Sovereign Compute
National AI infrastructure programmes: stated objectives, committed capital and delivered capacity.
Abstract
Twenty-three states have announced national AI compute programmes since 2023. This forthcoming paper assesses them against three questions: what capacity has been committed, what has been delivered, and does the delivered capacity serve the stated objective. Preliminary findings indicate a large and widening gap between announcement and energisation, a consistent underestimation of operating cost relative to capital cost in programme design, and a pattern in which delivered capacity is sized for inference and sovereignty objectives while stated objectives reference frontier training.
Key findings
- 01Preliminary: announced-to-energised conversion across national programmes is materially below the private sector equivalent.
- 02Preliminary: operating cost is under-provisioned relative to capital in the majority of programme designs examined.
- 03Preliminary: delivered capacity is generally sized for inference sovereignty rather than frontier training.
This paper is in preparation. Scheduled publication: September 2026.
Twenty-three states have announced national AI compute programmes since 2023, with committed capital that in aggregate exceeds the annual infrastructure investment of the largest private operators. The programmes differ enormously in design, in objective and in execution. Very little comparative assessment of them exists.
This paper will assess each programme against three questions. What capacity has been committed, in units that permit comparison? What has actually been energised and is operating? And does the delivered capacity match the objective the programme states — because a facility sized for national inference sovereignty is a different thing from one capable of frontier training, and the two are frequently described in the same language.
Scope
- Programme-level documentation for all 23 states, normalised to comparable capacity and cost units.
- Physical delivery verification through the Frontier Labs infrastructure dataset.
- Assessment of operating cost provisioning relative to capital commitment.
- Comparison of stated objective against delivered technical capability.
Preliminary findings are published here in advance of the full paper because several are already being cited from conference presentations. They are provisional and may change. The full methodology and all intervals will be published with the paper.
Sources and references
- [1]National programme documentation, 23 states.
- [2]Frontier Labs infrastructure dataset, release 2026.2.
Source assessments, provenance records and the full data appendix are described in the methodology. Corrections to this publication are recorded on the corrections page.
Keywords
- sovereignty
- national programmes
- compute
- policy
Cite this publication
Policy Desk (2026). Sovereign Compute. Frontier Labs. https://frontierlabs.research/research/sovereign-compute-programmes
Research infrastructure powered by AGI-S.