When Compute Becomes a Sovereign Reserve Asset

Setting up the experiment

Suppose compute capacity became something states held strategically — not merely procured for use, but stockpiled as a reserve, accounted for as a national asset, and drawn on in emergencies.

This is not a prediction. It is a device for examining what properties compute actually has, by asking what would follow if it were treated the way reserve assets are treated. Some of the consequences are counterintuitive and clarify something real about the underlying resource.

Where the analogy breaks

Compute fails the classical reserve-asset tests in instructive ways.

It depreciates rapidly. Gold held for fifty years is still gold; a data centre held for five is substantially obsolete. A compute reserve is therefore not a store of value but a maintained capability with a continuous replacement cost — closer to a standing military than to a vault.

It requires continuous energy. An idle reserve still consumes to remain ready, which means the reserve is only as strong as the power system beneath it. This makes compute reserves inseparable from energy policy in a way gold reserves never were.

And its value is contingent on the software and talent to use it. Idle compute produces nothing. A reserve without a trained population to direct it is a depreciating asset with an electricity bill.

What would actually follow

If states nonetheless treated compute this way, three consequences seem robust.

Energy policy would become AI policy explicitly rather than implicitly. The binding constraint on a compute reserve is generation and transmission, which would pull long-horizon energy investment into the centre of technology strategy — arguably where it already belongs.

Compute-sharing arrangements would become instruments of alliance, in the way basing rights and defence guarantees are. Access during a crisis is exactly the sort of contingent commitment alliances are built from.

And the accounting would be contested. A depreciating, energy-hungry, talent-dependent asset is genuinely hard to value, and the valuation method would become politically consequential — as it does for every strategic reserve.

The useful residue

Stripped of the framing, the experiment yields something practical: compute is best understood not as a stock but as a flow capability, jointly determined by hardware, energy and human capital, none of which substitutes for the others.

This has an immediate implication for states planning AI investment. A programme funding hardware without matching energy and talent is buying one factor of a three-factor input. The result is not a third of the capability. It is closer to none of it — which is the standard failure mode of compute-led national programmes, and the reason the thought experiment is worth the detour.