Microsoft, Alphabet, Amazon, and Meta will spend somewhere between $650 billion and $700 billion on AI infrastructure this year. Gartner projects worldwide AI spending at $2.52 trillion for 2026. These numbers have become so large they've lost the ability to mean anything. A billion dollars used to be noteworthy. Six hundred billion barely makes it past the earnings call.

The question that keeps nagging — the one the earnings presentations spend entire segments avoiding — is what, exactly, all of this money is buying.

The honest answer: cloud growth, mostly. Microsoft's Azure grew 40% year over year in Q2, with AI contributing about 16 percentage points of that growth. Google Cloud hit $17.7 billion in Q4 2025, up 48%. Those are real numbers. Real revenue. Real customers signing real contracts. However — and this is where the narrative curdles — the total direct AI revenue across the industry last year was roughly $51 billion against $527 billion in spending. That is a gap you could park a civilisation in.

An MIT study found that up to 95% of firms investing in AI have not yet seen tangible returns. Only 14% of CFOs report measurable ROI. Despite this, 68% of CEOs plan to increase spending again next year. The logic is circular: we must spend because our competitors are spending, and our competitors are spending because we must spend. Nobody wants to be the one who blinked and missed the platform shift.

I keep returning to the comparison with OpenAI's revenue panic. A company that raised hundreds of billions, has 800 million weekly users, and still can't make the economics work without plastering ads across a product its CEO called "uniquely unsettling" to monetise that way. The unit economics are a warning sign for the entire sector, not just one company.

What frustrates me is that the useful stuff gets buried. Barclays cut £2 billion through AI-driven efficiency programmes. Anthropic just embedded Claude into Excel and PowerPoint, which is boring and practical and probably where the actual value lives — in incremental productivity gains that never make investor presentations exciting. The flashy demos get the funding. The spreadsheet automation gets the results.

Analyst projections warn that Big Tech free cash flow could drop as much as 90% in 2026 as capex outpaces revenue. Ninety percent. That is not a rounding error. That's a structural choice to defer profitability on the bet that whoever builds the most data centres fastest wins the next decade. Maybe they're right. The companies making this bet have been right before — about cloud, about mobile, about search. But they've also been wrong before, about the metaverse and crypto and social audio and a dozen other things that consumed billions before quietly disappearing from earnings calls.

The money is real. The infrastructure is real. The revenue is not — not yet, not at the scale the spending demands.

Sources: