Update: Hyperscaler Capex Intensity Jumps Another +2.3 pp — Big-4 Now at 24%

What changed since the July research vintage

In late July we published the theme’s baseline: five hyperscalers pushing 18–37% of revenue into capex, with Meta and Oracle already inside the dot-com telecom peak band. That post answered the stock question — what fraction of sales is being reinvested? This update answers the flow question capital markets actually trade on: what moved in the newest official vintage, and does the free-cash-flow cushion still cover the step-up?

Three filing events force a refresh. Microsoft closed FY26 (June year-end), so the prior FY25 print of 22.9% is no longer the live full-year ratio. Amazon, Alphabet, and Meta reported H1 2026, which we annualize for a run-rate comparison against FY25. Oracle restated FY25 from the estimated 37.0% we carried in July to a disclosed 39.4%. The dashboard above is built as a vintage delta — prior bars, new bars, percentage-point changes, and an intensity×FCF scatter that overlays both vintages.

The headline table: prior print vs Aug 2026 update

Revenue-weighted across the big four (ex-Oracle), intensity moves from 21.8% → 24.1% — a +2.3 percentage-point vintage delta. That is not a rounding error. It is another full step from the pre-AI cloud norm (~11%) deeper into the historical wireline telecom band (~20%) and toward the stretched industrial range where foundries and peak-cycle carriers live.

Microsoft’s FY26 is the cleanest delta in the set

Calendar-year names still require annualization. Microsoft does not. FY26 capex of $88.4B on $329.8B of revenue prints 26.8% intensity — +3.9 pp versus FY25 and the largest single-name step-up in the update. Free-cash-flow margin compresses from 25.4% to 22.1%, still the healthiest cushion among the five, but the direction is unambiguous: Azure and AI capacity are taking a larger bite of each revenue dollar even as the cash engine remains intact.

For equity and credit desks that treat Microsoft as the “affordable” AI builder, the update reframes the debate. Affordability is no longer “low teens intensity with mid-20s FCF.” It is high-20s intensity with low-20s FCF — still survivable, no longer the old software multiple story. Pair the ratio with the absolute stack in the AI capex spend map: the same firm that looks resilient on intensity still contributes a rising share of the industry dollar total.

H1 annualization: Amazon, Alphabet, Meta keep climbing

Amazon’s H1×2 run-rate intensity of 21.2% (+2.8 pp vs FY25) coincides with an FCF margin of only 2.4%. That is the fragile edge of the cohort. Retail revenue still dilutes the percentage, so the ratio understates how AI-heavy AWS’s own P&L has become — but it also means there is almost no group-level cash buffer if utilization or pricing slips. Absolute annualized capex near $162B keeps Amazon the largest cash absorber in the set.

Alphabet’s annualized 25.4% (+2.7 pp) pushes the search-and-cloud complex firmly into elevated-reinvestment territory, with FCF margin sliding to 12.6%. Meta’s 36.5% (+1.8 pp) remains the advertising-funded extreme: still inside the 30–40% telecom-peak band, still printing mid-teens FCF, still asking equity holders to underwrite a foundry-like reinvestment rate on a consumer-platform revenue base.

Toggle the dashboard’s metric control from intensity Δ to FCF margin Δ and the pattern is uniform: every name’s cushion thins as intensity rises. The scatter panel (both vintages) shows the cohort drifting right and down — more reinvestment, less residual cash per dollar of sales.

Oracle’s restatement matters more than the guide

July’s research post flagged Oracle FY25 as estimated. The restated 10-K intensity of 39.4% (+2.4 pp versus that estimate) is a methodology correction, not a new build wave — but it relocates Oracle from “high-30s estimated” to “high-30s disclosed,” with FCF margin at 7.2%. Early FY26 guide language implies intensity can print near ~41% on the path panel’s H1’26* marker. At that level Oracle is no longer a software peer with a temporary spike; it is a capacity land-grab priced like industrial infrastructure, funded increasingly through the credit channel documented in our AI financing research.

Restatements also discipline the vintage narrative. When the prior post’s estimate was low, every subsequent “surge” headline risks double-counting. We separate restatement deltas (Oracle) from true period deltas (Microsoft FY26; H1 annualizations) so markets can tell measurement from momentum.

Sustainability: the hinge is still FCF, not the ratio alone

A firm can run 25%+ intensity for years if operating cash flow covers capex, buybacks, dividends, and debt service. The update’s scatter is the sustainability dashboard: prior points sit semi-transparent; new points sit opaque. The message is directional, not apocalyptic:

1. Microsoft remains the textbook high-intensity, still-cash-rich casebut the gap between intensity and FCF has narrowed.
2. Meta funds mid-30s intensity with mid-teens FCFuncomfortable versus its own history, survivable versus most industrials.
3. Alphabet is compressing toward the mid-teens FCF / mid-20s intensity quadrant where utilization and cloud pricing have to do more work.
4. Amazon is the stress case: intensity into the low-20s with FCF near 2%.
5. Oracle pairs nearly 40% intensity with single-digit FCFclassic build-now, harvest-later posture.

Sustainability is therefore not a single threshold. It is whether cash conversion survives the reinvestment rate long enough for utilization, power delivery, and model pricing to amortize the PP&E wave. Intensity can stay “affordable” for two years and still destroy equity value if incremental ROIC disappoints.

Why another +2 pp still matters for capital markets

Percentage points feel small next to trillion-dollar spend headlines. They are not small on multiples. Moving the big-four weighted intensity from ~22% to ~24% is the difference between “elevated cloud” and “telecom-steady-state” in many sell-side frameworks. Credit desks funding a rising share of the build (see the financing post) care because leverage capacity is a function of FCF, and FCF is intensity’s residual. Equity desks care because software multiples do not survive industrial capital intensity without industrial utilization stories.

The path chart — FY24 → FY25 → H1’26* — shows the climb is not mean-reverting yet. Every active name is higher than FY24. The pre-AI cloud reference line at 11% is historical color, not a forecast anchor. The telecom ~20% line is now a floor for the weighted cohort, not a ceiling.

Caveats and methodology

1. Gross PP&E ≠ AI-only spend. Filings do not cleanly split accelerators from warehouses, offices, or network gear. Amazon’s ratio mixes retail fulfillment with AWS; treat company-level intensity as an upper-bound proxy for “AI intensity.”
2. H1×2 annualization is not a forecast. Front-loaded GPU deliveries or seasonal cloud revenue can bias the run-rate. We label Amazon/Alphabet/Meta new vintages as annualized, not disclosed full-year.
3. Fiscal calendars differ. Microsoft’s FY26 ends June 2026; Oracle’s FY25 ends May 2025; calendar names use December fiscal years. Trajectory labels align periods, not months.
4. Oracle’s +2.4 pp versus the July post is partly a restatement of the prior estimate, not solely new spending. True period momentum for Oracle is better read from the FY26 guide marker (~41%) than from the FY25 restatement alone.
5. Lease accounting and partner capacity can move compute off the PP&E line. Intensity understates committed compute where operating leases dominate.
6. FCF margin here is (operating cash flow − capex) ÷ revenue. Definitions that add back SBC or exclude working-capital swings will differ.

Primary sources: Microsoft FY26 Form 10-K; Amazon, Alphabet, and Meta H1 2026 Forms 10-Q; Oracle FY25 Form 10-K (restated figures); prior theme baseline in capex intensity research. Sector reference bands (telecom ~20%, pre-AI cloud ~11%) carry forward from FCC ARMIS / TIA historical aggregates used in that post.

What to watch into year-end 2026

Three coincident signals will tell you whether the vintage delta stabilizes or accelerates: (1) full-year 2026 intensity for Amazon, Alphabet, and Meta versus these H1 annualizations — a print below the run-rate would be the first mean-reversion signal; (2) FCF margins, especially Amazon and Oracle — if intensity holds above 20% while FCF stops compressing, markets will treat industrial-level reinvestment as a new steady state; (3) guidance language that converts “multi-year build” into explicit intensity ceilings. Until then, the live number is not July’s 21.8% weighted print. It is 24.1% — and the slope is still up.