AI Capex Tops $700 Billion as Payoffs Separate From Promises

On April 29, within a 90-minute window after the closing bell, four of the world's largest companies reported their quarterly results. Alphabet raised its full-year capital expenditure forecast to as much as $190 billion. Microsoft lifted its own capex outlook to roughly the same figure. Meta stood firm on a range of $115 billion to $135 billion, nearly double what it spent in 2025. Amazon, which had already telegraphed a roughly $200 billion target, held the line. Combined, the four hyperscalers guided toward more than $700 billion in 2026 capital spending, a sum larger than the annual GDP of Switzerland. By the time the after-hours session ended, two of the four stocks had rallied sharply and two had sold off. The dividing line was not the size of the cheque each company was writing. It was whether anyone could see the revenue on the other side.

That sorting mechanism is the defining financial story of the current AI cycle. For two years, Big Tech's capex escalation drew undifferentiated enthusiasm. Every upward revision was treated as a signal of ambition, and ambition was its own reward. The Q1 2026 earnings season, which ran from late April through early May, changed that. As Forbes reported in its post-earnings assessment, "Alphabet surged while Microsoft and Meta fell as Big Tech earnings revealed a new divide: investors reward AI payoffs, not spending." The sentence is precise enough to serve as a thesis statement for the quarter. The capital allocation question is no longer about how much is being spent. It is about whose spending is producing visible, attributable revenue growth at a margin that justifies the outlay.

Alphabet provided the cleanest answer. Google Cloud revenue rose 63.4 percent year over year in the March quarter, accelerating from an already torrid pace. Search revenue, the business that for years faced predictions of obsolescence at the hands of AI-powered answer engines, grew 19.1 percent. The company disclosed a cloud backlog of $462 billion, providing multi-year visibility into future revenue, and reported a cloud operating margin of 32.9 percent. These are not the numbers of a business merely defending its turf against AI disruption. They are the numbers of a business whose AI investments are already pulling revenue through the income statement, at scale, with expanding profitability. In after-hours trading, Alphabet shares rose roughly 7 percent.

The capex guidance that accompanied those results was, on its face, staggering. Alphabet told investors it would spend as much as $190 billion in 2026 and expects to "significantly increase" that figure in 2027, CNBC reported. For any other company, a commitment to raise spending from $190 billion to an even larger number the following year would have prompted a reckoning. For Alphabet in the last week of April, it did not, because the revenue curve was bending upward faster than the capex curve. That is the arithmetic the market is now running on every hyperscaler: not the absolute level of spending, but the spread between the growth rate of AI-related revenue and the growth rate of AI-related capital outlays.

Microsoft's quarter illustrated the inverse of that dynamic. Azure revenue accelerated to 40 percent growth, topping the company's own forecast, and Microsoft disclosed a $37 billion annualised AI revenue run rate, GeekWire reported. By most conventional measures, those are excellent numbers. Yet Microsoft shares fell after the report, because the company simultaneously raised its 2026 capex outlook to nearly $190 billion, and the trajectory of that spending is now compressing operating margins and free cash flow at a rate that made investors uncomfortable. The AI revenue is real and growing, but it is not yet growing fast enough to outpace the infrastructure bill.

Microsoft's chief financial officer, Amy Hood, addressed the supply-side constraints directly, telling analysts that hardware and supply chain limitations "can be managed" and that "we are doing our best," Benzinga reported. The phrasing was notable for its absence of triumphalism. A year earlier, executives across the hyperscaler set had spoken about AI infrastructure spending as a generational opportunity. By late April 2026, the tone had shifted to operational reassurance, as if the primary message management needed to convey was not that the spending would produce wonders, but that it would not spin out of control.

Meta's quarter delivered the most dramatic repricing. The company posted its best sales growth since 2021, but shares fell roughly 10 percent in the following session, erasing approximately $175 billion in market capitalisation, Yahoo Finance reported. The proximate cause was capex. Meta guided to a range of $115 billion to $135 billion for full-year 2026, an increase of roughly 60 to 85 percent over the $72.2 billion it spent in 2025. Unlike Alphabet, whose cloud business gave investors a direct line of sight from infrastructure investment to invoiced revenue, Meta's AI spending is overwhelmingly directed at ad-ranking improvements, recommendation models, and the long-gestation bet on its Muse Spark assistant. Those projects may eventually produce returns, but the timeline is less legible and the revenue attribution less direct.

The Meta sell-off was not, in itself, evidence that the spending is misguided. It was evidence that the market has raised its burden of proof. Two years ago, a company that announced a doubling of AI capex could expect its stock to rise on the strength of the narrative alone. Today, that same announcement triggers a demand for visible payback, and the absence of it is punished in real time. Meta is essentially asking shareholders to fund a multi-year infrastructure build-out on the premise that AI-driven ad improvements will compound. The market, after one look at Alphabet's cloud margins, is no longer satisfied with a premise.

Amazon occupied a middle ground that, in the context of the week, registered as a qualified win. AWS revenue rose 28 percent to $37.6 billion in the quarter, driven by AI workload migration and an expanding inference business. CEO Andy Jassy told investors the company was holding its 2026 capex target at roughly $200 billion, a figure that includes spending on AI infrastructure, custom silicon, logistics robotics, and the Project Leo satellite broadband initiative. The key word was "holding." Amazon did not raise its capex guidance. In a week when Microsoft, Alphabet, and Meta all signalled higher spending, the absence of an increase was read as restraint, and restraint was rewarded.

The Economist captured the broader cashflow anxiety in a piece published shortly after the earnings wave, noting that "in short order America's biggest companies have gone from printing money to burning it" and that "Amazon, Meta and Microsoft are all" experiencing a deterioration in free cash flow as capex climbs. The framing is important because it shifts the analytical lens from the income statement to the cashflow statement. Reported earnings can be managed through depreciation schedules, capitalisation policies, and the timing of asset retirements. Free cash flow is harder to dress up. When free cash flow turns negative, or declines sharply, the company is consuming capital faster than its operations can replenish it. That is the territory several hyperscalers are now entering, and the market has begun to notice.

The Two Camps and the Capital That Divides Them

The earnings season crystallised a division that had been forming since mid-2025. On one side are the companies whose AI capex is visibly pulling revenue: Alphabet, with its cloud and search acceleration, and, to a lesser but still discernible extent, Amazon, with AWS workloads shifting toward higher-margin AI inference. On the other side are the companies whose AI capex is predominantly cost-side spending without equally visible top-line attribution: Meta, whose ad products are being re-engineered beneath the surface, and Microsoft, whose Azure AI growth is strong but whose cumulative infrastructure bill is growing faster than the associated revenue. This is not a judgment about which strategy will ultimately prevail. It is a description of what the market, in its quarterly verdicts, is actually rewarding.

The cognitive scientist and AI commentator Gary Marcus offered a more caustic assessment shortly after the earnings deluge. "Big Tech's AI spending is the 'greatest capital misallocation in history,'" Marcus said, Business Insider reported, warning that the returns on the hyperscalers' combined outlay would prove far lower than current valuations imply. Marcus's view sits at the far end of the scepticism spectrum, and it is worth noting that he has been making variants of this argument since 2023, a period during which the Nasdaq added roughly $10 trillion in market value. But his framing captures a legitimate question that the most bullish analysts have not fully answered: what is the aggregate return on the $700 billion the hyperscalers will deploy this year, and when will it be measurable in a way that is not subject to accounting discretion?

The bull case was articulated most vividly by Wedbush analyst Dan Ives, who told CNBC in mid-May that the Nasdaq will reach 30,000 points within the next year as the AI rally expands beyond the semiconductor names that dominated the first phase of the cycle. Ives used the earnings season as evidence that "a bumper earnings season continues to bolster enthusiasm for AI," and his note to clients, titled "The haters will hate," argued that the hyperscaler capex build-out is the foundation of a multi-year technology supercycle. The argument is internally consistent: if AI infrastructure spending is creating genuine productive capacity, the revenue acceleration visible at Alphabet and Amazon will eventually spread to Microsoft and Meta as well.

Big Tech's AI spending is the 'greatest capital misallocation in history.', Gary Marcus, cognitive scientist and AI commentator, as reported by Business Insider

The tension between the Marcus and Ives views is not resolvable with the data currently available. Both are extrapolating from the same set of quarterly filings, and both are projecting forward along different assumptions about the shape of the adoption curve. What makes the current moment distinctive is that the market is no longer treating the bullish extrapolation as the default. For the first time since the AI capex cycle began in earnest, investors are demanding evidence of payback on a company-by-company basis, and they are pricing the results with a granularity that was absent in 2024 and early 2025.

The capex figures themselves warrant disaggregation. Not all AI spending is created equal. A large portion of the hyperscalers' $700 billion-plus outlay is going to Nvidia GPUs and the data centre real estate required to house them. Another significant slice is being directed toward custom silicon, Amazon's Trainium and Inferentia chips, Alphabet's Trillium TPUs, Microsoft's Maia accelerators, each of which reduces long-term unit costs at the expense of near-term R&D and fabrication commitments. A third category covers the power infrastructure, networking fabric, and liquid cooling systems that make AI clusters operable. And a fourth, harder to isolate, category covers the inference capacity being built to serve AI workloads that have not yet materialised at scale. This last category is the one that most concerns the sceptics. It represents capital being deployed against demand that is still hypothetical.

The supply side of the equation introduces additional uncertainty. One of the least-discussed constraints in the current cycle is the memory chip supply chain. High-bandwidth memory, the specialised DRAM that sits adjacent to AI accelerators and largely determines their effective throughput, remains in tight supply. The same is true of advanced packaging capacity, which governs how quickly chips can be assembled into functional servers. These constraints mean that even the headline capex figures may understate the true level of investment intent: some of the spending being guided now is effectively prepayment for capacity that will not be delivered until 2027 or later. The capex numbers are large, in other words, but they are not necessarily efficient.

What to Watch When the Next Quarter Lands

The second quarter, which closes at the end of June, will provide the next data point in what is now a clearly established pattern. Three indicators warrant particular attention. The first is the trajectory of cloud revenue at Microsoft and Amazon relative to their respective capex growth rates. If Azure sustains 40 percent growth and AWS holds at 28 percent or accelerates, the market may begin to extend the Alphabet treatment to those names as well. If either decelerates, the sell-off that hit Meta could widen. The second indicator is free cash flow. Depreciation is a non-cash charge that smooths the impact of capital spending over several years, but free cash flow tells the story in real time. A further deterioration in free cash flow at Meta or Microsoft would test the patience even of investors who accept the long-term thesis.

The third indicator is the rate at which inference workloads are migrating from training clusters to production serving environments. Training a large model is capital-intensive but time-limited. Serving that model to millions of users, by contrast, creates ongoing demand for compute that maps directly to recurring revenue. The companies that can demonstrate a rising inference-to-training ratio in their capital allocation will have the strongest claim on investor confidence, because inference spending is demand-pull rather than supply-push. Alphabet's cloud backlog suggests it is further along this curve than its peers. Whether Microsoft, Amazon, and Meta can close the gap in the second half of 2026 will determine whether the capex cycle is remembered as the greatest capital misallocation in history or the foundation of the next decade's technology infrastructure. The market, for the first time, is demanding an answer before writing the next cheque.