AI Capex Debate Shifts From Spending Fears to Picking Winners

From "will it pay?" to "who gets paid?"

For most of the past year, investors asked one question about artificial intelligence: would Big Tech's huge infrastructure spending ever earn its money back? After the summer earnings season, the question changed. In a Reuters analysis published in mid-August, large asset managers said the rally in AI-linked stocks during reporting season had moved attention away from whether the spending would pay off and toward which companies would produce returns over the long run.14 Results from Microsoft and Amazon persuaded the market that demand for AI infrastructure is still strong, cloud growth is speeding up, and capacity remains tight.11

The change matters because it reflects how money is actually being moved. Wellington Management, which oversees about $1.3 trillion, said the hyperscalers remain core holdings and that it has recently added to many of those positions.12 Its co-head of technology, Brian Barbetta, described the big cloud providers as likely to be among the largest winners from what he called an AI paradigm shift.16

Nearly two months later, with third-quarter reports due soon, that calm deserves a second look. Coverage since August suggests the capex worry did not go away. It turned into a narrower question about which companies are turning spending into revenue and which are simply borrowing to keep building.

What the summer earnings showed

The numbers behind the August optimism were real. Alphabet's second-quarter capital spending roughly doubled to $44.92 billion. Over the same period, Google Cloud revenue rose 82% to $24.77 billion, which suggests the data center buildout is turning into sales to outside customers.1 Microsoft's Azure grew 43% on quarterly capex of $35.8 billion, and the company reported $19.64 billion in free cash flow and a commercial remaining performance obligation of $678 billion.7 Microsoft has also guided Azure to about 45% constant-currency growth for its fiscal first quarter and expects to stay free-cash-flow positive while spending roughly $175 billion.7

Capital Group's John Lamb explained why these results felt like a turning point. Data centers typically take 12 to 18 months to go from construction to revenue, he noted, and the latest quarter is the first to show that inflection clearly.19 Janus Henderson's Richard Clode put it more simply: today's capex is tomorrow's sales. He said Amazon is one of his fund's largest overweight positions.15

The stock market's odd split

One detail in the Reuters reporting stands out. The four biggest AI spenders all trailed a 75% jump in the Philadelphia Semiconductor Index. They also trailed the neocloud providers CoreWeave, which rose about 50%, and Nebius, which gained more than 200%. These companies rent computing power and have benefited from high spot prices for scarce AI capacity.16 In other words, the market has mostly rewarded the companies receiving the money, not the ones spending it.

That pattern also explains why hyperscaler valuations look modest. Their multiples have shrunk this year and remain below post-pandemic highs. At the time of the Reuters piece, Microsoft traded at about 24.6 times forward earnings, the highest of the group, and Meta at about 17.6, the lowest.13 Later commentary uses different multiples. One October analysis puts Alphabet near 16 times earnings and Meta near 26.1 Another puts Microsoft at about 29 times.7 The gaps come mainly from forward versus trailing measures and from Alphabet's earnings being lifted by a roughly $99 billion unrealized equity gain.7 Still, the overall point holds: the biggest builders are not priced like bubble stocks.

Agreement on the cash-flow gap, disagreement on what it means

All the coverage agrees on the basic arithmetic. Reuters estimates that hyperscalers will produce about $340 billion more in annual operating cash flow in 2027 than in 2025, while their capex rises by roughly $534 billion.11 That works out to about $1.57 of new investment for every new dollar of cash flow.17

Where the coverage splits is on what that gap means. Clode expects these companies to grow profits and cash flow faster than their added capex from late next year into 2028.15 Bank of America is much less hopeful. It projects that aggregate free cash flow across the five largest hyperscalers will swing from about +$180 billion in 2025 to about -$64 billion this year, then fall to -$144 billion in 2027 and -$186 billion in 2028.5 A Columbia working paper finds that in 2026, combined capex at Oracle, Microsoft, Amazon, Meta and Alphabet will exceed their combined operating cash flow for the first time.5

Alphabet shows the tension most clearly. Its free cash flow turned negative in the second quarter, the first time since its 2004 IPO.5 It raised about $70 billion in equity and debt and suspended buybacks.7 Its capex guidance has also moved. Some October commentary still cites $175 billion to $185 billion.1 Other trackers report that two upward revisions lifted the range to $195 billion to $205 billion.8 The higher figure appears to be the current one, and the revisions matter on their own: in this cycle, guidance has mostly gone up.

In aggregate, estimates run from about $725 billion in 2026 commitments, up about 77% from 20258, to a projected $780 billion that venture firm a16z expects to top $1 trillion a year from 2027.3 Goldman Sachs and S&P Global analysts have floated $1.2 trillion to $1.3 trillion for 2027.4

The hurdle the winners must clear

Goldman's recent framework puts a number on the bar. To earn a 15% return on invested capital from their 2026–2027 AI compute spending, six large builders would need about $1.42 trillion in cumulative revenue between 2028 and 2030, or roughly $11.6 billion per gigawatt per year.9 Goldman notes that the three major cloud providers held about $1.69 trillion in backlog as of the second quarter, which gives some support to the monetization case.9 The firm also says hyperscalers are almost certainly already earning more than 15% on their existing capital.9 Others are less patient. LGF+ZEST's Alberto Conca estimates that AI monetization must grow five- to thirteen-fold to justify current spending plans.12

Accounting adds another layer. Valuation scholar Aswath Damodaran points out that hyperscalers capitalize these outlays and use longer depreciation schedules, which softens the near-term hit to earnings even though the cash still leaves the business.2 Microsoft's reported calendar-2026 capex of about $175 billion came in under the roughly $190 billion the market expected, partly because extending building useful lives from 15 to 25 years moved some leases out of reported capex.8 Investors who judge companies by earnings alone may underestimate how capital-heavy these businesses have become.

Why hyperscalers over neoclouds

The Reuters reporting's sharpest call concerns which kinds of companies win. Clode argues that companies controlling both computing capacity and the software layers that help customers run AI efficiently across models will have an edge. He says Amazon, Microsoft and Google hold more durable advantages than neoclouds because of their scale and customer relationships.15 BCA Research's Noah Weisberger, whose name Reuters later corrected13, warned that neoclouds could be exposed if new capacity arrives and prices normalize, given their heavier reliance on debt. He recommends a long-hyperscalers, short-neoclouds trade, while cautioning that a more capital-intensive model could limit hyperscaler valuations.15

Barbetta expects the field to narrow as the market matures, with fewer winners eventually than there are players today.15

The reading: the angst has shifted, not faded

On balance, the August thesis holds up on direction but overstates the calm. The market's patience now depends on companies showing returns. According to recent commentary, hyperscalers with strong revenue growth have been rewarded, while those raising capex guidance without matching growth have been sold off hard.4 Wells Fargo's Ohsung Kwon expects the market to begin pricing a 2027 capex peak after third-quarter earnings and the midterm elections.5 There is also a feedback loop. Hyperscaler spending supports earnings at many suppliers, so a slowdown that helps the builders' cash flow could hurt the rest of the market.4

For upcoming earnings, the useful test is not how much a company spends. It is whether cloud backlog and growth keep pace with that spending. On that measure, Microsoft and Alphabet currently have the strongest external-revenue evidence.7 Meta's case depends on AI improving its own ad business, with ad prices up 12% last quarter.1 The neoclouds' strong run looks the most exposed if capacity shortages ease. The winners will be the companies whose revenue starts growing faster than their spending, and the coming quarter should show which ones are on that path.