“Way back” Tuesday, I noted that street forecasts for full-year hyper-scaler spending on the AI buildout were up more than $20 billion already in 2026, to $561 billion.

The next day, I said that capex estimate “will be revised higher, probably by a lot and beginning as soon as this week, with results from Alphabet and Amazon due.”

It’ll be a few days before it’s possible to tally company analysts’ updated projections, and Oracle won’t report until next month, but in the wake of Amazon’s $200 billion capex guide — which, you’re reminded, was 35% higher than expected — it’s safe to say we haven’t seen the last bump to full-year hyper-scaler spending forecasts.

We know, for example, that Meta’s in for at least $120 billion, and probably a lot more. Alphabet’s guide tipped $180 billion in outlays at the midpoint. And then there was Amazon’s $200 billion. That’s half a trillion right there, just between those three.

The figure shows you quarterly capex for the largest US tech companies ex-Apple. Between them, capex was just over $34 billion when Nvidia changed the world with one beat and raise. In the December quarter of 2025, that spending was up 270% at more than $126 billion.

Simply put: The AI revolution changed big-tech’s business model. And we’re talking here, about the most important, heaviest-weighted companies in the world, which’ve played Atlas to global equities for as long as some younger investors can remember.

“The flip from asset-light to asset-heavy business models [is a] major threat to the 2020s market leadership,” BofA’s Michael Hartnett remarked, in his latest, noting that “hyper-spender” capex could equate to more than 96% of those companies’ cash this year, up from 40% in 2023.

That’s not tenable. You can’t spend every free penny on data centers, bad as you might want to. Because where will the buybacks come from? That’s a rhetorical question. Here’s the answer:

The figure’s from Nomura’s Charlie McElligott and I encourage you to read the annotations: On the heels of a 400%+ increase in debt and loan supply in 2025 (think blockbuster offerings from Meta, Amazon and Oracle, as well as a huge debt package for the latter), the first month of 2026 saw more supply than the full-year totals for 2023 and 2o24.

Of course, that borrowing’s officially earmarked for data center building and leasing. But money’s fungible. Whether you’re burning through cash generated from operations or cash raised through debt issuance is irrelevant. Either way, you’re burning cash. “I’m spending your money on data centers” versus “I’m spending your money on buybacks I would’ve funded with the money I’m spending on data centers” is a distinction without a difference.

The bet — and this was underscored late Thursday by Andy Jassy — is that this is all almost certainly worth it given the world-changing nature of the technology. He might be (probably is) right, but make no mistake: It is a bet, which is to say it’s a gamble. As McElligott put it, “capex spend = growth opportunity versus cash burn / debt issuance.”

Demand for high grade supply’s voracious, and sans-Oracle, borrowers don’t get any more creditworthy than US big-tech. The supply overhang will demand a concession in the form of wider spreads, but it won’t be punitive.

Given that, it’s tempting to lose track of the fact that debt is… well, it’s debt. And when you issue debt to offset cash burn so you can — for example — continue to buy back stock even as you plow money into modern history’s biggest moonshot, you’re leveraging the balance sheet to fund buybacks.

That isn’t new. But it’s new at this scale, and as Hartnett wrote on Thursday evening, it means these companies may “no longer [have] the best balance sheets” going forward.