For hyperscalers like Amazon, how much capex is too much?

Investors have given the world’s tech giants a long leash when it comes to dumping hundreds of billions of dollars into their respective armadas of AI data centers.

We might be about to find out how much leash is left. 

Massive capex spenderAmazon is due to report results and refine its AI investment plans Thursday. Fellow hyperscaler Alphabet’s earnings land after the bell Wednesday .

How the market reacts to those numbers may go a long way toward resolving a key question that’s emerged in recent weeks: are investors growing impatient with airy executive assurances about AI dominance to come as tech companies go all in with billions of dollars of bets on the future? 

“The narrative to me is now: if you’re going to spend all this on capex, we want to see it come through in sales, and we want to see it come through now,” said Pat Tschosik, an analyst at Florida-based market research firm Ned Davis Research. “We better start to see some signs that AI is having an impact.” 

The idea that big capex outlays now need to be paired with concrete evidence of growth — positive sales surprises, guidance hikes, or forecasts for fatter margins or rising cash flows — has been gathering steam after hyperscalers Microsoft and Meta almost simultaneously announced much larger-than-expected capex in reports last week and received distinctly different feedback from the market. 

Microsoft tumbled 10%, its worst day since Covid hit back in 2020, with losses ballooning out to 15% in the following days. Meta rose 10.4% the next day — though it’s given back most of those gains as tech stocks have pulled back recently. 

Amazon and Alphabet’s quarterly results could shed even more light on the dynamic, depending on the matrix of capital expenditure, sales growth, and profit margins they present. 

We spoke with analysts and rifled through Wall Street reports to zero in on how investors and analysts now seem to be weighing such factors, in the hopes of identifying the single magic data point to watch. 

But at the end of the day, the assessment will likely vary by company, says Mark Moerdler, a longtime software analyst at Bernstein Research.

Take Microsoft, for instance. The slide that followed its fiscal Q2 numbers last week seemed to stem from the combination of a higher-than-expected capex figure alongside sales at its Azure cloud computing unit that analysts saw as soft. 

“In the case of Microsoft, the Street is looking at the capex and trying to gain comfort in both the revenue that’s going to be created, and the margin of that revenue, and the sustainability of that revenue,” Moerdler said. “What they want to see is alignment with spending and then converting that spending relatively quickly into revenue.”

In other words, what Meta did. 

Mark Zuckerberg’s advertising and hyperscaling behemoth is the most obvious comp for Microsoft, as it reported massive plans to boost capital spending in the coming year at precisely the same time last week. 

But unlike Microsoft, its shares surged, a response some attributed to the fact that Meta sharply lifted its sales guidance, forecasting growth of roughly 30% for the current quarter versus the roughly 20% analysts had been expecting. (Microsoft’s forecast for the current quarter was basically in line with analysts’ estimates.)

“Meta is increasingly demonstrating that revenue strength is offering significant returns to support elevated investments in both the core advertising platform and also the growing longer-dated AI ambitions,” analysts at Deutsche Bank wrote of Meta’s numbers.

Others on the Street suggest that some of the most important benchmarks to watch in assessing capex aren’t in the financials produced by the companies themselves, but rather by their customers. 

“The thing that I’m looking at very closely is enterprise AI demand. And what I mean by that is actual enterprises, not Microsoft, Google, OpenAI, Anthropic,” said Rishi Jaluria, a software analyst at RBC Capital. 

Virtually all companies have some sort of AI plan in place, Jaluria said. But most companies are very tentative, focusing on pilot programs or proof-of-concept AI experiments in relatively isolated contexts, such as IT help desk applications or assistance with coding. 

“It’s not widely deployed,” he said. “And I think the question that we have to ask is OK, well, what causes that leg up in enterprise AI demand? That’s what we’re looking for.”

As Jaluria’s comments suggest, the market also seems to be getting slightly jittery about the idea that some hyperscalers — Microsoft and Oracle among them — have been making their investments based, in part, on growing demand from privately held artificial intelligence companies like OpenAI. 

One of the big disclosures in Microsoft’s earnings report was that some 45% of Microsoft’s remaining performance obligations — a measure of the backlog of demand for its services — was tied to OpenAI. 

Sam Altman’s firm is reportedly burning cash at an alarming rate, making OpenAI commitments to spend hundreds of billions of dollars on Microsoft’s data center services for years to come seem less than ironclad. 

Some think that, alone, might have been a key differentiator between the market’s reaction to Meta and Microsoft’s results, with investors clearly taking more comfort in Meta’s plans to invest for the future. 

“Meta’s growth is conceptually more predictable,” said Vivek Arya, a Bank of America research analyst who covers semiconductor companies but keeps a close eye on the hyperscalers that buy significant amounts of the computer chips sold by such companies. “They’re not depending on this private entity.” 

Amazon and Alphabet’s numbers could provide a test of that theory. 

Amazon’s ties to OpenAI have grown in recent months, with the AI startup signing a partnership in which it plans to spend up to $38 billion to buy computing power from Amazon’s Web Services division. Reports have also emerged that Amazon is considering investing in OpenAI’s latest round of funding to the tune of $50 billion. Amazon has a large deal with Anthropic as well, in which the startup will deploy its Claude assistant using AWS servers, where Amazon is adding significant capacity. 

Those deals should help fill up some of those additional servers, boosting revenue and offsetting some of the expenses associated with capex. 

But it’s complicated. Profit margins associated with providing compute power for, say, training AI models for one large customer, such as OpenAI or Anthropic, tend to be quite low, Jaluria said, and that could affect the way the market interprets the results.

“Ultimately, the metric we’re looking at is effectively just going to be what cash flow can you generate over the long term from these investments?” he said. “Is it going to be enough to justify the spend or not? And the answers are not all created equal.”