Compute Concentration

Capacity is a physical fact before it is a financial one. This paper reconstructs it from the physical record.

There is no authoritative public register of AI training capacity. What exists instead is a patchwork: corporate announcements that conflate contracted, installed and operational capacity; national estimates built by aggregating those announcements; and analyst figures that inherit the same conflation. The result is a literature in which the same capacity is frequently counted more than once, and in which announced sites that have not yet been energised appear as though they were running.

We took a different approach. Rather than starting from what operators say, we started from what physical infrastructure exists and what it can support. Every facility in this dataset has an electrical interconnection record, a planning footprint, and a construction timeline reconstructed from sequenced imagery. Capacity is bounded above by delivered electrical supply and by cooling capability, both of which are matters of public record in most jurisdictions even when the deployment inside is not.

Method

For each site we establish four quantities: energised electrical capacity, contracted electrical capacity, gross building area suitable for high-density racking, and a construction completion state. From these we derive an upper bound on accelerator count under stated power and cooling assumptions, then narrow that bound using hardware import data at the regional level and, where available, disclosed deployment figures.

The resulting estimate is an interval, not a point. We publish the interval. Where the interval is wide enough that the site cannot be usefully classified, we mark it indeterminate rather than assigning a midpoint. Thirty-one of the 214 sites are indeterminate under this rule.

Concentration has increased

The dominant narrative of the past two years has been one of diffusion: more countries, more operators, more sites. All three are true in count terms and false in capacity terms. The number of facilities grew substantially. The share of capacity held by the largest facilities grew faster.

The mechanism is straightforward. Frontier training benefits from co-location — the interconnect bandwidth required between accelerators falls off sharply with physical distance, and the largest runs are correspondingly built at single sites. Each generation of frontier model has required a larger single-site deployment than the last. Smaller facilities absorb inference and fine-tuning load, which is growing quickly but does not concentrate in the same way.

The energisation gap

The most operationally significant finding in this dataset concerns the difference between announced and energised capacity. Roughly a quarter of the capacity that has been publicly reported as deployed sits at facilities whose energised electrical supply is below the level that deployment would require.

There are legitimate explanations for part of this. Phased build-outs are normal; capacity is often announced at final scale and delivered in stages. Behind-the-meter generation is not always visible in interconnection records. But the gap is larger than these explanations comfortably cover, and it is systematically in one direction.

Implications

• National compute statistics are a poor guide to national capability. Four countries in this dataset would lose more than 80 per cent of capacity with the loss of a single site.
• Capacity estimates derived from corporate announcement will overstate operational capability by a margin that has grown over time.
• The physical concentration of frontier training creates a set of single-point dependencies — electrical, thermal and geographic — that are not reflected in any published risk assessment we are aware of.
• Diffusion of inference capacity is genuine and is proceeding faster than diffusion of training capacity. Conflating the two produces materially wrong conclusions.

The full dataset schema, per-site intervals and the complete source list are published alongside this paper. Site identities are withheld where publication would create a security concern; aggregate figures include those sites.