US Runs 75% of World's Leading AI Supercomputers, China 15%, UN Report Finds
China holds about 15%, according to the 40 scientist panel behind the report. World leaders meet in Geneva on July 6–7 to decide what to do about that imbalance.
China holds about 15%, according to the 40 scientist panel behind the report. World leaders meet in Geneva on July 6–7 to decide what to do about that imbalance.
The United States runs roughly 75 percent of the world's leading AI supercomputers. China runs about 15 percent. When the UN's first independent global scientific assessment of artificial intelligence reaches world leaders at the UN Global Dialogue on AI Governance in Geneva on July 6–7, the people negotiating its findings will be sitting in two countries that already own most of the infrastructure the whole debate is about.
That asymmetry is not editorial speculation. It is the UN panel's own measurement, laid out in the preliminary report released this week by the Independent International Scientific Panel on AI, a 40-member body of international scientists and experts convened by the UN. The report frames the concentration as the central governance problem the world has not yet figured out how to solve.
"The science is here," UN Secretary-General António Guterres said at the launch press conference of the assessment, welcoming the panel's findings. The phrasing matters. The panel is not asking governments to trust a vendor, a regulator, or a think tank. It is asking them to act on a science-based reading of where AI capability actually sits, and where the levers for governing it do not.
The report is blunt about the gap. AI capabilities, the panel writes, are advancing faster than the governance frameworks meant to oversee them. The risks it names are concrete: cybersecurity exposure, labor-market disruption, misinformation, fraud, deceptive AI behavior, and a lack of reliable methods to keep advanced systems under meaningful human control. On the upside, the same assessment points to potential gains in healthcare, education, climate work, and economic development. The warning is not that AI is bad. The warning is that the world has not built the institutions to handle either the harms or the benefits at the speed they are arriving.
Two structural facts make this harder than a typical UN communique. First, the 75/15 split means that any governance regime negotiated in Geneva will be written largely by, or under the shadow of, the two countries with the most compute to lose. Smaller economies and developing countries, the report notes, face widening gaps in computing infrastructure, data resources, and technical capacity to participate in the AI build-out, let alone regulate it. Second, the panel's own warnings about human control are not abstract. They name the difficulty of keeping advanced systems aligned with human intent as a present-tense technical problem, not a future hypothetical.
The Geneva dialogue on July 6–7 is where the report lands. It is a two-day meeting of governments convened by the UN to discuss AI governance in advance of broader negotiations later this year. The panel's findings will be presented there. What happens after that is the open question. The report does not prescribe a binding regime. It lays out the case that the window for coordinated action is closing, and that without it, AI's benefits will concentrate in the same countries that already own the compute.
That framing carries its own risk. A UN panel warning about governance asymmetry can read, in some capitals, as a call to slow down the leaders. In others, as a call to catch up. Either reading assumes the 75/15 split is something the Geneva dialogue can renegotiate. The report does not say it can. It says it is the room Geneva is walking into.
For now, the most concrete fact is also the simplest. Two countries hold roughly 90 percent of the leading AI supercomputers the rest of the world is trying to govern. The UN has, for the first time, put a number on it. The question for July 6–7 is not whether the asymmetry exists. It is whether Geneva can write rules that survive it.