What the Numbers Say About the Biggest Bet in History

There's a lazy way to argue about whether AI is a bubble: techies squint at exponential curves, economists shrug about a modest bump to GDP a decade out, and everyone talks past everyone else. But there's a better way to settle it, because five companies have already made their answer public — not in words, but in money.

Call them GOMMA: Google, Oracle, Microsoft, Meta, and Amazon. (Nvidia doesn't count — Nvidia is the bookmaker, not a bettor.) Together, these five have committed roughly $2.5 trillion to building the physical infrastructure of AI — the datacenters, the chips, the power. Half of it is already spent. In raw dollars, it's the largest capital expenditure in history. As a share of the economy, the only real precedent is the 19th-century railroad boom — except the railroads took twenty years to build out, and GOMMA has done this in four.

That's the part worth sitting with. OpenAI and Anthropic can talk about AGI all they like — talk is free. GOMMA can't. They've bought the hardware. And hardware doesn't wait around to see if the thesis is right; it just depreciates, on schedule, whether or not the future arrives on time.

Here's where the math gets uncomfortable. AI chips lose value fast — about 25% a year. A dollar of chips bought today is worth 75 cents next year, and by 2029 it's down to a third of its original value. Run the numbers on GOMMA's $2.5 trillion outlay and by the time the buildout is finished, roughly $800 billion of it will already have evaporated, leaving a working pile worth about $1.7 trillion. That pile melts at around $420 billion a year just from depreciation. Add the 15% annual return that shareholders reasonably expect on capital this size, and the bill comes to about $670 billion a year — in pure profit, not revenue. Profit that has to come from somewhere beyond covering server costs and salaries: money that AI customers pay over and above what it costs to run the machines.

Where could $670 billion a year possibly come from?

The obvious answer: companies renting AI instead of paying human salaries, with GOMMA (and its dependents, OpenAI and Anthropic) taking a cut of the savings. Historically, technology vendors capture something like a third of the value they create — and a third is a generous estimate here, not a conservative one. But even a full third of what every software engineer in America currently earns only adds up to about $150 billion a year. Not remotely close to $670 billion.

So the bet has to be much bigger than "AI replaces programmers." It has to be a bet on AI eating a meaningful slice of the entire labor market. Once you total up every job that AI could plausibly augment — reasonable estimates put that at 59% of U.S. labor, representing about $8.5 trillion in wages — the math starts to work. GOMMA would only need to capture about 8% of those wages (which, at a one-third capture rate, means the underlying productivity gains would need to run about 24% higher than today) to hit their number.

Twenty-four percent. That's not "AI helps a bit at the margins." That's roughly fifty times the productivity impact that mainstream economists like Daron Acemoglu have projected for AI over the coming decade — and GOMMA is betting it happens within four years, not ten.

This is the tell. Frontier labs betting big on AGI is old news — it's their whole identity. But finance departments at Google and Amazon running the actual capital-return math and concluding "yes, commit the quarter-trillion" is a very different, much harder thing to explain away as hype. These are not visionaries; these are the people whose job is to be skeptical of visionaries. And they've signed off.

There's one more wrinkle that makes this bet unusually unforgiving: there's no comfortable middle outcome. Open-weight AI models are already selling at roughly a twentieth the price of frontier models, trailing the leaders by only months. That means GOMMA doesn't get to settle into a nice, stable, moderately-profitable future. Either AGI actually arrives and transforms the economy enough to justify the spend, or someone manages to legally lock in the advantage by restricting open-weight competition, or — the historical default — the gap closes and the profits get competed away regardless of how good the technology gets. That's exactly what happened to fiber-optic investors in the late '90s: they were completely right that the internet was coming, and they still lost everything, because being right about the technology and being the one who profits from it are two different bets.

So take your pick of endings. If AGI shows up on schedule, the world gets rewired and GOMMA gets paid. If it doesn't, we're not looking at a gentle cooldown — we're looking at a bust that makes the dot-com crash look like a fire drill. What we're not looking at is the boring middle scenario where AI quietly becomes a nice, modest addition to the economy and everyone goes home satisfied. The people writing the checks have already ruled that one out.

Distilled from LessWrong