For three years the market funded the AI buildout on faith. On July 23 it stopped. The S&P 500 fell 1.2% and a gauge of megacaps had its worst day since the April 2025 tariff meltdown — not because AI stopped working, but because investors finally put the capex line and the revenue line on the same page.
The numbers as of July 23, 2026
| Metric | Reading |
|---|---|
| Combined 2026 hyperscaler capex | >$452B |
| Top-5 estimate, raised | ~$750B (+67% YoY) |
| Third consecutive year of growth | >60% |
| Alphabet Q1 free cash flow | −47% to ~$10B |
| Amazon trailing FCF | −95% |
| Alphabet / Amazon on the day | −5-6% / −4% |
Capex as a share of sales
| Company | 2026 capex / revenue |
|---|---|
| Oracle | 86% |
| Meta | 54% |
| Microsoft | 47% |
| Alphabet | 46% |
| Amazon | 25% |
Why it moved
Nothing broke. No hyperscaler cut guidance — the opposite. Alphabet's Q2 capex hit $44.9B and management lifted the full-year outlook again. Amazon has guided to roughly $200B, Meta to as much as $145B, Microsoft above $120B. That was the problem. Every raise widened the gap between what these companies are spending and what AI is currently earning them.
The mechanism
Capex is not an expense — it lands on the balance sheet and drains cash, then hits earnings later as depreciation. That's why the damage shows up in free cash flow first. Alphabet's FCF falling 47% and Amazon's collapsing 95% is the same event viewed from the cash side: these businesses are converting enormous operating profit into buildings, power contracts and silicon. If the revenue arrives, the depreciation is absorbed and nobody remembers. If it's late, the depreciation still arrives on schedule and compresses margins for years.
When Oracle spends 86 cents of every revenue dollar on capex, it is no longer a software company in any meaningful financial sense. It is a leveraged infrastructure developer with a software brand attached.
The risk on the other side
I want to be honest about the bear case being wrong. Under-building would be far more expensive than over-building if AI demand is real — the hyperscalers lived through the 2010s learning that capacity constraints cost you the market. Every one of these management teams has said, in some form, that the risk of spending too little exceeds the risk of spending too much. They may be right. Amazon's AWS capex looked reckless in 2013 too.
The difference is that AWS was already selling everything it built. That's the test today's numbers don't yet pass.
My take
I'm not selling the theme, I'm changing what I'll pay for it. A capex cycle this size transfers value from the spenders to the suppliers — the power, cooling, grid and memory names get paid whether or not the buildout earns its return. I'd rather own the toll booth than the casino during the construction phase, and revisit the hyperscalers when free cash flow inflects rather than when guidance does.
What I'd watch is not the capex number but the ratio: capex growth against cloud revenue growth, quarter by quarter. As long as capex compounds at 60%+ and revenue doesn't, multiple compression is arithmetic, not sentiment.
Bottom line: the market didn't lose faith in AI on July 23, it lost patience with the funding gap. Capex above $452B against collapsing free cash flow is a timing problem, and timing problems get repriced.
This is analysis, not investment advice.