November 21, 2026
Heard in AI

RAND researchers explain what AI dominance would actually require

RAND researchers told podcast host Nathan Labenz that a decisive AI lead might need technological, economic and military advantages at once, possibly a 'wonder weapon', and a theory for avoiding war with nuclear-armed rivals.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on The Cognitive Revolution, episode published October 28, 2026

Nathan Labenz, host of The Cognitive Revolution, thinks the United States is already playing a dominance strategy in AI, whether or not anyone announced one. His evidence is chip denial: for what he called "a couple of years," Washington has restricted China's access to the advanced processors used to train and run AI systems. "I'm like definitely a skeptic of that policy," he said, and he asked his guests for the strongest version of the case for it.

The guests were Joel Predd, Ben Boudreaux and Matt Chessen of RAND's Center for the Geopolitics of AGI, a group studying how the arrival of artificial general intelligence, and eventually superintelligence, could reshape national security. In the episode, published October 28, 2026, they did not so much make the case for dominance as take it apart: what a "decisive strategic advantage" would have to mean, which conditions could make it achievable, and where pursuing it could backfire.

A skeptic's question

A decisive strategic advantage, in this debate, is a lead so large that a rival can no longer meaningfully contest it. Labenz said the best-known public picture of how the United States might get there leaves out the hard part. He praised the first half of Anthropic CEO Dario Amodei's essay Machines of Loving Grace, with its cured diseases and better mental health. On geopolitics, though, he said, "boy, you skipped a lot of points there along the way, Dario." In his paraphrase, the essay's plan amounts to "something, something" eternal 1991, putting China in a box and making it "an offer they can't refuse." To Labenz, the ending sounds "a lot like regime change."

That is his reading. The essay itself proposes an entente of democracies that secures its supply chains, scales up AI quickly and delays adversaries' access to critical inputs. Amodei pairs military superiority with an incentive: countries that join the coalition would share in AI's benefits. He also treats democratic leadership as something that has to be deliberately achieved rather than something technology delivers automatically.

Three different kinds of lead

Predd began by asking what the advantage is supposed to be in. People who advocate dominance, he said, often have different sources of advantage in mind.

The first is the technology itself. Even here there are two versions: keeping a steady lead over time, or preventing a rival from ever crossing some specific, hard-to-identify capability threshold. Because AI progress is driven by economic forces as much as technical ones, Predd said, the second version is hard to believe in, though not impossible. A steady lead seemed less difficult to imagine given the pace of things.

The second is economic: turning leadership at the frontier into lasting growth and scientific innovation. Predd pointed to his RAND colleague Tobias Sytsma, who has written that the ingredients for this are quite different from the ingredients of the technology. Sytsma's February 2026 report offers a framework for analyzing when economic feedback turns a temporary technological lead into a durable constraint on a rival's ability to compete. Across a simulation space spanning deep parameter uncertainty, it identifies two robust competitive regimes: in some combinations of assumptions the leads shrink as technology spreads and the follower invests; in others they widen. Widening can come from AI improving itself, but also from more ordinary loops, such as learning from deployment and reinvesting profits in infrastructure. The follower's interventions have the most leverage when they come early. Sytsma describes the model as deliberately stylized, and he leaves the step from economic advantage to military or political dominance to other mechanisms.

The third is military, which requires translating AI into weapon systems, concepts of operation and strategic doctrine, "all of which don't come for free," Predd said. The crux, in his account, comes when a decisive strategic advantage in an era of advanced AI is taken to mean all three at once.

Playing China

Labenz brought a personal data point. He once played China in a tabletop exercise, a structured war game in which participants act out how governments might respond to a scenario, built around a fairly fast AI takeoff. Like most players, he said, his side kept ending up "a step behind events," which left him thinking there is "at least some chance" that everyone is paralyzed by indecision until an outcome arrives before anyone plays their cards.

He also felt strong. "As China, I have a lot of cards to play," he said, and what limited him was his own restraint and his desire to seek a constructive deal before playing hardball. His example was the chip factories, or fabs, of TSMC in Taiwan. "They're awfully nice. It'd be a shame if something happened to them," he joked. "They're like right in my corner of the world."

That led to his central question: does dominance only become plausible if timelines are very short, in a "FOOM"-like scenario? FOOM is shorthand for an extremely fast takeoff in which AI improvement compounds so quickly that the leader's gap widens long enough to take some pivotal action before the other side can respond.

Control is the catch in a fast takeoff

Boudreaux said the answer depends on more than timelines. Building on Predd's breakdown, he noted that the United States might hold a decisive advantage in one domain while China holds it in another, and those advantages may not add up in any clear way.

A fast takeoff could stretch a leader's lead, he agreed. But it also raises the problem of keeping control of the systems doing the work. "A decisive advantage is only good insofar as the leader can continue to lead the AIs," Boudreaux said. A government unprepared to steer the technology might lose that ability, and a very fast takeoff might not leave time to build the control and steering mechanisms needed to use the lead at all. More time, in his view, might be necessary.

Boudreaux added that even a successful dominance strategy might not last. Through model theft, distillation (training a new model to imitate an existing one) or the general spread of the technology, the United States might achieve dominance for a period and then need to rebalance toward a different approach. Choosing among strategies, he said, may never be finished "for all eternity" and could instead be "a continual process of adaptation."

Sabotage, cooperation and the wonder weapon

The conversation then returned to game theory, the mathematical study of strategic choices, through a paper by RAND colleague Philip Root. His June 2026 analysis treats the race between the United States and China as something like a prisoner's dilemma, in which each side does better for itself by defecting even though both would gain more by cooperating. In his starting model, a single-shot prisoner's dilemma, mutual cooperation yields the highest economic growth, yet mutual sabotage of each other's AI programs is the dominant equilibrium.

Two changes alter that picture. Adding an assumed "wonder weapon," a capability that would let one side hold the other's AI program at risk, strengthens the incentive to seek a monopoly and to strike first. This is the condition under which, as it was put in the discussion, a dominance strategy would work. In later versions of Root's model, which add repeated interaction and accumulating catastrophic costs, shared risks that are severe enough make cooperative nonproliferation the better choice for each side individually. Root simplifies away companies and other countries to isolate these incentives.

Labenz asked what a wonder weapon would actually look like. The answer on the show admitted up front that it would not be very satisfying. In the history of military innovation, such weapons are very rare. Most new capabilities change the balance only through adoption, as militaries build force structures, operating concepts and doctrine around them; one machine gun or one tank rarely changes the game. Nuclear weapons are the canonical exception, upsetting the balance simply by existing.

The term has a history at RAND. In a February 2025 paper, Predd and Jim Mitre listed wonder weapons as one of five hard national security problems posed by artificial general intelligence, alongside shifts in global power, wider access to weapons of mass destruction, artificial entities with agency and instability. Their hypothetical examples include cyber capabilities that disable an adversary's ability to retaliate.

The discussion also noted that some have argued Mythos-class models were the first wonder weapon of the AI era, because of how quickly they can discover and exploit cyber vulnerabilities. Anthropic's April 2026 assessment of Claude Mythos Preview supports part of that picture. Over several weeks of testing, agents searching isolated copies of software found flaws such as a remotely triggerable crash in OpenBSD, and the model autonomously developed an exploit for FreeBSD. The results were uneven: it found a memory-corruption bug in a virtual-machine monitor but did not produce a working exploit for it. Anthropic limited early access to defenders hardening important software and said it expects defensive benefits to predominate in the longer run. The assessment is about software security; it does not test whether such a model could shift a military balance.

Beliefs as much as facts

One fundamental challenge, Labenz said, is how much everyone has to imagine about what a smarter-than-human AI might invent. That prompted a point from RAND's tabletop work, which the discussion described as hundreds of games, including crisis simulations and U.S.-versus-China strategy games. A recurring finding is that players' starting beliefs about a few pivotal questions shape how the games turn out. A player who believes a decisive strategic advantage is achievable, or that danger is near, produces a very different game; perceptions about the facts, it was said, "matter as much, if not more" than the facts themselves.

The nuclear era had uncertainties too, but also some certainties that are missing now. Scientists at Los Alamos could run a calculation that gave them confidence the Trinity test would not ignite the atmosphere. No comparable science exists today to settle whether danger is close, whether humans can coexist with superintelligence or whether a decisive strategic advantage is even real.

Born into a nuclear world

Labenz doubted that AI could deliver the level of power dominance would need, something that makes a rival's nuclear arsenal irrelevant. He pointed to a current war, which he did not name, between two countries with plenty of skilled programmers. It is being fought heavily with drones, he said, but "the cyber story, there hasn't been that big." He struggled to see cyber weapons taking an adversary's nuclear arsenal offline, and without that, the rival keeps "a pretty strong hammer."

The reply on the show began with a joke that Labenz, "as a good host," was trying to put the researchers in a corner and get them to take a position, and then partly agreed with him. Superintelligence will be "born in a world of nuclear weapons," the discussion went, which would be very different had it arrived in 1930. That points to one of dominance's failure modes: "you have to have a theory that you will not escalate to a catastrophic war in which you're trying to permanently suppress your rivals."

A lesson from nuclear history

The discussion also warned against expecting the decades-long stability of the nuclear era. AI strategy may move through similar phases, only faster. U.S. nuclear strategy was described as shifting over time without anyone deliberately planning for it: a sprint in the Manhattan Project; a temporary monopoly of about four years that the United States used for geopolitical advantage; a period of anxiety in which some proposed something resembling a moratorium, which "we didn't proceed with"; eventual verifiable agreements with rivals; and later, suppression efforts to keep the technology from Iran and North Korea "with limited success."

The episode did not name that moratorium-like proposal. A likely historical referent, in our reading, is the Baruch Plan. According to the State Department's history, Bernard Baruch presented a U.S. plan for international control of atomic energy on June 14, 1946. In December, ten members of the UN commission supported it, but the Soviet Union and Poland abstained, blocking the unanimous approval it needed. Soviet objections included inspections and the United States keeping its monopoly during the transition. If that is the episode being referred to, it went further than "didn't proceed" suggests.

The broader point in the discussion was that policymakers then had the luxury of time to work their way through these phases, and the AI era will not offer it.

Signals that harden resolve

Chessen took up the cost of pursuing dominance through moves like chip controls. Labenz had quoted a line from RAND's report urging the United States to weigh carefully any action that would harden a rival's resolve, and said that after a couple of weeks in China over the summer, the take he heard was that "these chip restrictions are like hardening our resolve."

Chessen said moving toward a suppression capability can tell rivals that their window for action is closing, which can push them to accelerate. He cited an argument that chip controls fed China's drive to build its own market for AI accelerator chips, because "now it's existential": with less access to GPUs on the international market, China has to make them itself. "I think there's some logic to that," he said.

The most dangerous signal, in Chessen's view, would be for the United States to suggest that a dominance strategy and the monopoly it brings would be used to threaten the Chinese Communist Party. Anything that becomes existential for a rival "hardens their resolve," he said, so Washington should be careful about how its messages are received and avoid "boxing your adversaries into a corner."

Labenz said he had already seen that hardening at one level: China is "going to have a chip industry at this point." His hope, he said, is that the resolve has not hardened much further than that.

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02
  3. 03
  4. 04
  5. 05
  6. 06

Connected ideas and articles

From the conversation

Podcast episodes

The Cognitive Revolution

Freedom of Action: America's ASI Strategy, w/ RAND's Joel Predd, Ben Boudreaux & Matt Chessen

Episode published This article draws on 29:41–38:08, 43:30–53:44, 1:18:42–1:19:39 and 1:23:04–1:25:04 (approximate times)

Article history

Updates to this article

Tags