Analysis
Three of the four biggest AI spenders reported second-quarter earnings this week, and the stock reactions revealed something the raw capex totals alone don't: investors have started grading how the money is spent, not just how much of it there is. Microsoft got rewarded. Meta got punished. Amazon landed somewhere in between, on the strength of a number the other two couldn't match -- AWS actually growing faster because of the spending, not just growing forecasts around it.
For two years, the market treated almost any capex number from a hyperscaler as evidence of AI conviction, and stocks mostly went up regardless of the split between spending and monetization. That pattern broke this week. Meta's quarterly capex hit $31.1 billion, consuming nearly all of its operating cash flow and leaving free cash flow at just $784 million, down from levels that would normally comfortably fund buybacks and dividends alongside growth investment. Shares fell roughly 8-10% in after-hours trading even though revenue came in at $60.8 billion, above the roughly $60.2-60.3 billion analysts expected, with ad revenue up 27% year over year. The market didn't dispute Meta's core business was strong -- it disputed whether the AI spending attached to it was paying for itself yet.
Microsoft told the opposite story with a similar-sized spending number. Its capex is tied directly to Azure, where AI-driven consumption revenue grew 43% year over year -- a metric investors could draw a straight line from spending to growth. The stock's reward ranged from roughly 3% in the immediate after-hours print to 8-9% by the next trading session, among the strongest earnings-week moves of any hyperscaler. Amazon, meanwhile, posted its own record: revenue topped $200 billion in a single quarter for the first time, and AWS grew 37%, its fastest pace in 18 quarters, with a backlog that reached $496 billion. Amazon also raised its 2026 capex outlook to $220 billion, an increase of $20 billion driven largely by rising memory contract prices, and CEO Andy Jassy told investors the company still won't have enough capacity to meet demand in 2026.
“Its capex is tied directly to Azure, where AI-driven consumption revenue grew 43% year over year -- a metric investors could draw a straight line from spending to growth.”
Put the three side by side and the pattern is capacity utilization, not capacity size. Amazon and Microsoft could both point to demand outrunning supply -- backlog, consumption growth, capacity constraints extending into 2027 -- while Meta's spending is aimed more at long-horizon model research and infrastructure that doesn't show up in a comparably direct revenue line yet. Alphabet's own cloud earnings the same week reinforced the same read: capex tied to visible, near-term cloud demand is getting a pass that capex tied to model development alone is not.
For founders and allocators, the practical shift is in what kind of AI infrastructure story now clears the bar with public-market investors -- and by extension, with the growth-stage investors who price off those signals. A pitch built entirely around future model capability, without a near-term revenue line attached to the compute being bought, is now competing against hyperscalers that just got punished for exactly that pattern at a much larger scale. Portfolio companies selling into AWS, Azure or Google Cloud should also expect that capacity, not just price, remains a real constraint into 2027.
The bear case against reading too much into this week is that all three companies are still spending record amounts, and Amazon's own trailing 12-month free cash flow swung to negative $7.6 billion from positive $18.2 billion a year earlier -- a warning sign if it persists, not just Meta's problem. Memory and chip cost inflation is adding real pressure across every hyperscaler's capex line regardless of how the market reacted this particular week.
What to watch: whether Meta's spending shows up in a visible product or ad-performance improvement over the next two quarters before investors lose patience again, whether Amazon's capacity constraints ease or worsen through 2027, and whether Alphabet or any other hyperscaler faces the same discipline test Meta just failed when they report next.
