Kurz gesagt: The four largest hyperscalers are converting a decade of surplus cash into fixed assets faster than AI revenue can justify, forcing investors to underwrite return on capital rather than the earnings growth that carried the trade until now.
Capex has roughly doubled in a year while the cash engine idles
Microsoft, Alphabet, Amazon and Meta have moved from disciplined reinvestment to an outright buildout. Combined capital expenditure across the four ran near $410 billion in 2025 and is guided toward roughly $725 billion in 2026, a step-up of about 77 per cent that most analysts now extend past $1 trillion by 2027. On a trailing-four-quarter basis through the first quarter of 2026, the four already spent $433.9 billion, with quarterly outlays of $129.8 billion growing 80 per cent year over year.
This is a different business, not a rounding change to a familiar one. Roughly half of the 2026 spend funds servers and silicon that depreciate over five to six years; the rest funds data-center shells and power infrastructure. The companies that spent the last decade compounding cash on other people’s balance sheets are now building the plant themselves.
| Unternehmen | 2025 capex (approx.) | 2026 guidance | YoY growth |
|---|---|---|---|
| Amazon | ~$100B | ~$200B | ~+100% |
| Microsoft | ~$95B | ~$190B | ~+100% |
| Alphabet | ~$85B | $175–185B | ~+110% |
| Meta | ~$70B | $125–145B | ~+80% |
| Combined | ~$410B | ~$725B | ~+77% |
Free cash flow, not earnings per share, is now the swing variable
Earnings per share (EPS) still look strong. The cash statement tells the story the income statement hides. Amazon reported first-quarter free cash flow of $1.2 billion, down roughly 95 per cent year over year, as quarterly capex reached $44.2 billion and nearly every incremental dollar of operating cash was redeployed into infrastructure. Meta generated $12.4 billion of free cash flow in the same quarter against $19 billion of capex, and one widely cited model has its free cash flow turning negative by 2027 — a company that produced $43.6 billion of it in 2025.
The mechanical consequence is the end of the buyback-and-dividend cushion that supported these multiples. When capital expenditure climbs above half of operating cash flow — the level Microsoft has now crossed — the return of capital to shareholders becomes the residual rather than the priority. Markets have begun to reprice both names on that trajectory rather than on their beat-and-raise headlines.
| Unternehmen | Q1 2026 free cash flow | Q1 2026 capex | Signal |
|---|---|---|---|
| Amazon | $1.2B (−95% YoY) | $44.2B | Amazon Web Services (AWS) margin compressed to 37.7% on AI depreciation |
| Meta | $12.4B | $19.0B | FCF modelled negative by 2027 (vs. $43.6B in 2025) |
A depreciation wall is scheduled for 2027 through 2029
The most underappreciated feature of this cycle is timing. Depreciation recognises spending gradually — five to six years for servers, twenty-five to forty for buildings — so today’s income statements capture only about one-third of current capex as expense. Trailing depreciation and amortisation of roughly $149 billion sits against $433.9 billion of trailing capex. The gap is deferred, not saved.
Meta’s infrastructure depreciation already rose from $7.32 billion in 2023 to $13.36 billion in 2025, an 83 per cent increase before the largest spending waves land. As the 2026 and 2027 cohorts deploy, that expense flows onto income statements through 2029, compressing reported margins mechanically — whether or not the revenue arrives. Firms are pre-funding the crossover: Meta, Alphabet and Amazon together raised well over $100 billion of debt across late 2025 and 2026, with sell-side desks projecting the sector may need $1.5 trillion of new debt over the coming years.
The revenue gap is the whole thesis
Every figure above resolves into one question: what is the return on this capital? AI revenue is real and growing quickly — AWS runs near a $150 billion annualised rate, Microsoft’s AI business is at roughly a $37 billion run rate growing 123 per cent, and Google Cloud grew 63 per cent in the quarter. Yet investment is scaling about 50 per cent faster than the revenue it is meant to produce, and Sequoia’s analysis puts the annual gap near $600 billion between infrastructure spend and what the AI ecosystem actually sells.
The demand-side evidence gives limited reassurance so far. MIT’s Project NANDA found that 95 per cent of enterprise generative-AI pilots produced no measurable profit-and-loss impact, against $30–40 billion of corporate spending. Capital intensity across the group now sits at 45 to 57 per cent of revenue — the range of a regulated utility rather than an asset-light platform — while sector enterprise-value-to-EBITDA (EV/EBITDA) multiples near 25 times approach the extremes last seen at the 2000 telecom peak. That comparison is a description of the risk the current price embeds, not a forecast.
What this asks of capital allocators
The bull case still holds. Compute demand is genuine, the leaders are the credible builders, and whoever owns the winning capacity may earn extraordinary returns. But the shape of the trade has changed: an investor buying these names in 2026 is underwriting a capital-allocation decision, not a cash-compounding machine. The asset-light model that justified premium multiples is being replaced, quarter by quarter, with a capital-intensive one — and the market has started to price the buildout rather than the boom.
Allegory Capital invests where capital intensity, returns on capital and the energy transition intersect, so this shift sits at the centre of how the firm reads the current cycle. Rather than close the argument, this piece opens it. Allegory Capital is seeking on-the-record and background perspectives from fund managers and technology-sector analysts on three questions:
1. When does capex intensity become a genuine concern, and what would prove the returns? Earnings are growing, but capex is absorbing a rising share of operating cash flow. At what point does that shift from a feature to a problem for you? What specific evidence would convince you that the hundreds of billions committed to AI infrastructure are earning an adequate return on capital — and which metrics (incremental return on invested capital (ROIC), capex-to-revenue, free-cash-flow conversion, cloud backlog, unit economics per token) are you actually watching?
2. Is revenue and EPS growth enough if free cash flow falls sharply? If buybacks and dividends shrink as cash is diverted into plant, does headline earnings growth still justify the cycle for you? And are investors underestimating how fundamentally AI is converting Big Tech from an asset-light, cash-generative model into a capital-intensive one that deserves a different multiple?
3. How long a lag between the dollar spent and the dollar earned is acceptable? What is the outer bound you will tolerate between capital deployed on AI infrastructure and the incremental revenue and cash flow needed to justify it — one year, three, five — before the thesis, rather than the timing, is in question?
Responses will inform an Allegory Capital briefing on the AI capital cycle. Contributions can be attributed or kept on background by request.
Referenzen
- Tom’s Hardware. Google, Microsoft, Meta and Amazon capex spending to hit $725 billion in 2026, up 77% from last year. https://www.tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion
- Silicon Analysts. Hyperscaler AI Capex 2026: $434B Trailing Four Quarters, D&A Lag, Debt Wave. https://siliconanalysts.com/analysis/hyperscaler-ai-capex-depreciation-wall-2026
- Global Data Center Hub. Amazon and Meta’s $335B Q1: The Free Cash Flow Inflection. https://www.globaldatacenterhub.com/p/amazon-and-metas-335b-q1-the-free
- ValueAdd VC. AI Hyperscaler Capex Compared. https://valueaddvc.com/blog/ai-hyperscaler-capex-compared-why-microsoft-google-meta-and-amazon-are-all-spending-at-once
- Forbes (Jason Kirsch). The AI Capex-to-Revenue Gap Is Widening. https://www.forbes.com/sites/jasonkirsch/2026/06/02/the-ai-capex-to-revenue-gap-is-widening—and-markets-are-starting-to-notice/
