The American economy keeps growing, and in the first quarter most of that growth came from a single place. According to an estimate by Morgan Stanley, roughly three quarters of the rise in GDP came from AI-related investment, while, in the same reconstruction, investment unrelated to AI was contracting. One engine running.
Within six weeks, that engine entered the munition regime. On 2 June an executive order established a framework for early government access to frontier models; on 12 June the Department of Commerce barred foreign nationals from accessing the two most capable models on the market, and Anthropic, unable to verify nationality in real time, switched them off for everyone; on 14 July a Treasury-led clearinghouse began collecting and coordinating vulnerabilities discovered partly with frontier capabilities such as Mythos, in the software that runs the world. The asset that drives GDP and the asset the state treats as a weapon are the same asset.
A weapon, here, is not what fires. It is a capability a state can grant, restrict or withdraw according to its own interests: a technology that enters this regime stops being a product and becomes a munition. And when a technological capability is at once the engine of growth, the centre of gravity of the markets and a national security resource, a new regime is born: the munition economy. AI does not define this regime: it is the first technology to make it visible.
This is not a war economy. There, war reorganises production; here, a capability born in the market becomes at once productive infrastructure and strategic resource. Its value depends on economic profitability and on geopolitical salience, and this dual nature draws the incentives of firms, markets and the state together without any need for direction.
History offers partial combinations of these elements, but not this configuration: an asset born in the market that drives growth, dominates the indices and can be revoked at a distance through jurisdiction over the producer. The railways of the nineteenth century were the engine of investment, the centre of the stock exchanges and a military resource, but controlling them required possession of territory or infrastructure: no state could revoke their use at a distance through the producer. Gulf oil combined economics and strategy, but control ran through possession of the resource, the routes and the territories; here it also runs through jurisdiction over the producer, who can revoke access to a globally distributed capability without occupying any infrastructure on the user’s side. And the cryptographic products placed in the 1990s under Category XIII of the United States Munitions List had exactly this legal regime, but not this economic weight. Plus one variable no precedent possesses: software can be copied, while the compute that turns it into capability remains scarce.
The circuit
Securitisation, the turning of a technology into a matter of national security, is not an inference, it is a declared strategy: theorised in a bestseller by the man who was building it, codified in an executive order, spelt out by its executors. What nobody designed is the economic circuit that strategy set in motion. The political strategy is deliberate; the economic effects emerge on their own.
The June chain of events shows it at work: the alarm starts with Amazon, Anthropic’s partner and investor, reaches the Treasury, produces the export control, and closes a month later with the company’s capability absorbed into the apparatus of the state. Every actor follows its own incentive: Amazon flags what it considers a risk, the Treasury steps in, the company complies. The nineteen days of lost revenue and the subsequent bipartisan bill show this was no plan: the first intervention imposed a cost on the company, while the second, after an OpenAI model escaped its own testing environment, tries to turn the kill switch from improvisation into legal infrastructure. Clear doctrine, makeshift tools: the signature of a strategy without a plan. The kill switch did not devalue the asset; it certified it.
The transaction
The most explicit theorist of this settlement, Alex Karp, poses the question of the quid pro quo: given an unprecedented concentration of technological wealth, what does the public receive in return? The answer has materialised on its own. The public receives points of GDP, rising indices, pension fund returns: the seven big technology stocks of the AI race account for roughly a third of the S&P 500, and passive indexing spreads the exposure to anyone with a retirement plan, whether they chose it or not. It is an implicit transaction that disciplines itself: to contest the complex is to contest one’s own pension. And it is a regressive transaction: the payment comes as asset appreciation, and in the United States the richest ten per cent hold over ninety per cent of equities, the poorer half around one. Consent is distributed by breadth, benefit by depth.
But the transaction carries a clause nobody signed. The sector has entered the phase in which the race demands investment that even enormous internal cash flows cannot always cover: hyperscaler capex absorbs a growing share of revenue and is increasingly financed with equity and debt as well. This is no marginal observation: the Bank for International Settlements warns in its annual report that hyperscaler commitments are outpacing their earnings and free cash flow, and that the same massive bet, made by everyone simultaneously, is a recipe for collective overcommitment. Alphabet shows the dynamic in its sharpest form: in the second quarter, negative free cash flow of 5.9 billion dollars, capex nearly doubled, close to 70 billion net raised in equity and debt, while the stock fell despite record revenues. And the credit market has begun to price the clause: the cost of insuring against default by the hyperscalers, the Financial Times reports on Bloomberg data, has climbed to record levels. The technological republic, once realised, is not a project of strength but a position in which slowing down costs more than accelerating. And it binds its builder first: it can no longer afford to deflate the asset that keeps growth running.
The reverse side
The Chinese case confirms the category from the outside, in three ideas. Abundance of models: shut out of America’s frontier capability, Beijing responds by giving the weights away, and Kimi K3 rivals American models on specific tasks. Scarcity of compute: days after launch, Moonshot suspends new subscriptions, demand outruns the clusters, and the abundance of software collides with the scarcity of silicon. The conditioning of openness: Washington accuses Moonshot of distillation, an accusation disputed by experts, and in Chengdu twenty-one economies, China and the United States included, tie open source to security assurances at ministerial level for the first time. The two economies converge not on closure, but on the principle that access must be governed; the difference lies in which layer each can afford to control.
Europe has no munition economy, and that is at once its weakness and its freedom: it has not yet tied its growth to that race, and it retains the room to choose a different contract between technology and society. But it buys capability inside someone else’s regime, and June showed what that means. Dependence no longer consists in purchasing a foreign technology. It consists in organising one’s own economy around a capability that another state considers part of its strategic apparatus. At that point the risk is no longer the price. It is the availability.
The information and views set out in this article are those of the author and do not necessarily reflect the official opinion of the European institutions.