What’s at stake in AI’s trillion-dollar gamble

Wharton professor Jessica Wachter frames the boom with a hard accounting question: how fast must hyperscaler earnings grow to justify spending that her team estimates near $1.1 trillion through 2027? To break even by 2030 with cost of capital and a 15% return, she finds firms need about 2.7× productivity growth — possible, but a lot of growth squeezed into a few years.

Gary Gensler notes AI revenues this year are roughly $150–200 billion against hundreds of billions in spend. Fail to deliver, Wachter warns, and you risk missed interest payments and what she calls the largest misallocation of capital in history.

We’re all in this bet now — via pension funds, credit markets, and power bills tied to data-center deals.

We’re watching a parlay, not a single wager. Gensler’s frame is right: hyperscalers need massive revenues, the broader economy needs productivity gains, and frontier models must keep winning against cheaper good-enough alternatives. Miss any leg and the story breaks.

The financing turn raises the stakes. Free cash flow for the group is heading negative; Alphabet posted a roughly $5.9 billion free-cash deficit in the latest quarter — its first since Google’s 2004 IPO — as AI infrastructure ate nearly $120 billion in revenue. Morgan Stanley estimates more than half of $2.9 trillion in 2025–2028 data-center spend will come from external capital. That risk migrates into private credit, guarantees, and instruments sitting inside ordinary portfolios.

Meta’s Hyperion project in Louisiana shows how byzantine this gets: stakes transferred to Blue Owl, joint ventures, four-year leases matched to GPU lifetimes, residual-value guarantees, and utilities planning multiple gas plants. Communities worry ratepayers eat stranded power costs if demand or tenants vanish.

Useful AI that actually lifts productivity deserves the capital. The downside if this scales as financial engineering without earnings is familiar from past bubbles — plus a unique twist: the industry has tied the future of AI progress to the fortunes of giant campuses. I’m watching depreciation clocks on GPUs, cheaper models eating frontier margin, and whether public backlash on power and jobs becomes a binding constraint.

Hyperscalers named in the piece include Alphabet, Microsoft, Amazon, Meta, and Oracle partnering with OpenAI. Some projections put total AI capex above $5 trillion over four years. Columbia’s Stijn Van Nieuwerburgh estimates that building ~183 GW of AI compute by 2032 at ~$41B per GW implies required annual revenues near $3.7 trillion at a 10% return.

Surveyed executives largely report no productivity lift yet over three years, while expecting modest gains ahead — often via higher sales and fewer employees.