The world is too loud. Read what matters.

The Cognitive Revolution 2026-09-15

A Pause on Frontier Development Is a Great Deal for China, So America Won't Take It

If a pause covers only frontier model development — not robotics, power, or semiconductor indigenization — China can finish running every other race and restart the competition at the moment it no longer lacks a chip advantage. That deal is plainly bad for the United States.

AI geopoliticsComputeExport controlsEuropean policyMiddle powers

The video won't play here. Listen to the audio instead:

Anton chains geopolitics, markets, European strategy and domestic politics into a single argument; the back half, on compute for access and the endgame for middle powers, is where the value is.

The argument · tap a timestamp to hear it

A pause on frontier development alone is a gift to China

Anton returns again and again to one judgment: if a pause covers only frontier model development and not the other domains of geopolitical competition, it is an ‘extremely good deal’ for China, which is why the United States is almost certain not to accept it. His reasoning: China is catching up or already ahead in robotics, AI diffusion, power buildout, semiconductor indigenization and data center construction. The only places the US leads by a wide margin are chip design, allied control over semiconductor equipment, and frontier model development. Pausing that one piece amounts to letting China finish running every other race, then restarting the competition at the moment the US has lost its decisive chip advantage.

— Anton Leicht

A fair version would require China to give up indigenization

In Anton's view, the only version of the deal that could count as ‘fair’ is one where the US demands that China slow not just its own frontier development but the entire supply chain tied to frontier AI — specifically chip production and EUV lithography equipment, i.e. ‘no substantive progress on indigenization.’ He concedes this ask is ‘really really hard to ask,’ because China already reads this round of talks as an American plan to contain its AI industry, and adding this condition makes it even less acceptable. So his conclusion: either the US makes a deal that is clearly bad for itself and good for China, or there is no deal.

— Anton Leicht

The host would trade a frontier pause for a chip catch-up

The host stakes out a position different from Anton's: he would trade a pause on US frontier scaling for a reciprocal pause on Chinese frontier scaling, and allow China to keep catching up on chips. His reason is the timescale — chip indigenization is a five-year-order problem, longer than any imaginable pause; by then we might have a better grip on the question and could negotiate again. Anton partly accepts this, arguing that even if the deal is slightly unfavorable to the US, from a risk-reduction standpoint it is still a net good for the world.

— Host

What a pause fears most is smuggling and stockpiling

Anton offers a concrete mechanism: if during the pause China keeps smuggling frontier chips and consolidating American chips into a single national-project data center, then when the pause ends China restarts from an integrated, high starting point — the pause helped it. So rather than demanding that China slow indigenization, he suggests making ‘actually enforcing export controls and shutting down smuggling’ the easier concession to extract. He admits the US has to handle its own side of this too.

— Anton Leicht

Valuations are betting on super-buyers, not accountants

The host proposes that demand is limited by human deployment capacity rather than model capability, so a pause need not be bad for the stock market. Anton partly agrees — if you switched all existing Western compute to inference, it would be enough to pay back the frontier model and chip investments already made. But he says current valuations ‘probably rest on us doing more than that’: the market is pricing in R&D acceleration contracts worth millions of dollars each in materials science and drug discovery, plus internal productivity gains and coding agents that can be sold at high prices. If those super-buyers don't show up, the valuations of labs and public companies — and the scale of the buildout — don't add up, and he expects the AI sector to drag downward at least.

— Anton Leicht

The market may run before you finish explaining

Anton thinks a pause is still read as a ‘radical policy proposal’ and an ‘unprecedented policy intervention,’ so conservative market analysts and smart money will think: the regulatory path is unclear, better to get out first. He distinguishes two cases — if the market believes everything will be business as usual in six months, fine; but if it reads this as a Bernie Sanders-style AI policy victory, the end of free development, and no idea when or on what terms it resumes, the market may already have run before you finish making yourself clear. He adds that if the five frontier companies announced a pause themselves and framed it as a reliability problem, it might instead be read as removing political risk, and could even support a bullish story.

— Anton Leicht

What AI most threatens is liberal democracies

Anton says many authoritarian, dysfunctional or absent states were never good at distributing and aggregating public opinion in the first place, and may instead use powerful AI to maintain stability — so they are not necessarily the most threatened. What he really worries about is liberal democracies: the state's monopoly on violence, dispute adjudication, and the aggregation of data and knowledge — these three functions get eroded by individual superintelligence. People stop going to court and have agents negotiate instead; data can no longer be scanned and read by the state; and the state therefore cannot respond to redistribution and social challenges. He says plainly he does not want to roll the dice on this bet, and would rather integrate capabilities into nation-states that still function.

— Anton Leicht

China can put AI in a vertically integrated cage

The other path is compute governance: efficiency gains happen only when they are permitted to happen, and that depends on concrete choices about allocating compute. China can put models into government-controlled data centers and vertically integrate the supply chain around them, so they ‘never see the light of day,’ used only to build products and strategic sovereignty, barely reaching citizens at all. Anton sees this as a stabilizer for China: strong state capacity, a thin line between public and private sectors, and capabilities that can diffuse along corporate hierarchies without diffusing to the broader market. By contrast the US market depends on models being broadly accessible, and the government is not in the habit of picking corporate winners, so the US has a harder time sustaining this kind of ‘limited access’ stable equilibrium.

— Anton Leicht

The open source is always X months behind pipeline may break

The default narrative: a capability appears at the frontier, open source catches up quickly, the efficiency curve then puts it in everyone's hands, and eventually it runs on a home GPU. Anton thinks neither transition point is guaranteed. Closed-to-open depends on someone continuously getting enough compute and a clean API; and if more future capability comes from highly proprietary, vertically integrated RL and post-training environments (like Anthropic's life sciences direction), then getting a pretrained model that is 6, 9 or 12 months behind won't get you to that level. The efficiency-curve end depends on the chip supply chain and who becomes the marginal buyer of compute — whether the gap between large servers and personal computing widens, he isn't sure.

— Anton Leicht

Congress is hopeless, but the executive can do two things tomorrow

Anton is pessimistic about legislation: a trade-and-AI omnibus like the Frontier Act didn't get through this Congress, Democrats will likely take the House after the midterms, and two chambers with grievances, subpoenas and hearings will make it even less productive. But he sees two low-hanging fruits at the executive level. First, institutionalize incident investigation — it shouldn't be up to OpenAI to voluntarily invite a third party; the government should publish a list of recognized third-party evaluators, let companies choose from it, and have evaluators write reports and feed back whether they got sufficient access. Second, put external evaluators in labs on a standing basis for continuous oversight — sitting in Slack channels, talking to safety and capabilities researchers, talking to executives, and able to escalate serious problems immediately.

— Anton Leicht

A third-party evaluator union is useless because it all rests on company goodwill

The host imagines third-party evaluation firms banding together to collectively bargain for full access. Anton thinks it won't work: there is currently no law, and not even much executive pressure, requiring labs to accept third-party investigation, and they can simply refuse on grounds of IP security, operational integrity, or fear of information leaking to competitors — and that still reads as a reasonable response today. The only possible pressure comes from internal employees, but if third parties are portrayed as unreasonable and overreaching, employee pressure won't be enough. There has to be an external incentive making labs willing to accept any investigation at all.

— Anton Leicht

Lock five companies in a room to set the pace; Meta and xAI are the variables

The host proposes a thought experiment: the president tells five or six companies they have 90 days to negotiate an agreement among themselves to coordinate control of the frontier pace, with mutual monitoring, and if they fail he will regulate them — and they will like that outcome much less. Anton thinks this is workable, and that it is a structured version of an SRO / industry self-regulation. But he notes Meta and xAI judge the risk differently, are more skeptical of industry coordination and voluntary standards, and have influence in government, so they would obstruct common standards; ironically, actually pacing the frontier gives them a faster catch-up path, so on instrumental rationality they should support it. The conclusion is that more substantive guidance is needed than ‘figure it out yourselves.’

— Anton Leicht

Antitrust and export controls are the two real gates to safety collaboration

The host says the two objections he hears most when pitching this kind of safety collaboration are: domestically it may violate antitrust, and internationally export controls are broadly worded and enforcement may be arbitrary. Anton separates the two. On export controls, he thinks this is clearly outside the authorized scope, and if it's worth doing you should fight that authority. But on antitrust, he thinks it genuinely could fall within substantive antitrust rules — industry coordinating not to compete on frontier development usually has adverse pricing effects in other areas. The administration could issue a no-action letter, but he doubts it will, because the administration enjoys finding new paths to target Anthropic. The only way out is to bring xAI, Meta and OpenAI on board early, so enforcement can't single out just one company.

— Anton Leicht

The real variable is political incentives, not another warning

Anton thinks change won't come from an external event or a shift in how people see the technology — everyone already thinks AI should be regulated and intervened on, and things are getting crazier. What matters is when politicians and policymakers act, and right now political incentives are insufficient. He pays more attention to political tipping points: a Democratic House will keep pushing and introducing bills, and how the GOP responds; and what achievements JD Vance and Marco Rubio want to run on — if they decide they can't run on a record of ‘the Trump administration did nothing about risk,’ they will push some kind of legislative or executive action in '27 or '28.

— Anton Leicht

AI's civilian and military uses can't be separated, and that is grounds for optimism

The host says he is genuinely afraid of AI now, having moved from ‘it could be terrifying’ to ‘it is terrifying now,’ but still won't accept the nuclear ending of ‘you get the weapons, not the power plants,’ which is why he hesitates on federal regulation. Anton's distinction: the entire nuclear weapons supply chain can produce zero civilian benefit, whereas AI systems are so general in the ways they are economically useful that it is very hard to build a model that is good only at winning geopolitical competition without incidentally becoming economic prosperity. Even if the US government only procured superintelligence to compete with China, it would incidentally build economically useful systems. So this time the strategic impulse is on the favorable side.

— Anton Leicht

Low-income countries' catch-up mechanism is incompatible with AI

The host asks: if Europe does nothing, what happens to Africa, Latin America, South Asia. Anton says it will be very hard, because the mechanism low- and middle-income countries have relied on for decades to catch up — betting on demographic differences, comparative advantage from cheap labor, attracting foreign investment, exporting into global supply chains — is deeply incompatible with advanced AI systems and automated manufacturing capacity. He points out in particular that this mechanism presupposes a labor force that can be put to work, and he isn't sure that premise still holds, at least not for most manual work.

— Anton Leicht

The middle powers' endgame is quasi-vassalage

Anton's medium-term picture: most countries will be in a state of ‘quasi vassalage’ toward the frontier AI-building states, and the world gets carved into spheres of influence. He concedes that in absolute terms people will be richer, live better, redistribution will be easier, and the street level will look more respectable — this really is the story of China's last few decades. But in relative terms they are substantially stripped of the ability to shape the world's trajectory. He calls this ‘a deeper kind of powerlessness,’ and says democratic expression, human autonomy and dignity are the dimensions actually worth worrying about, not the economic ledger.

— Anton Leicht

The economic ledger looks good; the state-capacity ledger looks bad

Anton splits the risk into two layers: the purely economic story is fairly positive, while the story of eroding state authority and capacity is far more worrying. The mechanism: defending against AI misuse (cyber, pathogen monitoring, scam filtering, ransomware-proofing infrastructure) itself requires a country to have AI systems, and middle powers neither have coordinated, reliably deployable access nor any answer to the capabilities terrorists and criminal groups get by stealing or fine-tuning open source models. Once the state can no longer protect citizens from AI-driven harm, people turn to mass emigration or private security structures — he cites the example of partly failed states in Latin America and thinks this is not an impossible medium-to-long-term outcome. His order of magnitude: roughly 70% of the world heads there, and Europe is the only place that might design a different ending for itself.

— Anton Leicht

Compute for access: data centers as collateral

This is the core mechanism of the strategy, which Anton says he wrote down as compute for access late last year and early this year and began pitching to governments. Europe builds data centers for US hyperscalers to use, in exchange for access to the models running on those data centers; as long as the US keeps supplying models, it keeps getting the data centers, and the moment it cuts off models, it loses the data centers. The second piece is making the US not nervous: alignment safety provisions, cyber and physical security, and KYC mechanisms with European firms. The third piece is Europe using semiconductor supply chain assets (ASML and others) as an anti-coercion tool — if everyone plays by the rules, those assets feed exclusively into the US supply chain and align with export controls on China; if the US uses cutting off frontier models as coercion, Europe presses the supply chain button in response.

— Anton Leicht

The biggest obstacle is that Europe doesn't believe the premise

Anton says the hard part isn't siting and building data centers, it's that in many rooms there is simply no shared understanding that ‘this is the realistic future we face.’ European policymakers deeply doubt that US model capability will continue along this trajectory, are more optimistic about open source competitors being widely available, and broadly doubt whether these models are really that strong or really a geopolitical issue. The more common rebuttal is: if it really matters that much, why not build it ourselves? It can't be that expensive; the Americans are wasting their own money, and we can surely build it for a few million dollars. Anton says these are a thorough misunderstanding of material reality, and puncturing that perception is the biggest obstacle; the remaining political problems (the triangle of the Dutch government, ASML and other member states, domestic suspicion of American tech companies, people who want to hedge between the US and China) he considers very easy to overcome.

— Anton Leicht

A strategy should be slightly more ambitious than the government currently wants

Anton describes two failure modes in writing strategy for a national government. One is writing only what fits the current budget (‘we only have 20 million’) — then it doesn't matter what you write; you could burn it or throw a party and the conversation wouldn't change. The other is honestly writing everything you think should be done, and getting back ‘we're sorry, we were wrong.’ His technique is to write slightly more ambitiously than the government currently wants, while factoring in that they will become more ambitious and need a push. His odds: the probability that the whole strategy is fully implemented within a year is not high, but the probability that several of its elements enter serious policy attempts and actually get built is quite high — which he says is the mark of a well-calibrated strategy.

— Anton Leicht

When Europe is really alarmed, it can get things done

Anton supports his optimism with two precedents. When he worked on German energy policy, after the Ukraine war broke out the German government made a ‘heroic effort,’ buying shadow LNG tankers around the world, building LNG terminals in the least suitable places, and integrating resources within six months — they got through the winter fine, and none of the doomsday scenarios happened. Earlier, when he worked on Covid policy, joint vaccine procurement faced enormous political pressure and member-state interests were extremely hard to align, yet Europe still procured fairly well and the rollout went reasonably (he says the non-pharmaceutical interventions part was messier). His inference: as long as Europe realizes this matters at the order of magnitude of Covid or Ukraine, the recommendations in the strategy can move.

— Anton Leicht

Australia is a sleeping compute giant

Anton names Australia: excellent conditions for building compute (land, construction, energy supply), plus deep national-security trust between Australian and US security agencies, with the US confident Australia won't tilt toward China and will cooperate on export controls toward China — which makes Australia an excellent place to run compute for access and build data centers. He says for a long time this was a sleeping giant, and now there is movement, but Australia should scale up its data center ambition by 5 to 10 times and run inference for half the world; that isn't excessive ambition. The counterexample is the UK: it has the densest concentration of government and civil-society talent outside the US, but isn't sure what to do with it, because it doubts alignment with the US and its relationships with the EU and other middle powers are damaged — Europe has assets but no awareness; the UK has awareness but no cards.

— Anton Leicht

China can't offer a full stack export, so the only choice is the US

Anton's judgment is blunt: the more you think AI matters, the less reason you have to choose China, because China currently has no AI export offering — they simply don't have the chips, can't supply data centers and chips, can't do a full-stack export. The paper he and colleagues wrote (a follow-up to the previous one) is titled closing window to win, about US AI export ambitions, and it predicts China will eventually get better at offering these kinds of export deals, just as it has with other international initiatives in South America, Africa and Central Asia — but it can't right now. So the real question isn't whom to choose, it's ‘how reluctant are you to buy American systems.’ The real strategic challenge is on the US side: every country is strategically incentivized to accept the deal, but they don't like an agreement the US can renege on at any time, and the US has to find a way to make its commitments credible — building data centers is one part, deep industrial integration is another.

— Anton Leicht

Home robots by 2030: he bets late, not early

Asked whether household service robots in 2030 is over or under, Anton says ‘that sounds about right,’ but it may be a bit later still, and the reason isn't technical maturity but ‘idiosyncratic, psychological and political resistance’ — psychological and political resistance. Technically he thinks robotics is an engineering problem, a scaling problem, and that physical bottlenecks will persist longer than software bottlenecks, but that it is ultimately ‘eminently resolvable.’

— Anton Leicht

Compute goes to space, and middle powers' compute for access expires

He splits compute in space into two versions. In the first, starting in 2029 space becomes one of the places compute is deployed, inference suits it better than training, launch capacity is the constraint, and a space-specific chip supply chain spins out; the result is compute more concentrated within US jurisdiction, a stronger SpaceX AI, and launch-site governance becoming more critical. In the second, marginal chips all go to space and ground data centers have no competitive case, at which point anti-satellite weapons become the core of deterrence. The most actionable inference: the compute for access strategy middle powers are pushing has a time limit, and they need to think through the compute endgame first.

— Anton Leicht

What's missing isn't capability, it's proprietary data and workflows

The host says it's now hard to point to exactly where Fable 5 or Astra is worse than hiring a person. Anton's answer: the problem is no longer capability gains, it's access to proprietary data and proprietary workflows. Models are already stronger than people at specific tasks, especially software engineering and some white-collar activities, but the labor market needs time to rearrange — you can't just swap the person in the seat for an agent; rather, the work of three people becomes one person directing agents, and that person still has to be stronger at the things agents can't do. This requires organizational and institutional restructuring, so in the short term he isn't sure it leads directly to displacement, but the disruptive impact is certain.

— Anton Leicht

If autonomous driving really lands, political friction grows first

Suppose Tesla licenses FSD and within 18 months every car drives itself — what happens to the four million Americans who make a living driving. Anton says the political reaction comes first: wage insurance, reduced hours, and regulations like ‘a human must be in the driver's seat’ even when the car drives itself. The economy may eventually absorb some of those people, but not necessarily into better jobs — quite possibly worse ones. He also notes driving is a rare exception: the whole job is a single task, so it can be replaced one-to-one by a single technology; most AI automation only eats part of a task profile, so the medium term is more likely augmentation and coexistence.

— Anton Leicht

The biggest problem with surveillance isn't privacy, it's laws being perfectly enforced

Asked whether there is an ‘American-style surveillance,’ Anton says his specific worry is this: the American legal system was never designed for perfect enforcement. If you enforced every law currently on the books, it would be a harsh regulatory regime. Many penalties and criminal statutes are calibrated for deterrence on the assumption that you catch one offender in a hundred or a thousand. So surveillance maxing is in effect a path to perfectly enforcing laws that were never meant to be perfectly enforced — unless you first rewrite the entire criminal code and enforcement practice.

— Anton Leicht

He doesn't want a grand plan, he wants to keep pulling power back

Asked what needle he most needs to thread, Anton says he doesn't want to write a grand narrative for AI labs; he just wants history to continue its long positive trend, and doesn't think AGI is needed to save anything. The concrete approach is to ‘muddle through’: make sure the balance of power continues to hold, that the balance of wealth continues to hold; that labs don't pull away from the US government in power and control; that the US government doesn't centrally control the world's intelligence; that other countries have a share of economic and power leverage. Whenever power concentrates too much and things look like they're going off the rails, pull it back a little and keep it on track.

— Anton Leicht

In their own words · checked verbatim

If you just call front 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.

Anton Leicht14:24

you just want there to be no substantive progress on indigenization. Of either chip production or semiconductor manufacturing equipment, extreme ultraviolet autoography production.

Anton Leicht16:46

The valuations of the companies, both of the non IPO companies and of the publicly listed companies that are in the AI supply chain. Probably rest on us doing more than that.

Anton Leicht25:00

The specific nation state concept that I'm most worried about in this context is also the nation state concept that kind of works best. Which is liberal democracy.

Anton Leicht31:59

You can vertically integrate supply chains around these AI models such that they never see the light of day.

Anton Leicht39:14

Currently, they're just too reliant on the good faith of the AI companies because of the voluntary dynamic, right?

Anton Leicht51:37

It's really hard to build a model that's just good at winning your geost competition that isn't accidentally also a big economic boom

Anton Leicht1:09:32

I think the pure economic story is pretty positive. Eroding the state, the authority and power of the state story, is I think a lot more concerning.

Anton Leicht1:17:51

I just currently don't think there is as long as China doesn't have the chips.

Anton Leicht1:37:53

I think the American law specific issues aren't made to be nearly perfectly enforced.If you enforced every law on the books in America.I think this would just be a draconian oversight regime.

Anton Leicht1:58:05

Make sure that.The labs don't pull away in terms of power and control from the US government.😊Make sure that US government doesn't centralize and control the entire, sort of like.Blow through of intelligence through the world.

Anton Leicht2:04:12

Figures

Anthropic price-to-sales ratio	30 to 1 (revenue multiple)	28:09
Average weekly time saved by Athena customers	15 hours	22:13
Faster account opening at KeyBank after using OutSystems	75%	22:13
Deadline the president gives frontier labs to reach a pacing agreement themselves	90 days	55:00
Lag of pretrained models Anton cites as an example	6, 9, 12 months	43:22
Share of the world Anton estimates heads down the ‘state capacity eroded’ path	about 70%	1:17:51
Multiple by which Australia should scale up its data center ambition	5 to 10 times	1:33:59
Time for Germany to integrate alternative gas supply after the Ukraine war	within six months	1:30:00
Over/under year for household service robots entering homes	2030	1:48:49
Year compute begins to be deployed to space	2029	1:50:00
Apprehension ratio for calibrating criminal-law deterrence	1 in every 100 or 1000 offenders	1:58:05

Glossary

compute for access: Middle powers build data centers for US hyperscalers to use, in exchange for access to the models running on them.
quasi vassalage: Anton's term for the subordinate status of most countries toward the frontier AI-building states in the AI era.
SRO: A self-regulatory organization that sets and enforces standards for its own industry; Anton sees it as the structured version of five companies setting the pace themselves.
surveillance maxing: Pushing surveillance capability to its extreme; Anton worries this leads to laws being perfectly enforced.
muddle through: Anton's preferred course: no grand plan, just continuously pulling over-concentrated power back.

How to listen

Who it's for

Founders, investors and policy researchers watching AI geopolitics, export controls, compute infrastructure and the direction of European policy.

Skip

The customer-case ad segment around 22:13 can be skipped.