M7 · Hyperscaler CapEx — Cash Flow Divergence: AI Infrastructure Capex vs. Shareholder Returns
Hyperscaler capex, quarter: $201.9 bn as of 2026-09-30.
Hyperscaler AI Capital Expenditure — Quarterly infrastructure capex tracking across top cloud hyperscalers driving the AI compute buildout.
The plate
Eight years of ramp: the four builders' quarterly capex
From 2018-Q1 to 2026-Q3, the four builders' combined quarterly capex went from ~$16B to ~$202B — roughly 13x. One line per company; the faint line above them is the four combined.
Latest-quarter CapEx
Most-recent reported quarter for each of the seven M7 names (US$ billions). As of 2026-Q3. The four cloud builders are running $33B-$78B per quarter; the rest sit at $2-5B — an order of magnitude inside a single basket.
What each company has actually committed to for CY2026
These aren't sell-side consensus numbers — each is the company's own framing from its latest earnings call. The four together commit to ~$730B for CY2026, roughly 1.9× CY2025 actual. GOOGL's CFO said CY2027 will "significantly increase" but didn't put a number on it, so no 2027 bar.
Where operating cash flow goes
Each bar is the latest fiscal year's OCF split into capex / shareholder return (buybacks + dividends) / residual. Bars are normalized to 100% of OCF so allocation, not size, is what's visible.
FCF / OCF conversion: from pre-AI to now
FCF = OCF + CapEx. Higher means more free cash falling out of every operating dollar. The four builders slide; the other three stay flat near the ceiling — the divergence is the AI-cycle cash story.
Eight years of ramp: the four builders' quarterly capex
From 2018-Q1 to 2026-Q3, the four builders' combined quarterly capex went from ~$16B to ~$202B — roughly 13x. One line per company; the faint line above them is the four combined.
Latest-quarter CapEx
Most-recent reported quarter for each of the seven M7 names (US$ billions). As of 2026-Q3. The four cloud builders are running $33B-$78B per quarter; the rest sit at $2-5B — an order of magnitude inside a single basket.
What each company has actually committed to for CY2026
These aren't sell-side consensus numbers — each is the company's own framing from its latest earnings call. The four together commit to ~$730B for CY2026, roughly 1.9× CY2025 actual. GOOGL's CFO said CY2027 will "significantly increase" but didn't put a number on it, so no 2027 bar.
Where operating cash flow goes
Each bar is the latest fiscal year's OCF split into capex / shareholder return (buybacks + dividends) / residual. Bars are normalized to 100% of OCF so allocation, not size, is what's visible.
FCF / OCF conversion: from pre-AI to now
FCF = OCF + CapEx. Higher means more free cash falling out of every operating dollar. The four builders slide; the other three stay flat near the ceiling — the divergence is the AI-cycle cash story.
The cohort has divided into distinct strategic paths: Microsoft, Alphabet, Amazon, and Meta channel massive operating cash flows into compute infrastructure, while Apple, Nvidia, and Tesla emphasize buybacks, fabless efficiency, or specialized industrial expansion.
What this page answers
This static page is built to answer searches for M7 · Hyperscaler CapEx. It summarizes the live dataset behind the Cash Flow Divergence: AI Infrastructure Capex vs. Shareholder Returns panel and links to the full interactive chart.
The cohort has divided into distinct strategic paths: Microsoft, Alphabet, Amazon, and Meta channel massive operating cash flows into compute infrastructure, while Apple, Nvidia, and Tesla emphasize buybacks, fabless efficiency, or specialized industrial expansion. The data is refreshed by the History of Market pipeline and published as a stable JSON endpoint for research, citation, and AI-agent use.
Data & Source
GET /api/mag7/ai-capex.json — Canonical dataset endpoint.
Exchange closing prices · Company filings · Robert Shiller · FRED · NBER.
FAQ
Where does this data come from?
History of Market combines public market and macro datasets — exchange closing prices, company filings, Robert Shiller, FRED, NBER, and the UBS Global Investment Returns Yearbook (Dimson–Marsh–Staunton). The exact endpoint for this panel is linked below.
How often is it updated?
Daily-tier datasets refresh after the U.S. market close, with a broader weekly refresh each weekend. The timestamp shown on this page comes from the JSON payload.
Can I use the data?
Yes, for research and education with attribution to History of Market. Upstream data sources retain their own terms.