Every hyperscaler capex headline this year has answered the “how much” question.
On September 21, at the AI Infra Summit in Santa Clara, Arete Research founding partner and Senior Analyst Jim Fontanelli tried to answer the harder one: who’s paying for the difference between what’s being spent and what’s coming in.
His math: Alphabet, Amazon, Meta and Microsoft will spend roughly $1.8 trillion on capital investment through 2028, against about $1.2 trillion in underlying cash flow the four businesses are expected to generate over the same period — a $600 billion gap that has to be closed with debt, equity, or off-balance-sheet financing vehicles. Fontanelli delivered the figures in a session bluntly titled “A Market Outlook: Is AI in a Capex Bubble?” and noted the gap understates the full financing need, since it excludes specialized AI cloud providers, enterprises and governments building their own capacity on top of hyperscaler spend.
He flagged a second, compounding risk one level down the stack: neither OpenAI nor Anthropic currently holds an investment-grade credit rating — the threshold institutional lenders typically require before buying debt freely — even though both companies’ compute commitments underpin much of the capacity now being built to serve them.
The figures track with, but sharpen other recent estimates.
CreditSights puts 2026 capex for the top five hyperscalers at $602 billion, a 36% jump over 2025, and J.P. Morgan’s John Servidea has called AI financing “the biggest secular theme” of his professional lifetime.
Bottom line: The AI capex debate has largely been fought over the size of the number. Fontanelli’s contribution reframes it as a financing-structure question — and puts a specific, sourced figure on the part of the buildout that can no longer be self-funded, which is the detail credit markets, not headline writers, will be pricing in next.
