AI Capex: In Our View, Priced to Perfection
Hyperscaler 2026 capex guidance totals roughly $320B. Industry data-center revenue is about $250B. The ROIC math is the part the market is not pricing.
The four US hyperscalers will spend roughly $320 billion on property and equipment in their 2026 fiscal years, based on the most recent guidance from Microsoft, Meta, Alphabet, and Amazon. That is the largest single-cycle, single-category capex wave in modern US corporate history.
In our view, that number is being priced as if every dollar earns 15%-plus ROIC. We do not think that is the right base case for a capex cycle this large, this concentrated, and this tied to a single end-market thesis. The market is not pricing a failure of AI. It is pricing the successful monetization of AI on hyperscaler depreciation schedules. Those are different risks.
This is an Opinion piece. We are not calling a crash, predicting a bubble, or making a short-term price forecast. We are saying the hyperscaler AI capex cycle is priced as if perfection is the base case, and we want to walk through the arithmetic that, in our view, makes that base case demanding.
The Capex Number, Source by Source
Here is the 2026 capex guidance as disclosed by each of the four US hyperscalers, rounded to the nearest $5 billion.
| Company | FY2026 capex guidance | Source |
|---|---|---|
| Microsoft (MSFT) | ~$80B+ | MSFT FY25 Q4 earnings release and FY26 guidance |
| Meta (META) | $60-65B | Meta Q4 2025 earnings call |
| Alphabet (GOOGL) | ~$75B | Alphabet Q4 2025 earnings release |
| Amazon (AMZN) | ~$100-105B | Amazon Q4 2025 earnings release |
| Combined | ~$315-325B | Sum |
We use $320B as the combined point estimate. That is roughly 2.4 times the four companies’ combined 2022 capex of about $135B (from each company’s 2022 10-K). In three years, the hyperscaler data-center build has more than doubled.
The comparative scale is not hypothetical. Synergy Research Group estimates the global cloud-infrastructure services market at about $250 billion in 2025 revenue. The four hyperscalers’ 2026 capex exceeds the entire industry’s current annual revenue.
The ROIC Bar
Start with a simple calculation. If $320B of capex is depreciated over six years (MSFT and GOOGL extended to six; AMZN and META use similar five-to-six-year schedules per their 10-Ks), annual depreciation on this single vintage is about $53B. For this vintage alone to earn a 15% unlevered pre-tax ROIC on the capital deployed, it needs to generate roughly $48B in incremental annual operating profit on top of covering the $53B depreciation.
Put differently: the 2026 vintage alone, plus the already-deployed prior vintages, needs to produce more than $100B in incremental AI-attributable operating profit just to earn mid-teens returns on what has been spent.
For context, total hyperscaler AI-attributable revenue in 2025 is estimated at somewhere between $50B and $70B across the four companies, based on the commentary each has given on its earnings calls about AI contribution to Azure, Google Cloud, and AWS. Meta does not sell AI as a standalone product; it monetizes through ads and engagement. That means incremental operating profit of $100B would require roughly doubling AI-attributable revenue in 12 months at industry-leading incremental margins.
We are not saying that is impossible. We are saying that is what the current multiple on the group implies.
Historical ROIC on Hyperscaler Capex
Hyperscaler ROIC on legacy data-center and cloud capex has historically run in the 11% to 13% range, based on analyst work compiled by Morgan Stanley and others (via Bloomberg and Reuters coverage of bank research). This is the prior regime: capacity capex earned decent but not exceptional returns, the business compounded at infrastructure-industry pace, and the multiple re-rated as AWS/Azure/GCP scaled.
Three things have changed at once in the current cycle.
- The absolute dollar is 2.4x larger. ROIC compounds on a larger base. Small percentage-point changes in ROIC now translate to bigger absolute profit-pool swings.
- The vintage is concentrated in GPUs and HBM memory, not general-purpose CPU servers. Depreciation schedules on the new mix are contested. MSFT and GOOGL extended to six years in 2024-2025, which, per MSFT’s own disclosure, flattered 2024 EPS by a few cents per share. If the useful-life assumption is wrong and actual useful lives are closer to four years, reported depreciation is understated.
- The revenue end-market is less diversified. Prior capex cycles served many use cases. This cycle’s marginal dollar is substantially tied to inference and training for generative-AI workloads.
If historical hyperscaler ROIC on capex runs 11%-13% and the current market is priced for 15%-plus, the gap is a narrow but real ROIC shortfall risk. In our view, the gap is what investors are implicitly under-pricing.
What a 20% Capex Cut Would Do
Assume, hypothetically, one of the four hyperscalers cuts 2026 capex guidance by 20% on the next earnings call (the Q1 2026 calls begin the week of April 28-30, 2026). That is $15B to $20B of removed spend in the back half of 2026, concentrated in data-center builds.
The direct revenue impact on NVDA is asymmetric. NVDA data-center revenue in fiscal 2026 runs around $130B, per NVDA filings, with more than 80% of that from hyperscaler customers. A $15-20B capex cut from one hyperscaler is not a $15-20B NVDA revenue cut, because some of the cut is in real estate, power contracts, and networking gear. But it is a meaningful signal about the run-rate.
More importantly, a single hyperscaler cut would compress the forward growth assumption across the whole AI compute stack: semiconductors, AI-adjacent power utilities, specialty cooling, and the BDC/credit financing that is increasingly funding data-center construction at the edges.
We do not need to forecast a capex crash. A speed-limit cut from any of the four, in our view, is enough to reset the multiple on the group.
Three Historical Parallels
Capex super-cycles of this magnitude have happened before. None of them ended with the market being wrong about the underlying technology. All three ended with the market being wrong about the timing of monetization.
- Telecom 1998-2001. Capex on fiber and wireless tripled from 1996 to 2000. The technology thesis was correct. Most of the deployed fiber sat unlit for a decade. The equity unwind in the stocks that funded the build (JDSU, Nortel, Lucent, Level 3) ran 80%-plus from peak.
- Solar manufacturing 2008-2011. Capex on polysilicon and module capacity ran ahead of end-market demand. The technology thesis was correct. China solved the cost curve. The equity unwind in US/European producers ran 70%-plus.
- Crypto mining 2020-2022. Capex on mining rigs and specialized ASICs ran ahead of mined-revenue economics in the drawdown. The technology thesis is still contested. The equity unwind in miners ran 80%-plus.
In each case, the correct intellectual bet was on the technology. The money was lost in the capex-heavy vehicles that funded the build. We are not saying AI is telecom 1999. We are saying the parent question, does this capex cycle earn its depreciation, is the same question.
The Depreciation Sleight of Hand
A footnote that matters. When MSFT extended its server-useful-life assumption from four to six years in its fiscal 2023 disclosure and GOOGL followed with a similar extension, each company flattered GAAP earnings by reducing annual depreciation. MSFT’s own 10-K commentary quantified the 2024 benefit at roughly five cents per share.
The issue is not that extending useful life is wrong. Some servers do genuinely run longer. The issue is that if AI-specific hardware (GPUs in particular) has a shorter useful life than general-purpose servers because of thermal, software-stack, and model-turnover factors, then the extension was optimized for legacy hardware and is being applied to a radically different mix.
If the true useful life of GPU-heavy capex is four years instead of six, reported 2026 earnings across the four hyperscalers are overstated by roughly $15-20B in aggregate. That is not a fraud claim. It is an accounting judgment, and it is not in our view an obviously conservative one.
What Would Change Our View
We are Opinion-category on this piece. We would update our view on the following.
- If Q1 2026 earnings calls show AI-attributable cloud revenue growing 70%-plus year-over-year at improving margin, the monetization path is tracking ahead of the capex build.
- If any of the four hyperscalers reduces 2026 capex guidance by more than 10%, the market will reset multiples on the ecosystem, and our “priced to perfection” framing becomes less distinguishable from consensus.
- If useful-life disclosures are revised back toward four or five years, reported earnings will take a one-time haircut, and the accounting-optimism concern we flagged above gets resolved the hard way.
- If hyperscaler capex concentration keeps rising (one buyer funds 90% of NVDA data-center), the concentration risk gets worse, not better.
In Our View
In our view, combined hyperscaler 2026 capex of roughly $320 billion is the largest single-cycle industrial deployment in recent corporate memory, and the market is pricing the payback on a schedule that assumes flawless monetization, steady margins, and accounting-useful-life assumptions that may or may not prove correct for GPU-heavy vintages.
We are not making a recommendation on any specific ticker. We are not predicting a crash. We are saying that the gap between historical hyperscaler ROIC (11%-13%) and the ROIC implied by current multiples (15%-plus) is the part of the picture the market, in our view, is not pricing.
For how this connects to broader index exposure, see our Mag 7 Concentration research piece. For a historical take on what capex-heavy super-cycles have looked like in fixed income, see our Private Credit 2001 vs 2008 opinion piece.
Ferrante Capital LLC is a registered investment adviser. Information presented is for educational purposes only and does not constitute investment advice, a solicitation, or a recommendation to buy or sell any security. All investing involves risk, including the possible loss of principal.
FC and its principals may hold positions in AMZN, GOOG, GOOGL, META, MSFT, NVDA. This analysis is for educational purposes only and does not constitute a recommendation to buy, sell, or hold any security.
Forward-looking statements reflect Ferrante Capital’s current analysis and involve assumptions and estimates. Actual results may differ materially. Past performance is not indicative of future results.
Please consult a qualified financial professional before making investment decisions.
