Big Tech’s AI Capex Has Outrun Its Revenue — and the Stock Market Is Beginning to Notice
The Nasdaq-100 is down approximately 8% from its February 2026 peak as of early June, and the companies carrying the most weight in that decline are the same ones running the largest AI infrastructure buildouts: Microsoft, Alphabet, and Amazon. The correlation is not coincidental. Investors who priced these stocks on the AI growth story through 2024 and early 2025 are now demanding evidence that the $700 billion in combined AI capital commitment is generating revenue at a pace that justifies the valuation multiples built into the stocks at their peaks.
The evidence, so far, is mixed in a way that markets find harder to interpret than either clear outperformance or clear failure.
Revenue Growth Has Not Kept Pace With Capital Spending
The structural problem that equity markets are re-pricing is straightforward when expressed in growth rates. Microsoft’s Azure AI revenue grew approximately 35% year-over-year in the most recent quarter — impressive by conventional enterprise software standards, but against a capital expenditure base that grew 79% year-over-year to approximately $21 billion in the same quarter. When capex grows at twice the rate of revenue, the return on invested capital is declining, not improving. At some point — and Wall Street analysts are debating when — the capex growth rate needs to slow or the revenue growth rate needs to accelerate for the investment to be financially rational.
The combined hyperscaler capex figures referenced throughout this piece are aggregated from each company’s most recent SEC quarterly filings on EDGAR, and the cross-cloud comparison data tracks Bloomberg’s ongoing coverage of AI infrastructure capital allocation.
Google’s situation is structurally similar. Alphabet reported Q1 2026 capex of $17.2 billion against cloud revenue growth of 28%. Amazon’s AWS grew 17% while AWS-related infrastructure capex consumed an estimated 40% of Amazon’s total capital expenditure. The pattern across all three companies is the same: unprecedented infrastructure investment delivering revenue growth that is strong by historical standards but insufficient to demonstrate that the specific AI infrastructure investment is generating returns proportionate to its cost.
FactSet consensus data as of June 2026 shows that the Magnificent Seven’s combined forward P/E ratio has compressed from approximately 34x at the February peak to approximately 28x currently — a multiple de-rating that reflects the market’s recalibration of AI revenue timeline expectations rather than any change in the underlying business quality of these companies.
The Return Timeline Question
The AI infrastructure investment debate has a timing dimension that makes it analytically unusual. The data centers being built in 2026 have 15-20 year economic lives. The GPUs being installed this year will run inference workloads for 5-7 years. The TSMC wafer commitments placed today will produce chips delivered in 2027. The revenue from these assets does not arrive in the quarter the capex is spent — it arrives in the quarters and years when the infrastructure is deployed and operational.
This mismatch between capex timing and revenue realisation is standard in infrastructure-intensive industries. Telecom companies that built 5G networks showed similar capex-revenue timing disconnects; utility companies building generation capacity routinely spend years ahead of revenue. The market has been willing to extend patience to Big Tech on AI capex because the historical track record — AWS, Azure, Google Cloud — shows that hyperscaler infrastructure investment eventually generates sustained, high-margin revenue streams.
The question markets are now asking is not whether AI infrastructure generates returns but when. The consensus estimate among technology equity analysts is that AI infrastructure capex begins generating proportionate returns in 2027-2028, as inference workloads scale to utilise the capacity being installed in 2025-2026. The $250 billion committed by Amazon, Microsoft, and Google for 2026 alone is predicated on demand forecasts that assume AI inference consumption grows at 40-60% annually through 2029.
If those demand forecasts are correct, the current stock weakness is a buying opportunity in established infrastructure — not a structural repricing. If the forecasts are wrong — if AI inference demand plateaus or if efficiency improvements (better models requiring less compute per query) reduce the infrastructure required per unit of AI consumption — the capex built ahead of that demand becomes stranded cost.
The Efficiency Risk: Models Getting Cheaper to Run
The most underappreciated risk in the AI capex story is the efficiency improvement rate of AI models themselves. Each generation of foundation model delivers comparable or superior capability at lower inference cost — a trend driven by architectural improvements (mixture-of-experts, better quantisation, distillation) and hardware improvements (each GPU generation delivers more FLOPS per watt). If inference cost per query declines 50% year-over-year as it has over the past two years, demand needs to more than double annually simply to keep total infrastructure utilisation constant.
Metcalfe’s Law arguments about network effects and AI adoption curves suggest this demand growth is achievable. But the efficiency trajectory creates a specific risk: hyperscalers who have committed to specific GPU architectures (Blackwell, MI350) may find those architectures underutilised sooner than planned if next-generation models run efficiently on less capable or less densely packed hardware. The capital intensity of AI infrastructure means these decisions lock in costs for years, not quarters.
What the Sell-Side Is Saying
The shift in analyst framing from pure growth enthusiasm to return-on-capital scrutiny has been notable in the past 90 days. Goldman Sachs’ technology research team and several prominent hedge funds have publicly highlighted the capex-revenue timing gap as the primary risk to sustained Nasdaq outperformance in H2 2026. Morgan Stanley’s June 2026 AI infrastructure note argued that the market is currently applying a “build-it-and-they-will-come” premium that was appropriate in 2023-2024 but requires proof-of-revenue delivery in 2026-2027 to sustain.
The dissent from this bearish recalibration comes from the companies themselves and from the bull case on AI adoption: enterprises are in early stages of AI deployment, the productivity gains from enterprise AI are becoming measurable, and the inference demand that is just beginning to emerge from agentic AI applications will eventually exceed the infrastructure being built today. On this view, the current capex is not ahead of demand — it is barely keeping pace with demand that will be fully visible only in retrospect.
Both arguments are internally coherent. The market is pricing somewhere between the two — a multiple compression that acknowledges the timing risk without fully pricing in the failure scenario. How that compression resolves depends on H2 2026 earnings calls: if Microsoft, Google, and Amazon begin showing accelerating AI revenue growth against stabilising capex in Q3 and Q4, the current de-rating was a buying opportunity. If capex continues to accelerate while revenue growth decelerates, the re-rating has further to run.
The market is not pricing in failure. It is pricing in impatience.
What Investors Get Wrong About CapEx Timing
MorganHousel’s observation: investors consistently underestimate the lag between spending and results. Not because they are unsophisticated, but because the lag between investment and return in platform-infrastructure businesses is longer and more variable than the mental models most investors apply. The AI CapEx cycle is the most visible current example of a pattern that recurs in every capital-intensive technology transition.
The market’s frustration with big tech AI CapEx in 2026 follows a recognisable shape. The companies that are spending — Microsoft, Google, Amazon, Meta — are spending at levels that are, by multiple metrics, unprecedented in the history of commercial technology infrastructure. The revenue that these investments are supposed to generate is growing, but not at a rate that makes the payback period obvious from the current numbers. The stock market’s response has been to discount the companies most aggressively investing relative to the companies investing more conservatively.
This response is not irrational. It is the historically appropriate reaction to large CapEx announcements in industries where the payback period is uncertain. Investors were right to be sceptical of the first wave of broadband infrastructure investment in 1999 and 2000; most of it was indeed overbuilt relative to demand at the time of construction. They were wrong to remain sceptical of the second wave of cloud infrastructure investment in 2010 and 2011, when AWS, Azure, and Google Cloud were building capacity for a workload transition that most investors couldn’t yet see.
The question the current moment poses is which pattern the AI CapEx cycle resembles. MorganHousel would resist answering that question with confidence — not because the answer is unknowable in principle, but because the most important things about long-term investments are always the ones that can’t be modelled from current data. The companies that are spending know things about their near-term pipeline that investors cannot see. They also face risks that neither they nor investors can fully price: regulatory friction, competitive commoditisation, and the possibility that the models being trained today are not the models that will carry the revenue in 2027.
The $250B combined hyperscaler CapEx commitment is the specific bet being placed. Whether it was the right bet will be visible in the 2027 and 2028 earnings reports, not in the 2026 quarters where the spending is occurring. The companies making the bet know this. The market is penalising them for it anyway — which is what markets do when the horizon is longer than the reporting period.
The behavioural trap to avoid: anchoring to the CapEx number as evidence of either certainty (“they must know something”) or profligacy (“they’re burning money”). The more useful frame is that large companies in capital-intensive transitions are making probabilistic bets on infrastructure that has a ten-year payback horizon, and the first three years of that horizon will always look like waste to observers who are optimising for shorter time frames. That is not a defence of any particular company’s specific allocation decisions. It is a description of the category.
