The Invisible Trillion · The Bill

The Compute War

Now the harder argument, because it’s the strongest case for building anyway, and because it’s true.

The claim is that the world is in a real contest over which kind of order runs the century — open and democratic, or authoritarian and state-controlled — that AI will be decisive in that contest, and that a world where authoritarian states hold the better AI is unacceptable on both economic and moral grounds. This is not a fig leaf. It’s the argument behind the CHIPS Act, the export controls, the federal money for domestic fabs, and much of the regulatory permissiveness that let the buildout proceed with so little federal friction. The people making it largely believe it, and some of them are reading classified assessments the rest of us aren’t.

The clearest evidence sits in Ukraine, which over four years became the most data-rich war in history. Both sides flew millions of drones; the character of combat changed under everyone’s feet. In one record month, Ukraine’s military reported that drones accounted for 96 percent of Russia’s battlefield casualties — unmanned systems outpacing every other weapon combined. Over the course of 2025, Ukraine’s defense minister, Mykhailo Fedorov, said drones killed or seriously wounded more than 240,000 Russian soldiers across some 820,000 video-confirmed strikes. And in the summer of 2025, for the first time in the history of warfare, a position was stormed and prisoners taken by ground robots and drones with no human soldier present — “the position was taken without a single shot being fired,” the Ukrainian commander told CNN.

Mass armor and massed infantry, the dominant forms of twentieth-century war, turned out to be vulnerable to swarms of machines costing a few hundred to a few thousand dollars apiece. And that kind of war runs on compute, at several layers: on the drone itself, for real-time perception and guidance under jamming — edge inference, the same architecture as a self-driving car; in field command posts, for sensor fusion and swarm coordination; in rear-area data centers, for training the perception models and grinding through the sensor data; and across the alliance, for shared intelligence and the rapid distribution of improvements.

The side with all four layers, integrated and at scale, has a decisive edge. That is a genuine argument for compute infrastructure, and it has nothing to do with consumer chatbots.

Which is exactly the catch. The military-relevant compute is a subset — heavy on edge silicon, focused training, and intelligence processing. It is not heavy on consumer inference, enterprise SaaS features, or slop generation, and those are most of what’s actually being built. Treating the whole buildout as a national-security project, when the bulk of it is consumer products subsidized to grab market share, is a category error — a convenient one for the operators, who get to drape commercial infrastructure in the flag, and an expensive one for the communities hosting facilities that mostly serve neither defense nor them.

Worse, the misallocation is self-defeating on its own terms: every dollar poured into hyperscale slop capacity is a dollar not spent on the drone silicon and ISR compute the actual contest requires. And the right architecture — distributed, edge-heavy, resilient — is more survivable in wartime than centralized hyperscale, which builds the enemy a concentrated target. A buildout sized for the security argument would look quite different from the one being built in its name.

The deeper point survives even if you set the war aside, and it is the one the “compute superiority” slogan tends to obscure. Strip the strategic argument down and what is actually scarce is not data centers. It is the ability to design and fabricate advanced chips — the fabs, the lithography, the packaging, the high-bandwidth memory, the supply chain that runs from a TSMC cleanroom in Taiwan through ASML in the Netherlands to the lines that turn finished wafers into parts. That capability is genuinely contested and genuinely worth a national effort to hold; it is what the CHIPS Act is actually about. A data center is a building full of those chips. The chips are the asset. The building is just where some of them happen to sit.

And the demand for the chips is durable in a way the demand for the buildings is not, because the future that needs them is mostly an edge future. The bet worth making is not that hundreds of millions of people will route their daily lives through frontier APIs in distant data centers. It is that AI moves into the physical world — robots in factories and, in time, homes; autonomous vehicles; on-device assistants; intelligence embedded in machines that act in real time and cannot wait on a round trip to Virginia. The timing is the real uncertainty there, not the direction. If that world arrives, it runs on staggering quantities of silicon, almost all of it sitting on the device and drawing the device’s own power.

So the honest version of the strategic case is narrower, and more defensible, than the one usually made: build the fabs, secure the supply chain, fund the research, keep the capacity to make the best chips in the world on the right side of the contest. That is the part of “compute superiority” worth defending. The hyperscale data center buildout is riding the same flag without the same warrant — sized for a centralized-inference present that the edge is already migrating away from, not for the distributed future the strategic argument itself points toward.

All 19 chapters of The Invisible Trillion · Library · All courses