Big Tech Strikes AI Gold, But at a Steep Cost

Microsoft and Alphabet reported earnings on April 29, 2026, and the numbers confirm a story that has been building for two years: Big Tech is spending massively on AI infrastructure, and the resulting depreciation charges are now visibly compressing operating margins[1].

When a company builds a data center, it does not write off the entire expense in one quarter. Instead, it spreads the cost over the useful life of the equipment, typically four to six years for servers and longer for buildings. That annual charge, called depreciation, flows through the income statement as an expense even though no cash left the building that quarter. For Microsoft, Alphabet, Amazon, and Meta, the depreciation piles are growing faster than revenue, compressing margins in a way that would have been unthinkable in the high-margin, low-capex tech environment of the 2010s.

The scale is staggering. Microsoft is guiding approximately $190 billion in capital expenditures for calendar 2026, the company announced on April 29[1]. Alphabet raised its 2026 capex guidance to $180-190 billion, and its Q1 2026 capex hit $35.7 billion[1]. Combined 2025 capex for Microsoft, Alphabet, Amazon, Meta, and Oracle reached roughly $448 billion, according to industry tallies, up from $162 billion in 2022[1].

The revenue side of the equation is real. Google Cloud grew 63% year-over-year in Q1 2026, reaching $20.0 billion, as businesses moved beyond experimenting with AI to integrating it into production workflows[1]. Microsoft Azure and other cloud services grew 40% on a reported basis, or 39% in constant currency[1]. But both companies conceded on their earnings calls that AI revenue, while growing fast, is not yet growing fast enough to offset the infrastructure cost drag.

Amazon Web Services remains the largest cloud provider, but its earnings call contained a more strategic story buried beneath the numbers[1]. According to Stratechery, Amazon’s bet on its custom Trainium chips is beginning to pay off as the AI market shifts from training large models to running inference and deploying agents[1]. During the training phase, which dominated 2023 and 2024, the priority was raw compute power, a domain where NVIDIA’s H100 chips ruled without serious competition. But inference runs a different computation: it requires cost-efficient, low-latency routing of predictions through already-trained models. Amazon’s homegrown silicon is designed for that exact workload[1].

The shift matters for the entire industry. If inference becomes the dominant AI compute pattern, as many analysts expect, the market will prize cost efficiency over peak performance. That favors companies like Amazon, which has a long history of turning high-margin technology markets into low-margin commodity businesses where it wins on scale. It also puts pressure on NVIDIA, whose GPU business relies heavily on training workloads, though the company also sells inference-capable chips like the H100 and H200.

Meta’s situation is different. It does not sell cloud services, so its AI investments must generate returns through improved user engagement and ad targeting rather than direct compute sales. That may be harder to measure. The company posted strong advertising revenue growth last quarter, but CFO Susan Li warned that the company has continued to underestimate its compute needs even as it has ramped capacity, and Meta raised its 2026 capex guidance to $125-145 billion[1]. CEO Mark Zuckerberg defended the spending, framing it as an investment in AI-powered features across Facebook, Instagram, and WhatsApp, with the payoff coming through sustained user engagement and higher ad conversion rates, metrics that are notoriously difficult to attribute to any single technology investment.

The fundamental question for investors is whether this spending cycle is different from past tech infrastructure booms. In the late 2010s, cloud providers built data centers expecting demand that eventually materialized, but the period between spending and payoff squeezed margins for several years. The current AI buildout is happening at a scale that dwarfs those prior cycles. Combined capex at the five hyperscalers has nearly tripled in three years, from $162 billion in 2022 to $448 billion in 2025, and consensus forecasts put the 2026 figure above $600 billion[1].

UBS analysts have flagged capex intensity, the ratio of capex to revenue, as the key metric to watch. For the big four US tech companies, that ratio has nearly doubled over the past five years to roughly 21%, with some forecasts seeing it rise toward 27% by 2030[1]. If revenue from AI services keeps growing faster than the capex bill, the math works. If capex continues to outpace revenue as it has for the past two quarters, margins will keep compressing and the market’s patience will eventually erode.

The next signal comes in late July, when Microsoft and Alphabet report their second-quarter earnings and give updated capex guidance for the remainder of 2026. If Microsoft announces that spending will level off, the market may interpret it as a sign that the buildout is peaking. If Meta raises its guidance again, investors will want to see concrete evidence that AI spending is translating into ad revenue growth. If Amazon’s Trainium chip begins to show up in its income statement as a cost advantage over GPU competitors, that will validate a thesis that is still unproven at scale.

For now, the gold rush is real and the revenue is real. But the cost of digging has never been higher, and the earnings reports in the coming quarters will determine whether this is a ten-year investment cycle or a five-year overbuild.