AI capex is hiding in the footnotes

The real bill is still coming

Everybody loves a good capex headline: more GPUs, more servers, more data centers, more “we’re investing for the future” speeches. But the juicier part of the AI boom may be hiding in the fine print — specifically, the contractual commitments that say the spending is already locked in, even if the cash hasn’t left the building yet.

Morgan Stanley estimates that Nvidia and five hyperscalers — Alphabet, Amazon, Meta, Microsoft and Oracle — have piled up about $1 trillion in purchase commitments. That’s a big number, but the punchline is even bigger: under current accounting rules, a lot of these obligations don’t fully show up on the balance sheet until the goods or services are delivered.

Why investors are squinting at the footnotes

Moody’s has made a similar point, estimating roughly $969 billion in future hyperscaler lease commitments earlier this year, including about $662 billion tied to leases that hadn’t even started yet. And the Bank for International Settlements has basically looked at some of these AI financing structures and said, “This is debt with better branding.”

That’s the uncomfortable reality here:

• Reported debt tells only part of the story.
• EPS can look tidy while future obligations quietly stack up.
• The balance sheet may be missing the most expensive part of the party.

Earnings season’s sneaky plot twist

This is why the upcoming earnings parade matters. Alphabet, Microsoft and Meta will probably keep talking up AI investment like it’s a religion. Fine. But the smarter question is less “How much did you spend last quarter?” and more “How much have you already promised to spend for the next few years?”

Nikkei’s own analysis of filings from Alphabet, Amazon, Meta, Microsoft and Oracle estimated about $1.65 trillion of AI-related commitments off the balance sheet. That figure isn’t independently verified by the companies, but it reinforces the same basic message: the AI buildout is not just a spending spree, it’s a long-term tab.

Big picture: the AI arms race isn’t only about who has the biggest cloud or the fastest chip. It’s about who can keep funding the infrastructure without turning the footnotes into a horror story.