AI's $1 trillion build-out: One economist's case that the math doesn't close

Microsoft, Alphabet, Amazon, and Meta have projected roughly $700 billion this year, and about $1 trillion in 2027, to build the data centers behind the next generation of AI. Steve Hanke, a Johns Hopkins economist who served on President Reagan's Council of Economic Advisers, says that math is the part the "cheap AI for everyone" story cannot answer.

Hanke, speaking by email to Business Insider, called the belief that AI will be virtually free "delusional and dumb" and "based on a disconnect from reality, as well as a good dose of idiotic economic reasoning." His thesis is that the cost of providing AI is bound by physical resources: water for cooling, electricity for compute, and the graphics chips at the heart of every model. Those inputs are scarce, geographically fixed, and bought at market rates. Software, by contrast, can be copied at near-zero cost. Hanke says conflating the two is the error behind most AI abundance claims.

"Businesses will not be firing everybody and replacing them with AI," Hanke told Business Insider, because in many cases human labor is cheaper than running the AI systems meant to replace them. "The cost of scarce resources that are gobbled up by AI will decide how far the AI revolution goes."

The capex commitments make that constraint concrete. Microsoft has guided to roughly $190 billion in calendar 2026 capex, a figure the company disclosed in its FY26 Q3 earnings statement. In the FY26 Q4 update, Microsoft shifted that number to approximately $175 billion after moving some financing to operating leases, an accounting change that keeps the same hardware spend on the books but classifies it differently. The headline capex figure is already moving, and the first signal is Microsoft itself. Alphabet has guided to up to $185 billion in 2026 capex, Amazon to about $200 billion, and Meta to a $125 billion to $145 billion range, according to Business Insider's synthesis of the four firms' earnings calls.

The long arc is heavier. Goldman Sachs projects a combined $5.3 trillion of capex across the four largest hyperscalers from FY2025 to FY2030. That is not realized spend. It is sell-side research projecting how the next five years of data-center build-out will land if the current trajectory holds. The trajectory has to clear the resource bill before any of it shows up as cheaper AI for the buyer.

Hanke also called many AI visionaries "charlatans and hucksters" for comparing AI to ordinary software. The distinction he draws is marginal cost: shipping a software update costs almost nothing, while running a frontier model costs whatever the local power and water rates happen to be, plus the depreciation on the chips themselves. That is why the resource argument is the load-bearing one. The technology story can keep compounding; the resource story cannot, because the inputs are physical and the prices are set by utilities and chip fabs, not by software teams.

The foil is Elon Musk, paraphrased in the same Business Insider piece as predicting AI will take most jobs and pushing for a "universal high income" of mailed checks to redistribute the gains. Musk argues that AI productivity will be large enough to redistribute, while Hanke argues the resource bill has to clear before the productivity gains arrive, and may not clear at all. Both treat AI as economically weighty; they disagree on whether the bill gets paid or the gains arrive first.

The test is the next earnings cycle. If Microsoft, Alphabet, Amazon, and Meta keep raising the capex line while the cost per token of running a frontier model keeps falling, the resource-constraint story loses. If capex guidance keeps shifting up while inference costs plateau, or if operating-lease moves start hiding what is effectively the same spend, the market has to sit with Hanke's resource math. Microsoft has already moved once. Goldman says there are five more years of this. The next 24 months of capex disclosures will show whether the resource math closes or the abundance story has to.