The AI Spending Bill Is Coming Due.... And Wall Street Wants Receipts

 

  I've been reading earnings transcripts for a long time, and there's a particular kind of silence that shows up when an analyst asks a CFO a question the company doesn't want to answer directly. That silence is going to be loud this July.

Here's the number that should stop you: hyperscalers are on pace to spend somewhere between $650 and $725 billion in 2026 on AI infrastructure alone. Data centers, chips, power contracts, cooling systems, the works. Microsoft is guiding toward roughly $190 billion. Amazon is close to $200 billion. Alphabet sits in the $175–190 billion range, and Meta is somewhere between $115 and $145 billion. Every one of those numbers is up 50 to 90 percent year over year.

That's not a rounding error in a corporate budget. That's four companies deciding, almost simultaneously, to bet a meaningful share of their market cap on a technology whose commercial payoff is still mostly theoretical.

Why This Earnings Season Actually Matters

I say "actually matters" because tech earnings seasons get called pivotal roughly every quarter, and most of the time nothing pivots. This one might be different, and here's my reasoning.

Capex of this size doesn't sit quietly on a balance sheet. It shows up in depreciation schedules, in free cash flow, in the margin story that every one of these companies has been telling investors for two years. Jensen Huang has said publicly that a single 1-gigawatt AI facility can run $55 billion once you account for full infrastructure, power, cooling, networking, not just the chips. Multiply that by the gigawatts these companies are chasing and you start to understand why analysts are getting nervous rather than excited.

The uncomfortable question underneath all of this: how much of that spend is actually generating revenue today, versus how much is a bet on revenue that may or may not show up in two or three years?

The Gap Between the Story and the Spreadsheet

Cloud revenue is genuinely strong right now. Google Cloud's growth numbers, for instance, look good on paper. But separating "AI revenue" from "cloud revenue that would have grown anyway" is close to impossible from the outside, and I suspect it's not much easier from the inside either.

What we do know is that depreciation and energy costs tied to this buildout are already leaning on near-term margins. That's the part of the story that doesn't make it into the keynote slides. Front-loaded spending with back-loaded payoff is a familiar pattern, the fiber-optic buildout of the dot-com era followed exactly this shape, and a lot of companies that built the infrastructure didn't survive long enough to benefit from it. The internet still got built. The shareholders who funded it didn't all get their money back.

This cycle has one structural advantage the dot-com era didn't: most of this spending is coming out of cash flow, not debt. That buys these companies patience. It doesn't guarantee a return.

There's real disagreement among people who follow this closely. Ed Zitron has been vocal that the unit economics of running these models, inference costs against what customers are actually paying, don't add up yet for a lot of the foundational model providers. Others argue we're simply early, the same way cloud computing looked expensive and unproven for years before it became the most profitable business line most of these companies have.

I don't think that debate gets settled in one earnings cycle. But the cycle will produce evidence either way, and that's what I'll be reading for.

What I'm Actually Watching For

A few specific things, beyond the headline beat-or-miss:

Whether AI revenue is separable from cloud revenue at all. If a company can't or won't break this out with any specificity, that's information in itself.

Capex guidance for the back half of the year. A pullback signals the companies themselves are getting nervous about utilization. An unchecked increase signals something closer to an arms race than a business plan. Meta's exploration of monetizing excess compute capacity is the kind of detail that tells you more than the topline number does.

Free cash flow trends. Capex as a share of operating cash flow has been climbing toward levels that would have been unthinkable five years ago. At some point that ratio has to plateau, or something else has to bend, margins, growth expectations, or stock prices.

What the suppliers say....Nvidia,Broadcom, and TSMC earnings will tell you more about real demand than anything a hyperscaler CFO says on a call, because supplier order books don't have investor relations teams shaping the message.

The Bigger Picture, Briefly

None of this is really about whether AI is "real." It obviously is. The question earnings season is actually testing is narrower and less exciting: whether the pace of spending is matched by the pace of monetization, or whether the gap between the two keeps widening until something forces a correction, either in guidance, in valuations, or in strategy.

Power availability is becoming a genuine constraint in some regions, which is its own story worth watching separately. And the eventual winners of this buildout may end up with a scale advantage that's very hard for anyone else to challenge, which raises its own set of questions about concentration that go beyond quarterly earnings.

For now, my advice to anyone watching this space is unglamorous: don't trade on the headline beat. Read the guidance language closely, watch what management doesn't say, and pay attention to the suppliers before the hyperscalers even report. That's usually where the real signal shows up first.