The Compute Bloc

I — From Strategic Input to Reserve Asset

The natural analogy for compute is oil, and for most of a century oil was the resource that decided outcomes. A state's access to oil set the limit on its industrial and military reach; the politics of the twentieth century were in large part the politics of securing it. Compute is taking that role for the century now beginning. The capability to train and run frontier models is becoming the input that sets the ceiling on what a state can do economically and militarily, and access to it is becoming the thing serious states organise around.

But the analogy is most useful at the point where it fails. Oil is extracted. Where it sits was decided by geology, and geology distributed it widely and somewhat randomly — across the Middle East, the Americas, Africa, Russia, the North Sea. Many states have some. The resulting politics, for all its violence, was a politics of many holders: a market, a cartel, a web of bilateral deals, and ultimately a set of treaties and institutions, because a resource that many parties possess eventually has to be governed by negotiation among them.

Compute is manufactured. Where it sits is decided not by geology but by a supply chain — and a supply chain has chokepoints that geology does not. The most advanced chips are designed by a few firms, fabricated by fewer still at the leading edge, using lithography equipment from effectively a single source. Nothing about this is distributed by nature. It is concentrated by industrial reality, and the concentration is more durable than any oilfield, because an oilfield can be discovered but a leading-edge fab must be built — at a cost of tens of billions, over many years, with knowledge held by a handful of organisations on earth. A thing this concentrated stops behaving like an input and starts behaving like a reserve asset: its possession, in quantity, constitutes power directly, and its distribution is narrow enough that simply holding it is a strategic position.

Oil was distributed by geology, and geology was promiscuous. Compute is distributed by a supply chain, and the supply chain has a door that a few hands can hold shut.

II — The Bifurcation and the Sorting

The trajectory that follows from concentrated, controllable compute is a world that sorts into tiers, and the sorting is already visible. At the top sit the actors who design the chips and control the fabrication and the lithography — the holders of the chokepoint itself. Below them sit the states that host frontier compute at scale, whether or not they control the supply chain that feeds it. Below those sit the compute-rich allies — states granted reliable access through alignment with a chokepoint holder. And below those sit the compute-dependent: states that consume frontier AI as a service, on terms set elsewhere, with no domestic capacity and no guaranteed access.

Almost every state now understands this and is reaching for what has come to be called sovereign AI — domestic compute, domestic models, a national capability not contingent on another power's permission. (Hogarth, 2023; OECD, 2024) The aspiration is universal. The achievement is not, and cannot be, because the fab chokepoint does not yield to national ambition on any relevant timescale. A state can announce a sovereign AI programme, fund it generously, and still find that the chips it needs come from a supply chain it does not control and cannot replicate. Building an independent leading-edge fabrication capability from a standing start is the work of a decade and a sum that few national budgets can sustain, and even then it depends on lithography that one company, in one country, supplies to the world. (Miller, 2022) Sovereign AI, for most states, means sovereign inference on imported chips — real, useful, and fundamentally dependent.

Competition will not erode these tiers, because they are not a passing disequilibrium. They are what a manufactured resource with a concentrated supply chain settles into, and they divide the world into the states that set the terms of access and the states that live with them.

The compute tiers

1. Chokepoint holders — Design the chips, control leading-edge fabrication and the lithography behind it. Set the terms of access for everyone below.
2. Compute hosts — Operate frontier compute at scale, whether or not they own the supply chain. Strategically powerful, but dependent upstream if they do not control the chokepoint.
3. Compute-rich allies — Granted reliable access through alignment with a chokepoint holder. Capability contingent on the relationship.
4. Compute-dependent — Consume frontier AI as a service on terms set elsewhere. No domestic capacity, no guaranteed access, no leverage over the terms.

III — The Choice Nations Face

Where a state lands in this structure is not entirely a matter of fate, and the interesting cases are the ones with a card to play. Begin with the states that have none. A small economy with no advanced manufacturing, no leading-edge fabrication, and a domestic market too small to matter to anyone is, on the most important input of the century, a price-taker in the fullest sense. It will consume frontier AI on whatever terms the holders offer, it will have no leverage over those terms, and it will have no recourse if access is conditioned or cut off. Such states do not make decisions on this axis at all; the decisions that shape their economic future are made elsewhere, by parties to whom they are not even a meaningful counterparty.

But factor endowment is not the only source of leverage, and this is where the structure becomes a negotiation rather than a hierarchy. A state may hold no chips, no fabs, and no frontier compute, and still hold something the chokepoint holders need: a large market. Frontier compute is built with enormous capital expenditure, and that expenditure has to be justified by returns, which means the models and the chips have to be sold into markets large enough to repay the investment. A state that controls access to a large consumer or industrial market therefore holds a card — the ability to grant or withhold entry to that market — that can be traded against access to compute. The holder needs the market; the market needs the compute; and between those two needs there is room to bargain.

The genuinely large markets that lack frontier compute of their own sit exactly here. They cannot match the chokepoint holders chip for chip, yet they can make access to hundreds of millions of consumers conditional — on technology transfer, on local data processing, on guaranteed compute allocation, on terms that a holder dependent on that market for returns cannot lightly refuse. The European Union is the paradigm case: a market too large to forgo, no frontier compute to speak of, and a demonstrated willingness to use regulatory control of market access as leverage. (Bradford, 2020) Its strength on this axis has nothing to do with technology. It is the size of the prize it can open or close. The same logic, in weaker form, is available to any state whose market is large enough that a compute holder would rather meet conditions than be excluded.

The states caught in the worst position are those with neither endowment nor market — too small in production to make their own compute, too small in consumption to bargain for anyone else's. They face the choice every minor power faces when a decisive resource is concentrated: align with a holder and accept dependence on terms, or remain unaligned and accept exclusion from the frontier. The non-aligned position, attractive in principle, is hard to sustain when the resource in question is the one that determines economic competitiveness across the whole economy, and when the holders have every incentive to reward alignment and little to reward neutrality. Most will choose a bloc, because the cost of choosing none is to fall behind on everything at once.

The fab takes a decade and a fortune that almost no one has. A large market takes nothing but the will to open or close it. Market size is the one form of leverage that does not require a single transistor.

— The Governance Question at the End of the Trajectory

The trajectory ends with a question of governance that no existing framework answers. Every prior strategic resource was, in the end, brought under some form of negotiated governance — treaties, institutions, cartels, shared standards — because the resource was held by enough parties that none could simply dictate, and governance was the alternative to perpetual conflict among holders. (Yergin, 1991; Krasner, 1983) The pressure that produced those institutions was distribution: because many had the resource, many had to be accommodated.

Compute may be the first decisive resource concentrated enough to escape that pressure. If the chokepoint holders are few, and the barrier to joining them is a decade and a fortune and knowledge held by almost no one, then the holders may never face the situation that forced oil and arms and trade into negotiated regimes — the situation in which there are simply too many holders to ignore. A resource that few possess and fewer can produce does not generate the distributed bargaining power out of which governance historically emerged. It generates the opposite: a structure in which the holders can set terms and the rest can take them.

That is not stability, though it may resemble it for a time. It is the absence of the condition that produces stability — the rough distribution of capability that obliges powers to build rules they can all live within. Two failure modes follow at the level of the international system. One is capture: a single bloc achieves durable dominance of the chokepoint and sets the terms of the machine age for everyone else. The other is fragmentation: rival blocs build incompatible compute worlds with no shared standard and no interoperability, and the global system splinters into parallel and mutually unintelligible technological orders. Neither is governed. Both are what happens in the absence of governance.

The question is the same at every scale, from the settlement of a single trade to the order of the international system: not who owns the decisive infrastructure, but who governs it, under what constitution, with what answer to the problem of accountability. In domain after domain, that question is already being answered the wrong way — by private actors, in the absence of public design. For compute, it has not begun to be asked at all. The metal is being assembled, the tiers are forming, and the constitution that would govern the result does not exist. No one has drafted it, no one has proposed it, and it sits on no agenda anywhere. That, and not the question of whether a model can be decentralized, is the thing worth being concerned about.

Every prior strategic resource was eventually governed by treaty, because eventually too many parties held it to allow anything else. Compute may be the first concentrated enough that the holders never have to sit down. The absence of conflict is not the same as the presence of order.