You pay $20 a month for ChatGPT or Claude. You assume — reasonably — that somebody with an MBA and a spreadsheet calculated whether the infrastructure behind that subscription makes financial sense. That the server farms, the cooling towers, the miles of fiber optic cable all pencil out to a rational business.

They don't. Not even close.

The numbers that don't add up

By the second week of April 2026, the aggregate picture had crystallized — and it was ugly. Amazon and Meta dropped their bombshells in April; Google and Microsoft had already shown their cards earlier in the year. Combined, the four biggest cloud companies committed somewhere between $610 billion and $690 billion in infrastructure spending for 2026 alone. Their combined AI-specific revenue? Roughly $76 billion — and that's being generous by counting all cloud revenue, not just AI. The actual ratio: the industry spends about eight dollars on concrete and copper for every dollar it earns from AI.

On April 9, Amazon CEO Andy Jassy published his annual shareholder letter defending a $200 billion capex — capital expenditure, meaning money spent on physical stuff like buildings and hardware. His exact words: "We're not investing approximately $200 billion in capex in 2026 on a hunch." That's the kind of sentence you write when your board is asking uncomfortable questions. The same day, Meta expanded its CoreWeave deal to $21 billion, pushing Meta's total 2026 infrastructure budget to $115–135 billion. Google disclosed $175–185 billion back on February 4, and its CEO admitted even that "still won't be enough." Microsoft CFO Amy Hood outlined $120 billion in planned 2026 capex during the company's January 29 earnings call, alongside an $80 billion backlog of Azure orders it literally cannot fulfill.

How we got here (spoiler: panic)

The GPU shortage of 2024 broke everyone's brains. When you can't buy chips, you panic-buy everything else: land, power contracts, construction crews. The cloud giants — Amazon, Microsoft, Google, Meta — locked themselves into multi-year power purchase agreements and construction contracts they can't cancel. They signed these deals with the urgency of someone buying toilet paper in March 2020. Now inference costs — the price of actually running an AI model — drop roughly 10x with each new chip generation, but the buildings those chips sit in are already paid for. You bought a mansion because you couldn't find an apartment.

And here's the twist nobody's discussing at Davos panels: the bottleneck isn't chips anymore. It's electricity.

US datacenters now draw about 41 gigawatts — consuming 4.4% of all US electricity, more than most individual states. Planned builds would double that by 2028. But connecting new facilities to the power grid takes 5–7 years — roughly the same timeline as "we'll figure out monetization later." In February, Bloomberg reported that US datacenter construction actually fell for the first time since 2020 — not because demand dropped, but because nobody can get permits or power.

Virginia, home to the world's densest datacenter cluster, has effectively halted new permits in several counties. Morgan Stanley projects a 49-gigawatt generation shortfall by 2028. You can't run a $690 billion AI empire on vibes and extension cords.

The ghost of 2001

The telecom parallel is the one nobody in Silicon Valley wants to discuss, probably because half of them were in middle school when it happened. In 2000, carriers laid fiber at 3x the rate demand justified, assuming internet traffic would catch up. Sixty percent of that fiber went dark for a decade. The internet did eventually justify the cables — but the companies that laid them went bankrupt before it mattered.

Morgan Stanley analyst Todd Castagno noted in February that the AI boom's capex-to-sales ratio has hit 34% — exceeding the dot-com peak of 32%. His assessment: "The AI buildout has become so large that it no longer supports paying any price for the companies driving it." That's analyst-speak for "this is insane and everyone knows it."

Alphabet posted its first $400-billion-revenue year on February 4 — and the stock dropped 7% anyway. Not because the earnings were bad. Because investors saw the capex number and flinched. When Wall Street punishes a $400 billion revenue year, the math has left the building.

What this means for your wallet

If you're choosing between AI tools and platforms, stop comparing features for five minutes and look at the balance sheets. Vendor survival now depends on filling those server racks. The cheapest model won't be the best model — it'll be the one whose provider most desperately needs to amortize a half-empty datacenter. When you see aggressive pricing, free tiers that seem too generous, and "strategic partnerships" announced with breathless press releases — that's not generosity. That's a company trying to justify the building it already paid for.

Jassy says this isn't a hunch. Sure. But the last time an industry spent $2 trillion on infrastructure while revenues lagged 8-to-1, it didn't end with a shareholder letter. It ended with a CFO on an earnings call admitting utilization was 40%, and a stock chart that looked like a cat knocked it off the table.