was not written for perfect enforcement, and catching one in a hundred is load-bearing. Asked for the needles that most need threading, he declines a lab-scale narrative. History has been on a positive trend without needing AGI to bail anyone out. Keep the balance of power and wealth from concentrating in labs or in a single government, keep other countries economically and politically in the game, and when too much power amasses in one place, pull it back. "Just muddle through."