AI Capex Focus Shifts to Picking Winners Ahead of Q3 Earnings

From 'will it pay off?' to 'who gets paid?'

For most of the past two years, the main question about artificial intelligence on Wall Street was whether Big Tech's huge infrastructure budgets would ever earn a return. After the summer earnings season, some of the world's largest money managers say that question has changed. Investors now want to know which companies will still be growing profits once today's shortage of AI computing capacity eases.49

A Reuters analysis published in mid-August described the change. It credited a rally in AI-linked stocks during reporting season, along with results from Microsoft and Amazon, with persuading markets that demand for AI infrastructure was still strong.35 Cloud growth was speeding up and capacity remained tight. That shifted attention away from the size of the spending and toward which companies are best placed to benefit from it.10

Nearly two months later, with third-quarter reports coming at the end of October, that shift looks real but incomplete. Fund managers have stopped arguing about whether AI demand exists. The disagreement now is over cash flow, debt and how long investors will have to wait for returns.

What the big money is doing

The positioning is surprisingly consistent. Many large investors kept big semiconductor holdings even after a July sell-off, which followed doubts about AI spending and worries about rising Chinese competition. At the same time, they have been adding to the hyperscalers, the largest cloud providers.6 Brian Barbetta, co-head of the technology platform at Wellington Management, which oversees about $1.3 trillion, called these companies very large beneficiaries of the AI shift. He said his firm had recently increased its holdings in many of them.49

That buying goes against recent share-price performance. The four biggest AI spenders all trailed the Philadelphia Semiconductor Index, which rose 75%. They also trailed the neocloud providers CoreWeave, up about 50%, and Nebius, up more than 200%. These companies rent computing power and have charged high spot prices for scarce AI capacity.37

Richard Clode, who manages Janus Henderson's Bankers Investment Trust, explained the case for the large platforms. He expects the hyperscalers' profits and cash flow to start growing faster than their additional capex from late next year into 2028. Amazon is one of his fund's largest overweight positions.9 John Lamb of Capital Group, which manages about $3.6 trillion, said investors don't have to choose between chips and cloud. His view is to own both. He noted that a data centre usually takes 12 to 18 months to go from construction to producing revenue, and said the latest results showed the start of that turning point.610

The earnings evidence behind the optimism

The best evidence for the bulls comes from Alphabet. Google Cloud revenue rose 82% to $24.77 billion in the second quarter, which suggests data-centre spending is turning into sales to outside customers.16 The segment also became much more profitable. Operating income more than tripled to $8.8 billion, and operating margin rose from 20.7% to 35.6%.13 Alphabet's backlog is reported at $514 billion.13 An older figure of more than $460 billion as of the first quarter shows how fast it has grown.16

In a strategist's terms, this is what Clode means by "today's capex is tomorrow's sales."9 Venture firm a16z goes further. It argues that this boom is driven by earnings growth rather than rising valuations, and puts the combined cloud backlog of Microsoft, Google and Amazon at about $1.7 trillion.18 The wider market numbers point the same way. FactSet expects S&P 500 companies to post a third straight quarter of earnings growth above 25%, and second-quarter profit margins hit 17%, the highest since its data begins in 2009.12

Where the coverage diverges: the cash-flow problem

The more closely you look at cash, the weaker the "capex angst fades" story becomes. The Reuters piece itself included the key figure. On consensus estimates, hyperscalers will generate about $340 billion more in annual operating cash flow in 2027 than in 2025, while their capex rises by about $534 billion.4 That works out to roughly $1.57 of new spending for each extra dollar of operating cash flow.20

Coverage since August has focused on that gap. Bank of America expects the hyperscalers' combined free cash flow to fall from about +$180 billion in 2025 to roughly -$64 billion in 2026, then -$144 billion in 2027 and -$186 billion in 2028.20 Goldman Sachs estimates the group will spend about $750 billion this year against roughly $778 billion of operating cash flow. That leaves almost nothing to spare, and Goldman expects debt to fund about a third of capex.19

The strain shows up company by company:

• Alphabet: free cash flow turned negative in the second quarter, at -$5.9 billion. That was its first negative quarter since it went public in 2004. Debt rose to $98.2 billion from $23.6 billion, and the company has said it may raise equity through an at-the-market program.13
• Amazon: trailing twelve-month free cash flow was -$7.6 billion at the end of June, down from +$18.2 billion a year earlier.19
• Oracle: sold $19.9 billion of stock and took $11.4 billion in customer prepayments to cover its cash shortfall.19

Forecasts for total spending also differ, depending on which companies and costs are counted. Company guidance for Microsoft, Alphabet, Amazon and Meta adds up to $720 billion to $745 billion for 2026, and Alphabet alone is guiding to $195 billion to $205 billion.17 With Oracle included, a16z puts the total at about $780 billion.18 Bank of America's estimate is close to $860 billion, rising toward $1.2 trillion in 2027.20 However the total is measured, spending has roughly doubled from about $410 billion in 2025.14

Hyperscalers versus neoclouds

The clearest strategy to come out of the debate is to favor the large platforms over the specialists. Clode argues that Amazon, Microsoft and Google have longer-lasting advantages than neoclouds because of their scale and customer relationships.7 Noah Weisberger, chief U.S. equity strategist at BCA Research, warned that neoclouds could struggle once new capacity comes online and prices return to normal, given their heavier debt and current high pricing. He recommends being long hyperscalers and short neoclouds. Reuters later corrected the spelling of his surname.26

The financing data supports that caution. CoreWeave carries $35.6 billion of debt plus a similar amount in leases that have not yet started. Reuters reported that Blue Owl had trouble syndicating a $4 billion data-centre deal anchored by CoreWeave.19 Weisberger also pointed out that the hyperscalers' move to a more capital-heavy business model could hold down their valuations even if earnings stay strong.7

The hyperscalers are not all the same, either. In August, Microsoft traded at about 24.6 times forward earnings, the highest of the group, and Meta at 17.6, the lowest.9 A more recent comparison on a different measure puts Alphabet at about 16 times earnings and Meta at 26. That analysis argues Alphabet earns its premium in investor confidence because it sells AI capacity to outside customers, while Meta uses its computing power mainly for its own advertising business.16 The two sets of multiples measure different things, but both lead to the same point. Investors are increasingly paying for AI spending they can see turning into revenue.

The skeptics haven't left

Not every investor accepts the optimistic case. Alberto Conca, CIO of Swiss wealth manager LGF+ZEST, estimates that AI revenue would need to grow five to thirteen times to justify current spending plans.7 Valuation expert Aswath Damodaran found that aggregate US corporate net income rose 57% between the second quarters of 2025 and 2026, but dividends and buybacks grew only about 5% and 9%. He expects the hyperscalers' earnings and free cash flow to suffer until the spending pays off.11 A strategist cited by The Motley Fool describes a "capex pickle." If hyperscalers keep raising spending, their own shares may fall. If they cut back, the suppliers whose earnings depend on that spending could be hurt.15

Credit markets are showing similar concern. Prices on credit derivatives tied to hyperscalers and chipmakers have risen to levels that imply a higher chance of default. Bond investors are asking for more compensation to lend to the sector.19

The reading: a narrower trade, not a resolved debate

The best way to read this is that capex worry has narrowed rather than gone away. Investors now accept that AI demand is real, and Alphabet's cloud margins show the spending can earn money. What they no longer accept is spending without visible returns. Hyperscalers that showed strong revenue growth were rewarded last quarter, while some that raised capex guidance without matching growth were sold off.15

That makes the late-October reports a test of a specific claim: whether data centres built over the past 12 to 18 months are now producing revenue fast enough to offset falling free cash flow. Wells Fargo's Ohsung Kwon expects markets to start pricing in a 2027 peak in capex after third-quarter results.20 If that happens, the winners will look like the companies Barbetta described. They will have the broadest technology portfolios, the deepest customer relationships and the most control over their own infrastructure. As he put it, there will be fewer winners in the future than there are players today.7 For now, the large-cap money's bet is that the biggest spenders will also be the ones that survive.