The Capex Supercycle: Why Investment Booms Build Winners, Then Gluts

By the most-cited estimates, the world's largest cloud companies will spend on the order of $700 billion on capital projects in 2026 — roughly a 36% jump over the prior year, with about three-quarters of it going straight into AI infrastructure: servers, accelerators, power, and the buildings to hold them. The demand behind that number is real. Compute is genuinely scarce, and the spending is a rational response to it. What's worth remembering is that almost every capital supercycle in history started exactly this way — with real demand, rational spenders, and a shortage that seemed like it would never end.

The pattern that follows is old enough to have a shape. A capex boom is a race: once demand outruns supply, every participant invests to capture it, and because capacity takes years to build, everyone builds at once. Each decision is justified at the moment it's made. The trouble is that the capacity all arrives together — often just as demand growth cools — and a shortage flips into a glut. Even now, rating agencies note that the largest spenders are deliberately pacing their builds to limit overbuild risk. That tells you the risk is already on their minds.

Why the winners change. During the build phase, the companies that sell the shovels win — the equipment makers, the component suppliers, the names with pricing power over a scarce input. Their revenue and margins expand because buyers are capacity-constrained and price-insensitive. A glut compresses exactly those margins. When capacity catches up, the scarce input becomes abundant, pricing power moves from seller to buyer, and the stocks that led the boom on the way up often lead it on the way down. The demand story can still be true — the internet did reshape the economy, and the fiber did eventually get used — while the capital that financed it is destroyed on the way to becoming useful.

Three earlier versions of the same movie. The fiber build of the late 1990s is the cleanest analogue: after the 1996 Telecom Act, carriers poured over half a trillion dollars into long-haul networks, and for years afterward the vast majority of that fiber sat dark — installed but unlit — while WorldCom and Global Crossing collapsed into what were then among the largest bankruptcies on record. Decades earlier, American railroads were overbuilt into the panics of 1873 and 1893; by 1894, roughly a quarter of the country's rail mileage had passed through receivership. More recently, the 2010s shale boom pumped capital into U.S. oil production until oversupply helped crack crude from around $100 to under $30 in 2014–2016, taking a wave of drillers with it. In each case the resource was real, the demand was real, and the capital cycle still turned.

What this looks like inside our own reads. A capex boom doesn't just concentrate spending; it concentrates the tape. On September 3, our engine's sealed morning reads had only 34 of 100 forward views pointing up, with 63 landing neutral — and the day played out accordingly: the AI-server and mega-cap names led while energy and cyclical names fell, even as the index closed higher. That is the market-level fingerprint of a supercycle — a narrow group of beneficiaries carrying a broad index — and some version of it shows up in every one of these episodes well before the capital cycle turns.

None of this dates the ending. The AI build-out may run for years, and the compute shortage today is not in question. The useful discipline is simply to know which phase of the cycle a given story belongs to — the scarcity phase that rewards the sellers, or the abundance phase that punishes them — rather than to assume the first phase is permanent because the demand is genuine. Booms are large precisely because the demand is real. That is what makes them supercycles; it is not what makes them safe.