Executive Summary

The "90%" the UN report cites is a powerful headline, but the story sharpens once you open it up. The breakdown reported by the first-wave coverage is roughly 75% for the United States, about 15% for China, and around 10% for the entire rest of the world. The 90% aggregate is the share two countries split between them; the world's remaining 190-odd nations jostle for position inside the leftover 10%. Divide that by country and the ratio collapses fast toward zero.

Put that distribution on a single bar and the shape of the gap becomes plain. The figure below simply transcribes the US, China, and rest-of-world shares of compute.

1.1This is not the "supercomputer ranking" you know

One common confusion needs clearing up first. This figure is a different metric from the Top500 supercomputer ranking we're all familiar with. The traditional Top500.org is a general-purpose HPC list based on the LINPACK benchmark and voluntary submissions, and China stopped submitting in 2022. The June 2026 Top500 was in fact topped by a Chinese system — but that is not a ranking of the compute used to train and verify AI. The value the UN cited is closer to a separate index tallied by GPUs and AI accelerators, and it lines up with the dataset from the private research institute Epoch AI to within rounding error. Epoch AI's own country-level compute shares come to 74.5% for the US and 14.1% for China — diverging from the UN's rounded 75% and 15% only after the decimal point.

In other words, "the world's fastest supercomputer" and "the compute to train and verify AI" are not the same thing. The very fact that the number-one changes depending on which metric you count by shows that this debate is ultimately about what we have decided to measure. The table below sets the two metrics side by side.

1.2Reading the phrase "191 countries at 0%" more carefully

Several outlets that carried this report attached the phrase "the remaining 191 countries are effectively at 0%." It's vivid, but it is not a figure that appears in the source. "191" shows up nowhere in the UN's primary materials. Secondary and tertiary coverage simply did the arithmetic — 193 UN member states minus the US and China leaves 191. This piece does not adopt that number as-is. The accurate statement is that "10% is split among 190-odd countries," and that once you break it down per nation, most countries' share rounds away to nothing.

There is a firmer figure that the report actually put forward. Count the countries that produce notable AI models and you get 59 for the US, 35 for China, and 13 for the entire rest of the world combined. This model-production count shows the concentration of capability more sharply than an estimated "compute share" ever could. The "118 countries" (absent from governance discussions) and "191 countries" (holding no compute) that different outlets attached are numbers answering different questions — mix them into one sentence and you invite misreading.

AI governance debates tend to fixate on "who makes the rules." But a rule only works when someone has the capacity to check it. Just as vehicle emissions standards mean nothing without the equipment and test labs to actually measure exhaust, a safety rule for frontier AI is only enforceable if there are hands that can pull the model apart and put it to the test. Those hands are compute and model access.

Verification comes in two layers: black-box verification, which probes a model's surface through cheap API queries, and white-box verification, which reaches into the internals by accessing weights, activations, and gradients. The two differ entirely in cost and in what they require. The figure below marks the boundary.

By the numbers alone, verification is far cheaper than training. A single training run for a frontier model can reach hundreds of millions of dollars, while red-teaming and evaluation run from hundreds to tens of thousands of dollars per task. A recent benchmark measuring autonomous red-teaming capability also put per-task costs in the tens-to-tens-of-thousands range. So it's easy to jump to "verification doesn't need money." But that cheapness comes with a condition: it assumes access to the model — an API or the weights — is already secured.

The real bottleneck is not money but access. And the depth of access determines the depth of verification. There is a class of backdoors — behaviors planted to trigger only under specific conditions — that has been proven computationally impossible to detect through black-box access alone. Deep safety verification — re-training, interpretability analysis, auditing internal representations — is only possible if you hold the weights in hand and can run compute on par with training. This isn't a simple "white-box costs a few times more than black-box" comparison; it means that for white-box, owning the compute and the weights is a precondition in the first place.

One figure helps gauge the scale of compute verification takes. The total development compute to bring a single model to completion runs 1.2 to 4 times a single final training run — a median of about 2.2 times. Experiments, fine-tuning, and evaluation all live inside that multiple. Put differently, a country without compute not only cannot reproduce the training run; it also lacks the evaluation and interpretability compute that runs to a multiple of it. Verification demands a material base as large as training — sometimes larger.

Overlay the previous two sections and one concept emerges: verification sovereignty — a country's capacity to independently audit and test frontier AI within its own borders. That capacity stands on a material base of compute, weights, and talent, and that base exists in only a tiny handful of countries. The authority to make rules can be shared by a vote; the capacity to verify cannot.

This monopoly is stamped directly onto the map of international cooperation. The international network of AI Safety Institutes (AISIs) was created to jointly test frontier models, but its members are a small club of advanced economies centered on the UK, US, and Japan AISIs and the EU AI Office. Even joint US–UK testing is still in early days. It carries no legal force and leans on voluntary cooperation. The result: "countries that hold compute" and "countries inside the verification network" almost perfectly overlap.

What does the world outside both circles look like? Data-center compute held by low-income countries is somewhere around 0.1% of the global total. Africa is home to roughly 18% of the world's population but under 1% of its data-center capacity. For these countries, frontier-AI verification is not a budget problem but a problem of physical absence: the equipment to test with is simply not inside their borders.

UN Secretary-General António Guterres summed up the danger as a risk that "the digital divide could harden into an AI divide." On top of the scars two decades of unequal internet access have left, a new inequality — of verification capacity — could now be layered. Countries that sign the rules but cannot check those rules for themselves make up the world's majority: that is the terrain we stand on.

Public investment aimed at narrowing the gap clearly exists. The US is trying to share compute with researchers through the National AI Research Resource (NAIRR); the European Union has earmarked tens of billions of euros for AI Factories and Gigafactories. India runs the IndiaAI Mission, and Gulf states are mounting large compute investments led by sovereign-wealth funds. The UN has proposed a global fund to close the compute gap. The problem is that all this public and international capital is off by orders of magnitude from the private sector.

Lay the numbers out on a log scale and the chasm comes into view. In the figure below, each step of a bar means a tenfold difference.

The four US hyperscalers' 2026 AI infrastructure investment comes to roughly $725 billion, sharply up year over year. The initial target of the UN's proposed global fund, by contrast, is on the order of $1 to $3 billion. The difference runs from about 240× to 700×. Even the US NAIRR has had only a tiny fraction of its recommended budget actually appropriated. When the orders of magnitude diverge this far, public investment doesn't close the gap so much as slow the rate at which it widens.

The more fundamental current is that the ownership structure of compute itself is tilting toward the private sector. The share of AI compute under public control fell from around 40% circa 2019 to under 20% by 2025, while over the same period the private share climbed toward 80%. The state where "191 countries sit near 0%" is not those countries' lack of will; it is the structural result of the world's compute concentrating into a handful of private firms. Within this structure, trying to close the verification gap through compute expansion alone is apt to become water poured into a leaky bucket.

So must the majority of countries and organizations short on compute give up on verification? No. The fact that white-box verification is blocked does not mean verification itself is impossible. Instead of the single direction of a compute arms race, there is another direction: maximizing verification power per unit of resources. The axis of that direction is data.

Data-centric verification boils down to three practices. First, well-curated evaluation sets. A small, high-quality eval set that precisely captures risk scenarios produces a sharper signal, with less compute, than indiscriminate large-scale querying. Second, data lineage and quality audits. Tracing what a model learned — the provenance and quality of its data — lets you indirectly verify a substantial part of the risk without opening the internals at all. Third, targeted benchmarks. A benchmark aimed at a specific failure mode, rather than general performance, concentrates limited resources on the sorest spot.

This approach also translates into the language of procurement. When an organization adopts AI, the question it should ask is not only "is this model good?" but "can we independently verify this model?" When the answer is no, data-based proxy verification — a well-designed eval set and a lineage audit — becomes the realistic second-best. If a perfect white-box audit is the privilege of a few, data-centric verification is the practical lever the rest of the world can actually get its hands on.

Primary sources (UN)

• 1.UN Independent International Scientific Panel on AI (2026). Preliminary Report. United Nations.
• 2.UN News (2026-07). UN AI panel report ahead of first global governance dialogue. news.un.org.
• 3.IPS News (2026-07). UN Artificial Intelligence Panel Launches Report Ahead of Global Conference. Inter Press Service.
• 4.Digital Watch Observatory (2025). UN Secretary-General report on voluntary financing options for AI. (Basis for the Global Fund for AI, $1–3B.)

Academic papers

• 5.Sastry, G., Heim, L., Belfield, H., et al. (2024). Computing Power and the Governance of AI. arXiv: 2402.08797.
• 6.Cottier, B., & Rahman, R. (2024). The Rising Costs of Training Frontier AI Models. arXiv: 2405.21015. (Total development compute ÷ final training run = median 2.2×.)
• 7.Casper, S., et al. (2024). Black-Box Access is Insufficient for Rigorous AI Audits. FAccT 2024 (Apollo Research).
• 8.AIRTBench (2025). Measuring Autonomous AI Red Teaming Capabilities in Language Models. arXiv: 2506.14682.
• 9.Epoch AI (2025). Trends in AI Supercomputers. arXiv: 2504.16026.
• 10.(2026). Expanding External Access to Frontier AI Models for Dangerous Capability Evaluations. arXiv: 2601.11916.

Data, policy, institutions

• 11.Epoch AI. Data on GPU Clusters · AI supercomputers performance share by country. epoch.ai.
• 12.Stanford HAI (2026). The 2026 AI Index Report.
• 13.TOP500.org (2026). 67th TOP500 List (June 2026).
• 14.NSF (2025). Foundations for Operating the NAIRR (NAIRR-OC), NSF 25-546.
• 15.European Commission. AI Factories / AI Gigafactories (InvestAI); IndiaAI Mission (indiaai.gov.in); Gulf sovereign-wealth funds (MGX, HUMAIN, QIA), public announcements.
• 16.CSIS / CFG. AI Safety Institute International Network; METR, Protocol FAQs / Frontier Risk Report.

Note: Epoch AI's dataset covers only about 10–20% of the world's AI compute (self-disclosed), and captures only about 2% of China's export-controlled chips. All country-level shares are estimates based on public information and may differ from the true distribution by more than five percentage points. "191 countries" is a secondary-media arithmetic estimate not present in the original report, so this piece uses "10% split among 190-odd countries" instead.