Big Tech Spent Its Way to a Reckoning, and This Week Delivers the Bill

For most of the past three years, the question investors asked about artificial intelligence was whether the technology worked. That question has been answered well enough. The question now is arithmetic, and it is being asked with increasing hostility: at what point does the money spent building the infrastructure stop being an investment in the future and start being a permanent tax on shareholder returns?

Alphabet supplied the sharpest version of that question last week. The company beat on nearly every line that used to matter. Second-quarter revenue came in at $119.8 billion, up 24% from a year earlier and ahead of the roughly $117 billion analysts expected. Google Cloud, the segment most directly exposed to AI demand, grew 82% with operating margin expanding to 35.6%. On any conventional reading, that is an exceptional quarter.

The stock fell about 5% in after-hours trading. The reason was a single revised number: full-year capital expenditure guidance raised to a range of $195 billion to $205 billion, up from $180 billion to $190 billion.

Microsoft and Meta report on Wednesday. Apple and Amazon follow on Thursday. Within 48 hours, roughly a third of the S&P 500 by market value will have told the market what it intends to spend, and the market has made it unusually clear which number it will be reading first.

The capital intensity problem in one ratio

At the midpoint of the revised range, Alphabet would spend about $200 billion on capital projects this year. Measured against its annualised revenue run rate, that is roughly 42%. It is worth sitting with that figure, because it describes something that has no recent precedent in the software business.

A capital-to-revenue ratio in the forties is the signature of a utility, a telecommunications carrier or a heavy industrial. It is not what investors have historically paid a software multiple for. The entire valuation logic of the large platform companies rested on the opposite characteristic: enormous operating leverage, minimal incremental cost per additional user, and free cash flow that compounded because growth did not require steel and concrete.

AI infrastructure inverts that. Serving a model consumes power, silicon and cooling in direct proportion to usage. Growth now costs money in a way it did not five years ago, and the market is being asked to decide whether the multiple should follow the economics.

The scale of the commitment across the sector

Alphabet is not an outlier. Meta has guided to full-year 2026 capital spending of $125 billion to $145 billion, citing higher component costs alongside additional data centre capacity for AI workloads. Combined capital expenditure across the four largest hyperscalers is on track to exceed $650 billion this year, with consensus estimates compiled across the sell side clustering around $724 billion for 2026 and approaching $950 billion for 2027.

Longer-range projections are larger still, with some estimates putting cumulative hyperscaler spending on AI and data centres at several trillion dollars by 2030. Those numbers should be treated as directional rather than precise, but even the conservative versions describe an industrial build-out comparable in scale to national infrastructure programmes.

The demand side is not imaginary. Microsoft has disclosed an AI business running at roughly a $37 billion annualised revenue rate, up 123% year on year. Google Cloud growth of 82% is not a rounding error. Revenue is arriving. The dispute is about the gap between the rate at which revenue arrives and the rate at which capital is committed.

Why the depreciation schedule is the argument

Capital spending does not hit the income statement when the cheque is written. It arrives later, spread across the assumed useful life of the asset, as depreciation. This is where the debate becomes genuinely technical and genuinely consequential.

If a company assumes a six-year life for AI accelerators, a $200 billion outlay contributes roughly $33 billion of annual depreciation once fully in service. If the true economic life is three years, because a successor generation of chips renders the installed base uncompetitive for frontier workloads, the annual charge doubles and reported margins fall accordingly. The cash has already left either way. What changes is when the market is forced to acknowledge it.

Several hyperscalers extended their assumed server lives in recent years, a change that flattered reported earnings at the time. If the AI hardware cycle turns out to be faster than the general-purpose server cycle those assumptions were built for, that flattering runs in reverse. This is the mechanism through which an AI capex disappointment would show up, and it would show up in operating margin long before it showed up in revenue.

What the historical analogues actually teach

Investors reaching for precedent usually land on the late-1990s telecommunications build-out, and the comparison is more instructive than the usual bubble shorthand suggests.

Carriers laid vast quantities of fibre on the correct belief that internet traffic would grow enormously. They were right about demand. They were wrong about timing, and about who would capture the value. Capacity arrived years ahead of the applications that would use it, prices for transit collapsed, and the companies that financed the build were destroyed while the companies that later ran services over cheap bandwidth prospered enormously. The fibre itself was not wasted. The capital structures wrapped around it were.

The nineteenth-century railway manias followed a similar pattern, as did the initial rollout of electrical generation. The recurring lesson is not that infrastructure booms are irrational. It is that the returns accrue to whoever owns the asset after the first set of investors has been wiped out, and that the decisive variable is how the build was financed rather than whether it was needed.

That is the material difference in 2026. The hyperscalers are, by the standards of any previous infrastructure cycle, extraordinarily well capitalised. They are funding much of this from operating cash flow rather than leverage. But the financing mix has been shifting toward debt and off-balance-sheet structures at the margin, and that shift is where the risk now concentrates.

What each report has to answer this week

• Microsoft: whether Azure capacity constraints are easing, and whether the AI revenue run rate is still compounding fast enough to justify a capital budget growing at a similar pace. Fiscal fourth-quarter results land Wednesday.
• Meta: whether the $125 billion to $145 billion guidance range holds or moves higher again. Meta has the least direct revenue attribution for its AI spending of the four, which makes its guidance the most exposed to investor scepticism.
• Amazon: whether AWS growth is reaccelerating enough to support the capital intensity, and how much of the spend is committed versus discretionary. Results are due Thursday.
• Apple: the counter-example. Apple has spent far less on AI infrastructure than its peers and has been criticised for it. If the market is genuinely repricing capital intensity, relative performance this week will show it.

The memory bottleneck nobody guided for

One component of the cost increases deserves separate attention. Meta explicitly attributed part of its higher guidance to component costs rather than to additional capacity. High-bandwidth memory and advanced packaging have become the binding constraints on AI system production, and pricing in those markets has moved sharply.

This matters because it breaks the assumption investors have carried since the beginning of the build-out: that unit costs decline over time. If the cost per unit of compute is rising rather than falling, then a given capital budget buys less capability each year, and the spending required to stay at the frontier increases even if ambitions do not. That is a materially worse setup than the one implied by simple extrapolation from semiconductor history.

Outlook: the market has changed the question

The most important shift this earnings season is not in any company financial statement. It is in what the market rewards. A year ago, raising AI capital expenditure guidance was a signal of confidence and the stock went up. Alphabet raised its guidance last week off the back of an excellent quarter and the stock went down about 5%. Meta shares have fallen roughly 6% and Microsoft about 2.5% in comparable episodes, with beats on revenue failing to offset concern about payback periods.

That is a regime change in how the sector is valued, and it has arrived at an awkward moment. It coincides with a Federal Reserve meeting where the discount rate applied to distant cash flows is under active discussion, which our coverage of the July policy decision examines in detail. Long-duration assets financed by heavy near-term spending are precisely the assets most sensitive to that discussion.

None of this establishes that the spending is wrong. Demand signals from cloud revenue remain strong, and a company that underinvests in a genuine platform shift faces a worse outcome than one that overinvests. But the burden of proof has moved. For the past three years, executives were asked to explain why they were not spending more. Over the next 48 hours, they will be asked to explain when the spending pays for itself, and vague answers will be expensive.

---

About the data: Revenue, margin and capital expenditure figures are taken from Alphabet second-quarter 2026 results and guidance, Meta published 2026 capital spending range, and Microsoft disclosed AI revenue run rate. Reporting dates follow company earnings calendars. Sector-wide capital spending totals for 2026 and 2027 are analyst estimates, not company guidance, and are identified as such in the text.