The biggest bet ever made is being stacked in real time. Here is what's holding it up — and what it looks like if it comes down.
The headline is the spending; the real question is where it comes from. Only about half of the on-the-books build-out is paid in cash. The rest is borrowed or promised — and a far bigger sum sits off the balance sheet entirely, where it's hardest to see.
The Wall Street Journal went through the footnotes and found about $3.1 trillion more in commitments that never touch the balance sheet — roughly $1.9 trillion in locked-in purchases of chips and compute, plus $1.2 trillion in signed-but-unstarted data-centre leases. Alphabet alone jumped to $811 billion in a single quarter. A separate tally by Nikkei puts the hidden layer at ~$1.65 trillion — two ways of measuring the same iceberg, not to be added. Every dollar rests on the AI revenue turning up to pay for it.
The money goes first, in days. The people go last, over months. That lag is the whole story — the damage is nearly invisible right when it's least reversible.
It doesn't take a crash to start one. Enterprises decide the AI spend isn't paying its way and trim their budgets; one hyperscaler softens its build-out guidance. Because everything is leaning on a few names, that small pull is the first domino.
The AI names fall together and trading halts on the circuit-breakers. Because they're a third of the index, the loss is instantly inside pensions, 401(k)s and superannuation — reaching people who never chose these companies by name.
reaches everyone at onceOrders dry up, so the chipmakers fall. Data-centre construction freezes and the layoffs begin. A debt-financed 'neocloud' can't refinance — and the collateral behind its loans, the chips, has quietly lost most of its value.
lags the crash by weeksRetirements too close to recover. The pass-through crossing an ocean into super funds. The towns that bet on the plant — their power, their water, their bond ratings. And, as in 2008, the public quietly ends up holding the risk.
the slowest, widest damageIf the bet breaks, the wealth destroyed runs into the tens of trillions — several times the entire dot-com crash.
Three independent estimates on different bases (household wealth · US equity value · P/E reversion) — shown separately, not summed · Gopinath, Oliver Wyman, Dean Baker, 2025–26
Nobody can stop stacking. No company can stop building without ceding to a rival; no country can stop without "coming second to China." The fear is the fuel. The only real argument is what's left in the rubble.
Most of the railway track and the internet fibre survived their own busts and powered the era that followed. The bubble kills the speculation, not the technology — and the fire-sale finally makes the compute cheap enough to be useful.
A GPU is not dark fibre. Fibre sat in the ground for twenty years and cost nothing to keep; a chip ages out and burns power whether or not anyone uses it. "There is no future for it," the loudest bear argues — the capacity may not outlive the crash.
We don't get to know which yet. That's exactly why it's worth watching the gauges now.