Compute as currency

Compute as currency. Chips, models and cloud capacity are no longer ordinary exports but instruments of statecraft—granted to allies, withheld from rivals, traded for investment and alignment. Three developments in recent weeks suggest the transactional era has arrived; the harder question is how long such leverage survives being used. For most of the past decade, advanced computing behaved like any other high-value export: designed in one country, fabricated in another, sold to whoever could pay. That world is closing. Chips, frontier models and cloud capacity are increasingly treated as strategic assets—allocated to allies, denied to rivals, and exchanged for investment, security commitments and political alignment. The shift has been gradual enough to miss. Three developments in recent weeks make it difficult to miss any longer. The first is that the United States is selectively opening its gates. Earlier proposals in Washington had floated making major chip allocations conditional—tied to security guarantees, or to investment in American data-centre capacity. The decision to permit advanced Nvidia exports to the United Arab Emirates suggests that this transactional logic is no longer hypothetical. Access is becoming something to be negotiated for, priced, and periodically reviewed, rather than simply purchased. What was once trade is starting to look like foreign policy conducted by other means. The second is that China may be preparing to close gates of its own. Beijing is reported to be weighing restrictions on overseas access to its most capable AI models, treating model weights and the techniques behind them as national-security assets rather than exportable products. This is the mirror image of the American move, and it matters as a signal: the country that has spent three years on the receiving end of export controls appears to be concluding that its own frontier capabilities are worth guarding by the same logic. The instrument, once demonstrated, invites imitation. The third is that Europe has recognised that being an ally guarantees nothing. The European Union's technological-sovereignty package, unveiled in June, spans chips, cloud, artificial intelligence and open source. Its underlying premise is uncomfortable but clear: dependence on American platforms has become a strategic vulnerability even for friendly states. A partner who controls the switch is still someone who controls the switch. Europe is not being denied access; it is reasoning about a future in which access might be conditioned, and deciding that the risk alone is intolerable. Read separately, these are three national stories. Read together, they describe a single system taking shape—one in which computational capacity functions as a currency of statecraft. And like any currency, its value depends on confidence. That is where the logic turns on itself. Used as a lever, dependence stops being a fact and becomes a problem to be solved. Leverage of this kind is powerful precisely because it is scarce and hard to substitute. But each use advertises the vulnerability it exploits. When Washington conditions access, it teaches every recipient that dependence is a liability to be engineered away. When Beijing walls off its models, it confirms that the frontier is a contested border rather than a common good. When Brussels funds sovereign alternatives, it signals to others that self-reliance is now a defensible national expense. The result is a steady, rational drift toward fragmentation: indigenous fabrication, domestic model development, regional cloud, and the grey-market workarounds that always accompany hard controls. Every one of these erodes the chokepoint that made the leverage work in the first place. This is the central tension of compute-as-statecraft. The tool degrades with use. Export controls buy time and extract concessions in the near term, but they also set a clock running on the very dependence they exploit, funding the rivals and the redundancies that will eventually make the controls toothless. The short-term alignment is real. So is the long-term incentive it creates to escape. It is worth stating the counter-case plainly, because it may be the stronger one. Some chokepoints are not so easily engineered away. The most advanced lithography passes through a handful of firms—one Dutch, one Taiwanese foremost among them—whose lead rests on decades of accumulated tooling that money alone has not yet been shown to buy quickly. If the critical bottleneck is durable enough, fragmentation remains mostly rhetorical: rivals announce sovereign ambitions, spend heavily, and still arrive years behind at the leading edge. On that reading, the leverage does not decay so much as persist, and the drift I have described is real only in the commodity layers while the frontier stays firmly held. I find this partly persuasive, and it should temper any confident prediction of a single endpoint. What it does not do is remove the incentive to escape; it only lengthens the time the escape takes. None of this argues that states should refrain. A government that declined to use the leverage it holds would simply be leaving an instrument on the table for others. The more useful observation is that this is not a stable equilibrium but a transition—one whose endpoint is probably not a single integrated market restored, nor a clean bifurcation into two blocs, but a messier landscape of partial stacks, hedged dependencies and duplicated capacity. Efficiency will be the casualty; resilience, the stated justification. The language of the coming years will be the language of access: who is granted it, on what terms, and at what price. It is worth being clear-eyed about what that language describes. Treating computation as an instrument of power is not a distortion of the technology's trajectory but, increasingly, its main line. The question is no longer whether advanced AI is an object of statecraft. It is how much of the open system that produced it will survive being governed as one.