RAND Warns the AI Race Will Leave Middle Powers Behind

The compute, capital, and talent needed to field a sovereign AI system are now steep enough to function as a foreign-policy problem, and the leverage is concentrating in two capitals. That is the case Matan Chorev, the analyst who runs RAND's Global and Emerging Risks program and once ran strategic planning at the State Department, made in a recent RAND Q&A co-authored with colleague Salman Ahmed.

RAND is the federally funded, U.S. national-security research center with deep Pentagon and State Department ties. That lineage matters: Chorev is not a futurist on a podcast. He is a working policy analyst with a federal-planning track record, and the argument is a thesis he is willing to put his name on.

The thesis turns on a single phrase: strategic autonomy. In plain terms, that is a country's ability to set its own course in economics, security, and technology without leaning on a great power. Chorev and Ahmed argue that AI is the first technology in decades whose basic requirements, including compute, model training, state-of-the-art data, energy, and the engineering talent to wire all of it together, are concentrated in so few hands that the rest of the world defaults to dependency rather than choice. The most visible artifact of that default is the chip export-control regime Washington has run since 2022, and the mirror tools Beijing has built around its own rare-earth and foundry choke points. The less visible artifact is procurement: when a mid-sized country wants a state-of-the-art model, a hyperscale training cluster, or a credible domestic AI safety office, the qualified-supplier list is short.

The mechanism is not a single wall. It is a stack: cutting-edge chips designed by a handful of firms, cloud training routed through a small set of hyperscalers, state-of-the-art datasets curated at a scale no national institute can match, the energy footprint of a frontier training run, and the standards bodies where model evaluation and safety norms get written. Each rung adds a dependency. Add them up and a country that wants to be its own AI power has to win on every rung at once, which is the part of the Chorev–Ahmed thesis that does most of the work and the part that is hardest for an outside reader to falsify.

That is the case for a two-country AI world. The case against is real. Open-weight models, where the trained parameters are published so any lab can run or fine-tune them, including releases from labs like Meta and the French startup Mistral, have repeatedly shown that capability can leak downward faster than gatekeeping can catch up, and several middle-power programs have built credible domain-specific stacks in fields like biomedicine, geospatial intelligence, and language. A separate RAND commentary on planning for a tech-driven geopolitical crisis, "America Should Assume the Worst About AI," concedes the same point: diffusion may run ahead of any one state's policy levers, and the Chorev–Ahmed thesis is best read as a tendency, not a verdict.

The middle position is the one worth holding provisionally. Sovereign AI is now a continuum, and where a country lands on it depends on which rungs of the stack it owns and which it rents. The reader's job is to watch three specific signals and update accordingly: the next middle-power AI procurement decision (especially a defense or intelligence agency buying frontier compute from a non-domestic vendor), the next open-weight release from a lab outside the U.S.–China axis that meaningfully raises the capability floor, and the next round of chip controls, which will signal whether Washington is widening or narrowing the gate.

Chorev is not arguing the game is already lost. He is arguing that the planning assumption U.S. policymakers carried out of the Cold War, that strategic autonomy would spread rather than concentrate, is the assumption that has to be retired. Whether the rest of the world can build, rent, or refuse its way out is the question the next eighteen months will answer.