HWNEWS
The Third Tier Of Compute: How Mid-Sized Countries Are Building Their Own AI Stack
AI compute is the physical infrastructure behind advanced AI systems: chips, hyperscale data centers, and the power grids that feed them. The U.S.
AI compute is the physical infrastructure behind advanced AI systems: chips, hyperscale data centers, and the power grids that feed them. The U.S.
The race for AI compute has been miscast. For most of the last three years, the story has been framed as a contest between two giants, one slowing the other down with export controls and the other trying to outbuild the first with subsidized factories. That frame is not wrong, but it is also not where the next consequential decisions are being made. The real contest is over who can permit, finance, and interconnect a gigawatt-scale data center on a timeline measured in years rather than decades. On that contest, the U.S.–China duopoly is simultaneously the most capable and the most constrained actor in the world.
The numbers explain why. According to a 2025 analysis of 500 supercomputers cited by Forbes contributor Nili Gilbert, the United States and China together account for roughly 90 percent of global AI compute performance, split about 75 percent to the U.S. and 15 percent to China. Those figures compress a stack of inputs (frontier model training capacity, hyperscaler data-center buildouts, advanced chip production, and a tightening export-control regime) into a single ranking. On June 12, 2026, the U.S. government suspended foreign access to Anthropic's two most capable AI models under export control authority, the most concrete signal yet that compute is being treated as a controlled technology rather than a commercial input.
Both governments now treat AI compute as an existential competitive challenge, on par with economic dominance, scientific leadership, and military capability. The 75/15 split, however, hides a harder problem: the physical substrate beneath the numbers. Building frontier-class AI capacity requires gigawatts of firm power, multi-year grid interconnections, water for cooling, and a supply of large transformers that is itself a multi-year wait. As Gilbert argues in her analysis for Forbes, the industrial policy response on both sides of the duopoly is straining against those constraints rather than closing them. The same wall is being hit from opposite directions, and neither superpower can scale around it quickly.
That wall is the opening. A third tier of countries, large enough to need sovereign AI capacity but not large enough to subsidize the entire stack, is now designing policy in the gap the dominant pair cannot close. Sovereignty is the largest of those gaps: every government that depends on foreign-controlled AI systems, for defense planning, financial infrastructure, public health modeling, or scientific compute, is structurally exposed to a foreign power that can change the terms of access at will. The June 12 export-control action made that exposure legible to anyone who had not been paying attention.
The third tier's response is concrete and ongoing, even if it does not yet have a shared name. Mid-sized economies are signing sovereign-cloud deals with domestic and regional operators, commissioning national data centers, passing compute-sovereignty bills that require sensitive workloads to run on infrastructure under national jurisdiction, and tying grid-expansion planning to projected AI load. These are not symbolic gestures. They are procurement decisions with multi-billion-dollar price tags, multi-year build horizons, and explicit conditions about where the bits physically live.
For investors and operators, the practical question is no longer whether the duopoly will widen its lead. The numbers suggest it will, and the policy posture in both Washington and Beijing confirms it. The practical question is which mid-sized actors can convert policy intent into commissioned megawatts before the export-control regime freezes their procurement options. A window of roughly 18 to 36 months — the time, in the reporter's analysis, it takes to permit and energize a hyperscale site — is also the window in which sovereign-cloud contracts, chip-of-origin rules, and grid-interconnection queues will harden into facts on the ground.
The compute race is no longer a contest over who owns the chip or the model. It is a contest over who owns the next binding constraint, the gigawatt, the interconnect, the transformer, and the rule that decides where sensitive compute can physically sit. The duopoly will set the ceiling. The third tier will decide how much of its own AI future it can run on infrastructure it actually controls.