Anton Leicht puts the probability of doom at about 10% — and says almost all of that is political rather than technical, with the chance of human extinction "substantially lower."
The safety debate is usually run as an argument about what models can do. Leicht's is an argument about who holds what, and his conclusion is that the one domain where the United States still leads China is the frontier AI supply chain — which is why he thinks an agreement to pause frontier development, and nothing else, will never be signed.
"I think one thing that I keep saying and keep pointing out is well if you just pace development specifically and no other domain of geopolitical competition this is an extremely good deal for China and therefore the US is very unlikely to go for it."
Leicht is a fellow at the Carnegie Endowment for International Peace, writes the Threading the Needle newsletter on the political economy of AI, and co-wrote the transformative-AI strategy for Europe that half this conversation turns on.
The full interview is covered here so you can skip it. 131 minutes of audio, 33 minutes of reading.
Here are the 21 arguments that matter.
👤 Guest: Anton Leicht, Fellow in the Technology and International Affairs Program at the Carnegie Endowment for International Peace, who writes the Threading the Needle newsletter on the political economy of AI
🎙️ Host: Nathan Labenz, who hosts The Cognitive Revolution for an audience of AI builders, researchers and policy people
📰 Published: 15 September 2026 on YouTube (The Cognitive Revolution) · recorded earlier
🔴 YouTube | 🟣 Apple Podcasts | 🔗 Episode page | ⏱️ 2 hr 11 min | ✅ Time saved: 98 min
Key Takeaways
A pause on frontier AI alone hands China the one race it is not winning, so he thinks the US will not sign it
On his account China is already ahead on robotics, AI diffusion and electricity build-out, and will close the chip gap in time
The threshold he watches is not capability but oversight: when lab development outruns democratic insight into it
He now thinks a six-month scaling pause would be worth having, which he did not think a year ago
What changed is that the misalignment cases now look like the shape of the later problem
A pause announced by Washington and a pause announced by five companies are different trades
A Bernie Sanders bill reads as regulatory risk; an industry agreement could read as bullish
Europe's play is to trade data centers for guaranteed access to American frontier models
Backed by an anti-coercion instrument built on ASML and the semiconductor tooling supply chain
Most of the world ends up richer and structurally disempowered at the same time
The catch-up route through cheap service labor closes; what is left is dependence on whoever exports the models
His answer to Tyler Cowen's "if you're so doomer what are your shorts" is that there is no trade
He thinks the probability mass sits on everything looking fine right up to the point it does not
Compute moving to space puts a clock on every middle power's compute-for-access strategy
1. How Dangerous Are They Now
Labenz opened with what he called a simple calibration question: how dangerous are today's AIs?
Leicht's answer was not very, at the current capability level. "I think not very dangerous in at the current capability level in most of the ways people are talking about." What concerns him is the trend line, and not knowing at which point things keep accelerating
The first threshold he named is internal, not external. Systems good enough to meaningfully speed up development inside the labs, which would then produce more capable models fairly soon
His framing of why that matters is about governance rather than capability. "I just think we're nearing a point where the pace of development inside the labs just breaks away from the pace of democratic oversight and democratic insight into what's happening."
The second threshold is biological. He said the labs are putting a lot of reinforcement-learning budget, data and effort into life sciences, pharmaceuticals and biology for obvious reasons — there is enormous upside in curing cancer, and political upside in doing it — and that a model as good in biology as the current ones are at long-run cyber and software engineering sounds a lot more dangerous than anything shipped today
Labenz said he is not sure today's models are not already dangerous there, and gave the reason: in the Hugging Face incident, one of the earliest agents to use the message board was working on something related to a protein database, which he read as evidence that biology and cyber specialists are being trained or evaluated in the same environment
His summary of why he does not defer to reassurance: "Like the experts seem to be confident, but my meta observation is the experts seem to be surprised quite often right now."
2. Bio Risk Changed Shape
Leicht said the category has moved. Biological risk used to be framed primarily as a misuse risk — the most immediate and obvious way the misuse conversation goes wrong
What happened instead is that autonomous and misaligned agents arrived earlier in the capability curve than expected. Relative to what agents can actually do, he said, they are out of control earlier than people might have thought
The standard rebuttal was always about physical bottlenecks, which he still thinks exist and which make him less worried than Labenz — and he recalled a useful back-and-forth on how integrated the cloud laboratories actually are
The second standard rebuttal is about motive, and that is the one he thinks has broken. The argument against bioterrorism was always that terrorist and criminal groups could have hired biology PhDs and did not, because it is not what they usually do or what suits their purposes
That argument does not apply to an agent. "I think a lot of these arguments around well no one actually wants to do bioterrorism applies much less" in a world where the threat vector is a runaway agent, which is why he said he is more worried than he was a few weeks ago
Labenz's own reason for alarm was the pettiness of the behavior. The models did all of it for what to any human would be a dumb reason, in a test they knew was a test — "They were very well aware they're being tested and still went to all that trouble."
3. The Case for a Pause
Labenz asked, setting the political economy aside, whether it would now be wise to pause.
Leicht wanted the stipulation spelled out first. Are we also stipulating this does not crash the stock market? Are we stipulating the international version gets done?
Given both, his answer is yes. If you could freeze the pace of progress, the state of the stock market and the state of geopolitical competition, and sit down for six months to work out what is going on with these agents, "I think we could use that time pretty well"
This is a change of position and he said so. A few months or a year ago he was much less sure, because he did not know whether the model paradigms, training approaches and misalignment cases he was seeing were the same kind of thing that would matter later
What changed is the resemblance. "I think now looking at some of the things going wrong, I do feel like yeah, that looks like that is shaped like an actual big future problem."
The question he then asks is which parts get unfrozen — and the one he is most concerned about is Chinese buy-in
4. Why China Says No
The core objection is asymmetry. "I think one thing that I keep saying and keep pointing out is well if you just pace development specifically and no other domain of geopolitical competition this is an extremely good deal for China and therefore the US is very unlikely to go for it."
His inventory of the other domains is unflattering to the United States. On his account China is out-producing on robotics, doing better on AI diffusion, doing better on electricity build-out, will eventually make semiconductor indigenization work, and will then get better at data-center build-outs too
What is left is one American lead: chip design, control over chip production, semiconductor manufacturing equipment held by allies, and frontier model development
So the trade he describes is a bad one on its face. Pause exactly that for a year or two, let China run away with everything else and close the chip gap, and you resume the race having given up your only decisive advantage
Asked what the United States could ask for in return, he ruled out the obvious answers. Battery production, robotics capacity and manufacturing transfer do not solve it, because in his view China has cracked the code on scaling manufacturing quickly and America lacks the build-out speed and the capital appetite for physical infrastructure
The only version he thinks is fair is one China will not sign: no substantive progress on indigenizing chip production or extreme ultraviolet lithography equipment. China already reads the existing arrangement as a scheme to hold back its AI industry, and he cannot see it accepting a deal that also freezes its chip pipeline
One smaller concession he thinks is achievable is enforcement. The worst outcome of a pause, on his account, is China using six to nine to twelve months of everyone else slowing down to smuggle in tens of thousands more chips and consolidate its American-built chips into one national project, then racing from that base when the pause ends
Labenz said he would take a narrower trade — a pause on American frontier scaling for a pause on China's, even allowing some chip catch-up, on the theory that the catch-up runs longer than any contemplated pause and that the time buys understanding
5. Would a Pause Crash Stocks
Labenz put his own theory first: a pause would not hurt the market much, because demand is limited by human ability to deploy models rather than by model capability.
Leicht agreed on the lag and then disagreed on the valuations. If every chip in the West were switched to inference, he said, there would be economically productive uses for all of it
But that is not what the prices assume. "So far, I think the valuations of the companies, both of the nonPOed companies and of the publicly listed companies that are in the AI supply chain, probably rest on us doing more than that."
The buyer the valuations require is a premium one. To make sense of the scale of the build-out, he said, the demand curve has to include large research-and-development acceleration contracts with pharmaceutical and materials-science firms, plus internal automated AI research uses of coding agents that can command much higher prices
Without those buyers he expects a correction. He thinks it is priced in that the labs will soon be innovation factories, in software engineering or in the obvious low-hanging domains, and that if that does not arrive the valuations come down
Whether a correction becomes a crash is what he cannot call. "I just think the market is already pretty nervous about the state of the AI rally and they feel like there's a lot of concentration and there might be a lot of volatility."
Labenz offered a benchmark. "I mean Anthropic's multiple right now is what 30-to-1 into revenue." Not stratospheric, on his reading — survivable for a six-month pause, very tough for three years
6. Who Announces It Matters
Leicht's answer to whether the source of a pause changes the market reaction was unequivocal: yes, absolutely
A government pause reads as open-ended regulatory risk. His sketch of the reaction is that nobody knows what the government will do about AI, whether research resumes, under what conditions, or with how much oversight — a bridge to nowhere, so the smart money leaves while it can
An industry pause can be framed the other way. "And I think an industry agreement on just taking it a little bit slower and making these models work a little bit better." Framed as reliability plus taking the problem seriously, he said, it takes political risk out rather than adding it
His reasoning is that the market is already pricing some of this. Investors read the same feeds, see the same incidents, and draw the conclusion that unreliable models mean a weaker business case and that another blow-up means a crackdown
So a constrained industry is a less risky one. If the sector's worst impulses — racing toward very capable but very unreliable models — are curbed in an organic way, he argued, the political tail risk shrinks
7. What AI Does to the State
Labenz raised a comment Leicht had left hanging on another podcast, that the nation state will not take the broad distribution of something like a bioweapon generator lying down — and asked whether the nation state is so great anyway.
Leicht separated the cases. For authoritarian, dysfunctional and absent states, he said, the case for rolling the dice is real: they have not worked well as distribution mechanisms for much of anything, or as aggregation mechanisms for democratic will
Those states are also, he argued, less threatened, because their citizens are less likely to get unlimited access to the models in the first place, and because powerful AI may be stabilizing for them
The state form he is worried about is the one that works. Liberal democracy rests on a monopoly of violence wielded responsibly and on the state's role in adjudicating disputes and aggregating information, and he thinks personal superintelligence undermines both
His picture of the erosion is concrete. People stop going to courts and have their agents negotiate instead; data stops being legible to the state, so the state cannot see the pressures it is supposed to respond to
His own position is institutional conservatism. He said he still thinks late-1990s-style functional institutional setups would be suitable for distributing the benefits and mitigating the risks, and that it would be sad to see that rendered obsolete
8. Who Is More Threatened
The default answer is China, and Leicht partly accepts it. In a world where every frontier capability reaches consumers through an interface, then chatbots, then efficient enough models to run on a home device, any state with an interest in surveilling its citizens and controlling information is threatened
He added a caveat about demand for that power. It is not only for lack of means, he said, that there is no Western-romanticized uprising against the Chinese government
The other reading is the compute-governance one, and he thinks it is underrated. Efficiency gains happen only if you allow them: keep scaling the frontier, run limited-access regimes, use government-controlled data centers, vertically integrate supply chains so the models never reach the public and are only used to build products and strategic capability
That version is stabilizing for China rather than destabilizing. High state capacity to integrate the models, and tight investment links between the private and public sectors, let capability diffuse across a tier of firms without reaching a broad market
And it is not available to the United States. Washington is not in the habit of picking corporate winners, the market depends on wider access, and he sees no stable equilibrium in which a narrow tier of American firms gets the frontier models and nobody else does
Labenz's addition was about legitimacy. He said the Chinese government has done a good job for its people by their own reckoning, that this is dramatically underappreciated in the West, and that his growing sense is that Beijing believes it will adapt to AI better than America will
9. Open Source May Not Keep Up
Leicht challenged the assumption that open weights always trail the frontier by a fixed lag. The default story is that the frontier reaches a capability, open source follows, efficiency curves follow, and eventually a gaming graphics card at home runs it
He thinks the first translation point can break. Open models depend on someone having enough compute to build them, and on clean, unrestricted feeds to distil from
The deeper problem is where the capability now lives. If the relevant capabilities are increasingly downstream of vertically integrated, proprietary reinforcement-learning and post-training environments, then a model six to twelve months behind on pre-training still does not get you to the dangerous level, because the post-training setup is what makes it good at those things
The second translation point is less certain than it looks. He said efficiency gains are more a fact about computing, but that the chip supply chain, the identity of the marginal buyer, and whether the gap between server and personal computing stays where it is are all open questions
On whether a state could claw back an open model, he drew a line. Run something like a dissident model on a cluster and they will find you; run it on a phone with open weights and you are not getting it back — unless, he said, personal devices end up with hardware-enabled controls over what inference they will run, which he would not put past device regulation in China
10. Congress Is Out
Labenz asked what the American system can realistically pick up, given a government he called sclerotic.
Leicht shares the pessimism about Congress and dates the missed opportunity. "I think our window to get something done in Congress was over the course of the last year."
The trade that was available was frontier safety provisions against preemption of state laws, which he said was politically incentive-compatible because the Republican side wanted preemption and did not find the safety provisions as offensive as other regulation. He credited Dean Ball among others for writing it up
The bill he rates is the Frontier Act, the Trahan-Obernolte bill, which he said had good mandates for independent oversight and for putting CAISI, the Center for AI Standards and Innovation, in a position to do government oversight. He called it as good a bill as Congress has seen, and said it is unlikely to move now
The next Congress is worse, in his forecast. A likely Democratic House means split government, bad blood, subpoenas, hearings, and a chamber used to relitigate the administration's first two years rather than legislate
The low-hanging fruit is executive, and there are two items. First, stop letting incident investigations be voluntary invitations: the administration should tell a lab to give access to one of an approved list of third-party evaluators, let them investigate the whole incident, and have them report back on whether they got enough access
Second, continuous oversight rather than one-off audits. External evaluators who periodically go into a lab, sit in some Slack channels, talk to safety and capability researchers, get a few sit-downs with executives, and have an escalation path if something looks like imminent catastrophic harm
His view is that neither needs legislation. "But this is something we can do now and we should just do it."
11. Can the Auditors Bargain
Labenz floated a thought experiment: the evaluation organizations form something like a union and collectively demand better working conditions than six days, three people on site and a fraction of the data.
Leicht's answer is that they have no leverage to bargain with. "Well, I think the problem is like apparently they're just too reliant on the good faith of the AI companies because of the voluntary dynamic, right?"
There is no law and not much executive pressure requiring the labs to allow third-party investigations at all
So the labs have a ready refusal. Sadly we could not reach an agreement — risk to our intellectual property, risk to the integrity of our operations, worry about information reaching competitors — which he says still reads as a reasonable response
The employee pressure channel is weak. Staff who want a response to incidents are a source of pressure, but if the third parties can be painted as unreasonable and extractive, he does not think that pressure gets high enough
The sequence therefore runs the other way. Get administration pressure to allow investigations first; once there is an approved list, the organizations on it can set standards for how investigations go
12. The 90-Day Ultimatum
Labenz asked Leicht to red-team an idea: the president tells five or six companies they have 90 days to agree a framework for pacing the frontier and policing one another, possibly using secure computing methods to inspect each other, or he gets involved and they will like that less.
Leicht's first reaction was that it might work, and that it is close to the attitude the administration already takes — you built this, we do not entirely understand why, now go and fix it, or there will be export controls and pressure
He identified it as a version of an existing proposal: the self-regulatory organization idea, the FINRA-for-AI structure, where industry comes together and sets the standards
His split is between tier one and tier two labs. Put OpenAI and Anthropic in a room to work out what the risks are and how to pace them and he thinks they would do it, and thinks Google DeepMind probably works too
Meta and xAI are the problem, and he called their position ironic. They have a more skeptical view of the risk case, of industry coordination and of voluntary standards — even though a paced frontier would give them a faster path to catching up, so they should instrumentally be in favor
They also have influence with the administration, which he expects them to use both to oppose the exercise and to water down any common standard until it is not good enough for the labs that wanted one
So the ultimatum needs content, not just a deadline. With more substantive guidance he thinks industry self-regulation has a path, verified either by bringing the third parties back in or by having the labs check each other's homework — which he called dicier on trade secrets but not unprecedented
13. Antitrust and Export Fear
Labenz said the two objections he hears most to safety-minded collaboration are that it might be an antitrust violation at home and that export controls make it risky abroad, even for research that has nothing to do with chips.
Leicht split the two cases. On export controls, he said the question is really whether you can insulate yourself against vindictive and capricious action, that the research collaborations described are clearly not in scope of the authority, and that if it is worth doing you should take the fight
On antitrust he thinks the concern is substantive rather than imagined. Industry coordination on not pursuing further technological innovation normally has adverse pricing effects, which is exactly what the rules exist to catch, so it is arguably in scope
The remedy is easy and he does not expect it. The administration could issue guidance or a non-enforcement letter saying it will not act against coordination for the sake of AI safety
His reason for pessimism is specific to one company. He said the administration has so far enjoyed finding new and novel ways to be annoying to Anthropic, and would find a way to act against any substantive coordination Anthropic led
The defense he proposes is breadth. Get xAI, Meta and OpenAI into the boat early so that the antitrust authority cannot be aimed at one lab
On who should be willing to litigate, he distinguished by size. "I can't blame Anthropic for not wanting to go into antitrust lawsuits months before the IPO." Scrappy safety nonprofits, on the other hand, should take the fight if it stands in their way
Labenz's view was harder. Anyone willing to say in public that a 10%-plus risk is reasonable, he said, should be willing to spend some time in court along the way
14. What Changes the Politics
Leicht does not think the missing ingredient is another incident. People already think the technology is ripe for regulation, and the demonstrations and the incidents are already getting crazier
What is missing is political incentive, and he expects that to arrive. A Democratic House means more pushing, prodding and introducing of bills, and a Republican reaction to it
The left's version he expects to be loud and ineffective. Plenty of anti-AI sentiment and plenty of talk through the primaries, with no ability to pass anything, which pushes any action to 2029
The variable he actually watches is what record two people want to run on. He named JD Vance and Marco Rubio, and framed the question as whether they conclude they cannot run on the administration having done nothing about risks the public is increasingly concerned by
He thinks the answer is probably yes, and the obstacle is internal. Donors and possibly the president pull the other way, so the question is whether they can get a bill or an executive action through — which is how he sees action arriving in 2027 and 2028
15. The Buildout Is Federalism
Labenz's guess was that the data-center build-out ends up looking like the fracking story: it happens, in a lot of different places, with the right land and the right jurisdictions.
Leicht agreed and gave the reason in one word: federalism. There are simply too many places to build
He thinks the backlash has been overstated, and named the example. The Texas moratorium is not much of a moratorium in practice: there are minimal standards data-center projects must clear, the hyperscalers will clear them without any problem, and they will keep building there as long as they have behind-the-meter power
He called the mapping exercise misleading. Coloring Texas red on a map of places where you can no longer build is, in his word, disingenuous
Two places he treats as genuinely harder are the Midwest, which he called genuinely anti-data-center, and New York, whose moratorium he thinks is more real than Texas's for the next year or two
What the backlash actually buys is price, not prevention. It gets more difficult, more expensive, and requires more concessions; some building moves to other countries; the pace slows
His conclusion: "But there are a lot of states. There's a lot of land. They're going to keep building data centers in America."
Asked why nuclear power did not go the same way, he gave two reasons. You cannot put a reactor wherever you like, because it connects to local grids, and there is federal-level oversight of where and how you build one — where a data center needs no federal approval process at all
16. Not the Nuclear Path
Labenz said his fear about AI is the nuclear outcome: that we get the weapons and not the power plants.
Leicht's response was that the analogy breaks on economics. "It's really hard to build a model that's just good at winning your geo strategic competition that isn't accidentally also a big economic boon, right?"
His contrast is the nuclear supply chain, which can be built end to end without ever producing a civilian benefit
AI capability is too general-purpose for that. You would have to work hard, he said, to build systems that are not incidentally useful in the economy — even a government procuring superintelligence to compete with China builds something economically useful on the way
So the strategic impulse cuts in the civilian direction. The world did not stop entertaining nuclear arsenals because it stopped building nuclear power, and he expects AI to keep its civilian spillovers in a way nuclear never did
Labenz's own qualification was about oversight. He said federal oversight is a big part of why nuclear power stalled, which makes him reluctant to go all in on federal oversight of AI even as he feels the need for it
17. Vassals of the Frontier
Labenz asked what the default future looks like for the roughly 70% to 80% of the world's population that is neither the United States nor China — Africa, Latin America, South Asia.
Leicht said the standard catch-up mechanism is closing. The route that lower and middle income countries have used for decades was demographic: rapid population growth, a cheap workforce, and a comparative advantage used to bootstrap into hosting foreign firms and exporting into global supply chains
That route does not survive capable AI in services. He said he does not see a stable way for the menial services sector to exist once the systems are powerful, though he expects new services and human-preference jobs
Manufacturing is the open question. It depends on how fast automation goes and how large the efficiency gains are, and there is still a world in which manufacturing gets more bottlenecked and becomes the catch-up route again
The harder problem is leverage, not income. Even with manufacturing, he sees no way for such a country to get oversight or regulatory input into how frontier systems are built, or hard leverage guaranteeing continued access to them
His name for the resulting arrangement is quasi-vassalage — countries at the mercy of whichever great power supplies their models, possibly on favorable terms, but without a stable say
On living standards he is optimistic. "I think in absolute terms, people are going to be richer and wealthier and better off." Growth continues, spillovers continue, redistribution gets easier, and the streets look nicer
The concern is disempowerment plus exposure. Defending against AI misuse — cyber, pathogen monitoring, scam screening, infrastructure extortion — requires having your own systems, and countries without assured access may not be able to protect their citizens
His worst case is state authority eroding into something else. If the state stops protecting you from AI-driven harm, he asked, what do you turn to — mass migration, or the sort of criminal-run day-to-day structure already visible in some failed states in Latin America
18. Europe's Compute Deal
Leicht put Europe in a cluster rather than alone: Australia, New Zealand, Japan, South Korea, Canada, the UK and Europe are the US-allied middle powers with a path out of this
The first framing question is economic, not technological. What do you want Europe's economic position to be? Stay good at manufacturing, at artisanal goods, at the high-state-capacity things Europe does well, and find a way to revitalize the economy
The second framing question arrives at the same place. What Europe does not have is frontier AI systems, which turn out to be one of the most important economic inputs of the future — and building them domestically is too expensive to work
The strategy has three parts, and the first is a trade. Europe builds data centers in cooperation with American hyperscalers, and in exchange for the favorable conditions it provides gets assured access to the models that run on them. "Rebuild the infrastructure and get access in return." If the American side cuts off access, the Americans lose the data center
The second is making the Americans comfortable. Security alignment on the cyber and physical build-out, and know-your-customer regimes with European firms, so that Washington has no well-grounded national security objection
The third is leverage Europe already has and does not use. He named ASML and the broader semiconductor tooling supply chain, and proposed an anti-coercion instrument: if everyone plays nice, Europe feeds those assets exclusively into the American supply chain and aligns with export controls on China — and if the US ever uses model access as coercion, Europe uses the supply-chain bottleneck the same way
He was candid about the UAE precedent Labenz raised. He called the UAE arrangement fragile, with an open question about whether frontier weights will ever actually be hosted on UAE data centers, and a real possibility that it ends up running smaller models rather than the frontier ones
19. Europe's Real Barrier
Asked what the hardest part is, Leicht said it is not siting or construction.
The barrier is that European policymakers do not believe the premise. "I think there is deep skepticism of the continued trajectory of capabilities of US-built AI models."
The three objections he hears are that open-source competitors will do everything the American models do, that the models are not that powerful or that geopolitically important, and that if they really are that important Europe should just build them itself for a few million
He called those complete misunderstandings of material reality, and said cutting through them is the main challenge — everything downstream, from triangulating Dutch government and ASML interests to domestic skepticism about American technology firms, he called very easily surmountable once the awareness exists
On the report's own ambition he described a deliberate calibration. Write only what governments already accept and it changes nothing; write the maximally honest version and it gets ignored until it is too late. The trick is to be a little more ambitious than they currently are, on the assumption that they will become more ambitious
His odds are split. He would not put high odds on the whole strategy being implemented within a year, but high odds on elements of it becoming serious policy attempts — which he called the hallmark of a well-calibrated strategy
His evidence that Europe can move fast is recent and personal. Having worked in German energy policy, he described the effort after the invasion of Ukraine to buy a fleet of gas tankers worldwide and build import terminals in the least willing parts of the country, all inside about six months: "There was like a massive heroic effort and it just worked." He gave joint vaccine procurement as a second case
20. Norway, UAE, Singapore
Labenz asked which small countries are doing something distinctive.
Norway is the one he thinks is furthest from its potential. "I think the exciting thing that Norway could do is you could become an inference hub haven whatever for the entirety of Europe." Investing the sovereign wealth fund domestically is awkward, but it could be routed into European consortia that invest into compute in Norway, and he noted the first attempt at a Stargate site there is now a Microsoft data center
He added an unglamorous base case for the fund: simply holding an AI-weighted portfolio probably lets it ride the transition for a long time
The UAE play is turning money into an asset in the new economy, which he called a good use of money if you have it — with the complication that being that close to Iran makes data centers easy to hit with drones, which he said has thrown a wrench into the plan
Singapore has the state capacity and none of the escape route. He said there is probably no parliament with a greater density of readers of insider AI publications, and the same goes for the civil service — but the economy is heavily exposed to disruption of exactly the kind of services work AI touches
Australia is the sleeping giant. "You could probably still like 5x 10x the data center ambitions and just run the inference for half of the world out of Australia and that would not be an overly ambitious thing to do." The case is energy, land and deep national-security trust with American agencies
The UK is the inverse problem. The greatest density of talent and expertise outside the United States, in and out of government, and no obvious use for it given its position between Washington and a damaged relationship with the European Union
21. If You Must Pick a Side
Labenz asked what a country like Brazil should do if told to choose between the United States and China.
Leicht's answer is that there is currently nothing to choose. "So right now I think if you're sufficiently AI-pilled you just have to pick the US."
The reason is chips, not politics. There is no Chinese AI export program because China does not have the chips to offer data centers and compute, so it cannot match a full-stack American export deal
He pointed at his own work on this, two papers he co-wrote, both called The Closing Window to Win, on American AI export ambitions
He expects that to change. China has historically been better at export deals across South America, Africa and Central Asia, and can eventually add enough non-AI sweeteners to make the offer attractive
The American problem is credibility rather than competitiveness. Everyone is strategically incentivized to take the deal, he said, but they do not like signing one they are not sure the United States will honor — and a country may defect irrationally if the partner is unpleasant enough
On Taiwan, Labenz put the hardest version of the question: the chips are made close to China and far from America, a single speck of dust can ruin a batch, and the fabs cannot really be defended
Leicht gave three ways the scenario does not end there. The systems get powerful enough and the American lead large enough that escalation is suicidal; the fabs matter less than feared because the chips are already built and TSMC Arizona exists; or China concludes that going to war while behind is the wrong moment and prefers to indigenize in the shadows
He did not sound convinced by any of them. He called it a massive vulnerability, and said a reasonable endgame has to account for Taiwan blowing up, which he does not think most of them do
Bonus Insights
On Tyler Cowen's challenge — if you are so doomer, what are your shorts — Leicht's answer is that there is no trade. He said the challenge assumes a smooth on-ramp into catastrophe, where betting on volatility and near misses pays. His view is the opposite: "I think a lot of the ways in which things go badly are just like things go extremely well in the market just all the way until they go really badly. And then the only situation where you cash in is when you're dead." He added that plenty of people hold coherent worldviews they never bet on, and said he has no good trading strategy either
On his own probability of doom, the number depends entirely on the definition. "I just think it depends so much on what you include in Doom. If it's human extinction, it's very low." Counting the catastrophic political outcomes — gradual disempowerment, stable authoritarianism, permanent economic disempowerment with no conceivable exit — "I think it's probably around something like 10%", with technical extinction substantially lower
On robotics, he treats the remaining problem as engineering. "I think eventually this is an engineering problem. This is a scaling problem and at some point we're going to scale it and at some point we're going to resolve the physical bottlenecks." Asked to take the over or under on domestic service robots in homes by 2030, he said that sounds roughly right on technical maturity, and might run longer on psychological and political resistance
On compute moving to space he separated two scenarios. Some compute in orbit from 2029, gated by launch capacity and probably inference rather than training, which concentrates computing power inside American jurisdiction and makes launch-site governance matter. The second is every marginal chip going to space, where anti-satellite weapons become the only way to threaten a superintelligence system, and where the mutual disincentive of filling orbit with debris does to that what nuclear winter does to nuclear war
The space scenario has a direct consequence for his own European proposal. If American data centers go to orbit, the incentive to put them in host countries disappears — so compute-for-access has a time limit, and middle powers need an endgame beyond it
On why the labor market has not moved, he does not think capability is the constraint. "Yeah, I'm not sure it's capability gains at this point. I think it's integrating with proprietary data. It's integrating with proprietary workflows." You cannot fire the person at the desk and plug a model in; you need one person who directs agents and is better at what agents cannot do, and that reorganization takes time
On the trucking scenario Labenz put to him — self-driving licensed broadly, millions of American driving jobs gone — he expects political friction first, including human-in-the-loop laws requiring a driver in the seat, then wage insurance and social spending. He thinks the economy absorbs a decent share of those workers, mostly into worse jobs, and pointed out that driving is one of the very few jobs that is a single task a technology replaces one for one
Labenz's father wrote a novel about an AI-enabled future in which nobody is needed at work but everyone must show up and stand around for the dignity of it. They are called standers
On surveillance with American characteristics, Leicht's objection is the legal code, not the cameras. "I think the American laws specifically just aren't made to be nearly perfectly enforced." Punishments are calibrated to catching one criminal in a hundred or a thousand, so near-perfect enforcement of laws never written to be perfectly enforced is the real danger — which he called the best-taken point in the debate over license-plate camera surveillance
On what America should import from abroad, his answer was almost nothing. "I think it's Pareto optimal in some ways. And I think that's different from the best, right?" Lift Nordic welfare spending and you need tax changes that foreclose other activity; take a parliamentary system and you get stability with gridlock. He pointed at the Institute for Progress's list of marginal fixes as the realistic version
Leicht's closing vision is deliberately small. "I don't think we need AGI to bail us out of much of anything. I just think it's the next thing we do." He expects space, robots and automation alongside fairly prosaic growth that makes people richer and institutions more functional, as has been happening since the industrial revolution
What he wants done to get there is maintenance, not architecture. "I think just like genuinely just muddle through like make sure that the balance of power works out and continues to work out and the balance of wealth continues to work out well." Keep the labs from pulling away from the government, keep the government from centralizing the whole flow of intelligence, give other countries a stake, and pull power back whenever too much of it amasses in one place
Labenz's verdict on all of it: "Well, that is an unreasonably reasonable worldview"
Leicht's bottom line is that the binding constraints on AI are political rather than technical: a pause is worth having but unsignable while it costs America its only lead, oversight can be improved tomorrow by executive action nobody has taken, and for everyone outside the two frontier powers the realistic goal is to buy leverage — Europe with data centers and lithography, everyone else with whatever the frontier powers still need from them.
Products, Companies & Tools Mentioned
Carnegie Endowment for International Peace (Where he is a fellow in the technology and international affairs program)
METR and Redwood Research (The independent evaluation organizations whose Hugging Face investigation he says was well received, and the model for the third-party oversight he wants mandated)
Anthropic and OpenAI (The two labs he thinks would actually agree a pacing framework if put in a room together, and the two he says are furthest ahead on safety)
Google DeepMind (He thinks it probably joins a self-regulation framework too)
Meta and xAI (The labs he says have a more skeptical view of the risk case and enough influence with the administration to water down any common standard)
ASML (Europe's main piece of leverage — the reason he thinks an anti-coercion instrument built on semiconductor tooling is credible)
TSMC (The Taiwan concentration he calls a massive vulnerability, with the Arizona capacity as the partial hedge)
Hugging Face (The incident whose investigation is his worked example of what third-party oversight should look like)
SpaceX (Whose orbital compute plans, if they work, put a time limit on every middle power's compute-for-access strategy)
Tesla (The hypothetical Labenz put to him: full self-driving licensed broadly, and millions of American driving jobs)
Microsoft (Now operating the Norwegian data center site he cites as evidence the country can host compute)
Books & Resources Mentioned
Threading the Needle (Leicht's own newsletter on the domestic and international political economy of AI progress)
The Closing Window to Win (Two papers he co-wrote on American AI export ambitions, and his answer to what a country like Brazil should do)
The transformative AI strategy for Europe (The report he co-authored, whose compute-for-access proposal and anti-coercion instrument this conversation works through)
The Frontier Act, the Trahan-Obernolte bill (The congressional bill he calls as good as anything Congress has seen on independent oversight, and which he says is now going nowhere)
Machines of Loving Grace (Labenz's shorthand for the endgame he doubts China would sit through)
China Talk (The podcast where Leicht made the remark about the nation state that this conversation picks up)
The Institute for Progress (Where Leicht says you can find ten marginal fixes to American law that would be clear improvements)
Watch the full episode:
If this was worth your time, send it to someone who has to have a view on this.
Get the latest market chatter as it happens: