Anton Leicht on AI Geopolitics and Power

Ask Anton Leicht how dangerous today's AI models are, and he does not reach for a doomsday number. Leicht is a fellow with the Technology and International Affairs program at the Carnegie Endowment for International Peace and the author of a newsletter called Threading the Needle. His answer is calm: not very dangerous yet, in most of the ways people worry about (08:00). What concerns him is the trend line. He thinks models are nearing two thresholds at once. The first is a point where a model could meaningfully speed up AI research inside the very labs that build it, letting development outrun anyone's ability to watch it happen. The second is a point where models get very good at biology and pharmaceutical chemistry, the domain where a capable, misaligned system looks most frightening if something goes wrong (08:23).

"The pace of development inside the labs just breaks away from the pace of democratic oversight and democratic insight into what's happening." — Anton Leicht [08:50]

That gap, between what a lab can do and what an outsider can verify, runs underneath almost everything else Leicht has to say, across a conversation that covers Beijing's industrial strategy, a Norwegian sovereign wealth fund, and the geopolitics of shooting down satellites.

The Pause That Helps the Wrong Side

Leicht is unusual among safety-minded analysts in worrying that a popular idea, a temporary pause on frontier AI scaling, the practice of training ever-larger models on ever-more computing power, could be a bad deal for the United States. His reasoning has little to do with AI risk itself. It has to do with everything else. China, he says, is "eating America's lunch" in nearly every other domain of strategic competition: robotics, manufacturing, electricity build-out, AI diffusion (14:40). The one place the US is decisively ahead is what he calls the core frontier AI supply chain: chip design, chip production controlled by allies, and frontier model development itself (15:20). Freeze that one advantage for six months or a year, he argues, and China gets that much more time to build its own semiconductor manufacturing capacity, closing the single gap that still favors Washington.

"If you pause specifically that part of the development and let China run away with the entire rest of it... that's just a very geopolitically lopsided deal." — Anton Leicht [14:55]

A pause built on trust rather than enforcement carries an added risk, Leicht says: China could spend the frozen months smuggling in more American-made chips and consolidating them into a single national project, so that when the pause ends, the race resumes from a stronger Chinese position (20:14). A genuinely fair deal, he suggests, would require China to also freeze its chip indigenization efforts, something he thinks Beijing would never accept (17:42).

The same asymmetry shadows the stock market conversation. Leicht thinks a six-month pause would not be catastrophic for AI valuations if it read as a deliberate, industry-led move toward reliability. But if it reads as the start of an unpredictable regulatory crackdown, of the kind that spooked the market after recent political rhetoric, investors would rather exit early than wait to see how it resolves (28:42). That distinction matters because current AI valuations, in his view, already assume a lot. Anthropic's revenue multiple of roughly 30 to 1 only makes sense, he argues, if labs become "innovation factories" that accelerate drug discovery and materials science, not just tools that help an accountant work faster (27:20).

Leicht also worries about a subtler kind of erosion: powerful AI available to individuals could weaken the nation state's two core functions, its monopoly on legitimate violence and its role adjudicating disputes and gathering the information needed to govern. He thinks liberal democracies are more exposed to this than authoritarian states, since market-driven diffusion pushes capable models toward everyone's pocket, while a state like China can keep frontier capabilities inside a small, tightly controlled tier of firms and still deliver the economic benefits (39:07). That, he says, could make China's model of AI governance more stable than America's, at least for a while (40:03).

On the question of oversight, Leicht has a concrete, near-term proposal. Independent evaluators such as Meter or Redwood Research, nonprofit organizations that investigate AI safety incidents, currently depend on voluntary cooperation from the labs they study. Leicht argues the US executive branch could change that immediately, by directing frontier companies to grant a pre-approved list of third-party evaluators real access, rather than letting each company decide for itself who gets to look and for how long (49:40). A further step would let evaluators sit inside labs on an ongoing basis, with the ability to escalate to the government if something looks seriously wrong (50:06).

What the Rest of the World Gets

Beyond the US and China, Leicht's picture turns sober. The old path for low- and middle-income countries, cheap labor that draws in foreign manufacturing and, later, services work, depends on human labor staying economically necessary. Automated manufacturing and AI-driven services erode that advantage, he argues, and roughly 70 percent of the world's population lives in countries that depended on it (71:50). Life will likely get materially better even so, he says, but political leverage over how the technology develops will not follow. "It is very difficult to figure out how any country in that spot ever gets any leverage over what happens at the frontier," he says (74:10). Countries in this position become dependent on whichever great power supplies their AI, with wealth rising but with no seat in decisions about how the technology is built or governed.

Europe, Leicht believes, is one of the few places positioned to avoid that fate. His proposal, laid out in a recent report on transformative AI strategy for Europe that he co-authored, is what he calls compute for access: European governments offer American hyperscalers favorable energy and permitting terms to build data centers on European soil, in exchange for a guaranteed, revocable promise of continued access to frontier models (82:00). The arrangement is meant to be self-enforcing. If the US side ever cuts off model access, Europe can cut off the data centers. Backing that up, Leicht points to Europe's control of the semiconductor supply chain, including the Dutch lithography maker ASML, as leverage the continent could use if it ever needed to push back against American coercion (83:35). The obstacle, he says, is less technical than psychological: many European policymakers still doubt that frontier AI is as consequential as it looks, or assume Europe could simply build its own competitor cheaply (85:05).

Smaller countries face narrower but sharper choices. Norway has a large, already AI-savvy sovereign wealth fund and good conditions for building data centers, but has not yet turned that into a coherent national compute strategy (91:10). Singapore has extraordinary depth of AI expertise inside its civil service and parliament, but a service-heavy economy that leaves it exposed if white-collar automation arrives as fast as some expect (93:30). Australia, Leicht argues, is a sleeping giant: excellent land and energy for data centers, plus a level of security trust with US agencies that makes it a natural home for compute-for-access deals (94:15). The United Kingdom has the opposite problem, an unmatched density of AI talent outside the US, but no clear plan for using it (95:25). For countries without frontier chips of their own, the choice of ally is close to automatic for now, since China cannot yet offer the full-stack deal, chips plus data centers plus models, that the US can (97:00).

Even that calculus has a shelf life. Leicht expects some AI computing infrastructure to move into orbit starting around 2029 (109:25), a shift that would make anti-satellite weapons a meaningful part of national deterrence, since shooting down a satellite becomes the only way to physically threaten a system that isn't sitting in a data center on the ground (110:16). If enough compute eventually moves to space, the incentive for the US to build data centers inside allied countries shrinks, and with it, the leverage that middle powers are currently trying to build through deals like Europe's (112:10).

None of this adds up, in Leicht's telling, to a grand plan for AI's future. When asked to sketch his best-case outcome, he does not describe a utopia. He estimates something like a 10 percent chance of broad catastrophic outcomes, including stable authoritarianism and long-term political disempowerment, a category he treats as distinct from human extinction, which he considers far less likely (106:45). His hope is narrower and, he admits, deliberately modest.

"I think just, like, genuinely just muddle through... enjoy all of the obvious upside that AI has to offer." — Anton Leicht [123:54]

Muddling through, in his description, means constant small corrections: keeping labs from pulling away from government oversight, keeping any one government from monopolizing the flow of intelligence through the world, and keeping enough other countries with a real stake in the outcome that power does not simply pool in one place. It is not a vision built to inspire. It is a vision built to survive contact with a world that, as he keeps noting, is more likely to surprise everyone than to behave.