MTS

Wiki/Operators

Hyperscaler Capital Expenditure: A Primary-Source Reference

Scope: Microsoft, Meta, Alphabet/Google, Amazon, Oracle, with selective notes on Apple and ByteDance. Every figure below is sourced from SEC filings, issuer press releases, earnings transcripts, or contemporaneous wire/newspaper reporting on bond prospectuses and SPV structures. Numbers reflect filings available through Q1 2026 earnings season.

The headline story — half a trillion dollars of annual AI infrastructure spend by 2026 — is well covered. The more interesting material is in the footnotes: useful-life elections that reverse themselves within 24 months, finance-lease "not yet commenced" disclosures that dwarf reported PP&E, and SPV structures that route tens of billions of build cost around the parent's balance sheet entirely.

1. Quarterly Capex by Company (USD billions)

Microsoft fiscal year ends June 30; Meta, Alphabet, Amazon, Apple use calendar year; Oracle fiscal year ends May 31. "Capex" is purchases of property and equipment per cash flow statement unless otherwise noted. Meta's reported capex includes principal payments on finance leases — a definitional choice peer companies do not follow, and material when comparing.

Period MSFT (cash) META (incl. fin. lease princ.) GOOGL AMZN ORCL
CY22 Q1 / FY22 Q3 6.3 5.46 9.69 14.32 0.49
CY22 Q2 / FY22 Q4 6.6 7.69 6.83 14.16 0.55
CY22 Q3 / FY23 Q1 6.6 9.36 7.28 16.38 0.71
CY22 Q4 / FY23 Q2 6.8 8.92 7.60 ~16.6 0.91
FY22 total / CY22 total ~24.0 (FY22 cash capex on PP&E) 31.43 31.49 ~63.6 4.50 (FY23)
CY23 Q1 / FY23 Q3 6.6 6.82 6.29 13.97 ~1.7
CY23 Q2 / FY23 Q4 8.9 6.13 6.89 ~12.7 2.2
CY23 Q3 / FY24 Q1 9.9 6.76 8.06 12.5 ~1.0
CY23 Q4 / FY24 Q2 9.7 7.57 11.02 14.6 1.3
FY23 / CY23 total 28.1 (MSFT FY23) / 32.2 CY23 27.27 32.25 52.7 6.9 (FY24)
CY24 Q1 / FY24 Q3 14.0 6.40 12.01 15.0 2.3
CY24 Q2 / FY24 Q4 13.9 8.47 13.19 17.6 2.4
CY24 Q3 / FY25 Q1 14.9 9.20 13.06 22.6 4.0
CY24 Q4 / FY25 Q2 22.6 14.42 14.27 27.8 5.0
FY24 / CY24 total 55.7 (FY24) / 65.3 CY24 37.26 52.54 83.0 ~14.0 (FY25)
CY25 Q1 / FY25 Q3 21.4 12.94 17.20 24.3 8.5
CY25 Q2 / FY25 Q4 24.2 16.54 22.45 31.4 8.5
CY25 Q3 / FY26 Q1 19.4 19.20 24.18 ~34.0 8.6
CY25 Q4 / FY26 Q2 23.7 22.14 27.55 ~36.0 n/a
CY25 / FY25 totals 88.7 (FY25 incl. fin. lease) 72.22 91.4 ~125 ~21 (FY26 guide)

Source notes: Microsoft cash capex from FY22–FY25 10-Ks (msft-20250630.htm); MSFT FY25 total of $88.7B includes finance-lease principal and exceeds the $80B target Amy Hood reaffirmed on the Q2 FY25 call (MSFT FY25 Q2 webcast). Meta quarterly figures from earnings releases on investor.atmeta.com; FY25 total $72.22B and Q4 $22.14B from the Q4 2025 release (SEC Ex 99.1). Alphabet quarterly capex from each 10-Q's cash flow statement; CY25 total $91.4B per Q4 2025 release. Amazon quarterly capex (purchases of PP&E net of incentives) from 10-Qs and Jassy's Q4 2024 call language that "the $26.3B Q4 figure is reasonably representative of annualized run-rate" (CNBC, Feb 6 2025). Oracle from press releases on investor.oracle.com; Q1 FY26 $8.5B per Q1 FY26 release.

AI-specific carve-outs

Only Meta breaks out an AI-infrastructure line in its disclosures, and even then only at the cost-of-revenue level via depreciation commentary. Microsoft references "the majority" or "more than half" of capex going to AI server and GPU procurement on calls (Hood, FY24 Q4 and FY25 Q1 calls) but does not segment it in filings. Amazon's Olsavsky said on the Q4 2024 call that "the majority [of capex] is for AWS, to support demand for our AI services." Alphabet does not segment. Oracle effectively is an AI capex line — Catz has publicly said FY26 capex will be "around $35bn, could be a little higher" with capacity going to OpenAI, xAI, Meta, Nvidia, AMD (DCD coverage of Q1 FY26 call).

2. Capex Intensity: Capex / Revenue and Capex / Operating Cash Flow

Company FY22 capex/rev FY23 FY24 FY25 FY26E (street/guidance midpoint)
Microsoft ~12% 13% 22% 28% ~32% (FY26 trending toward $120B+ on ~$310B rev)
Meta 26.9% 20.2% 22.6% 32.7% 41-48% (guide $125-145B on $215-225B rev)
Alphabet 11.0% 10.5% 15.0% 23.0% ~37% (guide $180-190B)
Amazon 12.4% 9.2% 13.0% 17.0% ~21-24% ($200B on ~$870B)
Oracle 9.5% 16.4% ~28% 50%+ ~70-80% (Stargate-driven)

Sources: revenue from each company's 10-K Item 8 income statement; 2026 estimates cross-referenced against CreditSights "Hyperscaler Capex 2026 Estimates" which has Oracle at ~86%, Meta 54%, Microsoft 47%, Alphabet 46%, Amazon 25% as of mid-2026 revisions.

Capex/operating cash flow tells the more pointed story. Meta CY25: capex of $72.2B vs. operating cash flow of $97.3B — a coverage ratio of 1.35x. By Q1 2026, with capex run-rate at $20-25B/quarter and operating cash flow stable around $25B, the company is hovering at coverage near 1.0x. Amazon's CY25 capex of ~$125B against operating cash flow of ~$112B already crossed: free cash flow turned slightly negative for the trailing-twelve-months as of Q3 2025. Microsoft's FY25 capex of $88.7B against $130.7B operating cash flow leaves cushion, but FY26 guidance of >$120B capex against operating cash flow likely ~$140B compresses the margin to ~$20B of free cash — its narrowest in a decade.

3. Useful-Life Policy Changes (and Reversals)

The accounting story splits into two phases: 2022-2024 extensions that flattered operating income, then 2025 partial reversals as AI/GPU obsolescence accelerated.

Co. Change Effective Impact disclosed Source
MSFT Server & network gear: 4 → 6 years FY23 (Jul 2022) Hood: +~$3.7B FY23 operating income; $1.1B in Q1 FY23 alone FY22 Q4 webcast / FY23 10-K MD&A
MSFT Further refinement, certain server gear: 6 → various (some GPUs shorter implicit) FY25 Not separately quantified; disclosed in FY25 10-K under "Change in accounting estimate" MSFT FY25 10-K
GOOGL Servers 4→6, network equipment 5→6 Jan 2023 -$3.9B depreciation FY23, +$3.0B net income FY23; Q1 23 alone -$988M depr / +$770M NI Alphabet Q4 23 release, Q1 23 IR note
META Various servers and network gear 4 → 4.5 years Jan 2022 Modest, disclosed in FY22 10-K META 2022 10-K
META Extended to 5 years (subset) 2023 Disclosed FY23 10-K
META Extended to 5.5 years (subset) Jan 2025 "Approximately $2.9B reduction in 2025 depreciation expense" on assets in service as of 12/31/24 META Q4 24 press release
ORCL Servers 4 → 5 years FY24 Disclosed in FY24 10-K
AMZN Servers and network 5 → 6 years Jan 2024 +$0.9B operating income in Q1 2024; +~$3B full-year benefit AMZN 2023 10-K
AMZN Reversal: 6 → 5 years for "certain servers and networking equipment" Jan 1, 2025 Reduces 2025 operating income by ~$0.7B, cited "increased pace of technology development, particularly in the area of artificial intelligence and machine learning" AMZN 2024 10-K, Feb 7 2025 filing
AMZN Early retirement of subset of equipment Q4 2024 $920M accelerated depreciation Q4 2024 + additional $0.6B 2025 hit AMZN 2024 10-K

Cumulative bookend: the extensions across 2022-2024 added roughly $10-12B/year in combined GAAP operating income across the five hyperscalers vs. counterfactual unchanged lives (deepquarry analysis). Amazon's 2025 reversal is the first crack. Whether Meta and Google follow is the open question — the GPU resale market implies useful economic lives closer to 2-3 years for H100-class accelerators, which is not what any GAAP filing reflects.

4. Off-Balance-Sheet Structures

These are the underexplored mechanisms by which announced "capex" understates the actual capital being committed against a hyperscaler's economic interest.

Meta–Blue Owl–PIMCO Hyperion SPV (Louisiana, October 2025). Approximately $30B financing for the Hyperion campus in Richland Parish, Louisiana. Structure: a Morgan Stanley-arranged SPV in which Blue Owl is the lead equity sponsor alongside PIMCO and Meta itself; the vehicle is capitalized roughly $2.5B equity and $27.3B fully-amortizing senior secured notes due 2049 (per TheValueist's bond-doc summary and PE-Insights). Meta owns ~20% of the SPV's equity and signs a long-term triple-net lease to occupy the campus; lease payments service the bonds. The debt sits on the SPV, not on Meta's consolidated balance sheet. Meta's lease obligation is recognized as an operating or finance lease (treatment not yet definitively disclosed in Meta's filings as of Q4 2025 10-K).

Microsoft / BlackRock / GIP / MGX — Global AI Infrastructure Investment Partnership (GAIIP) announced September 2024. Initial target $30B, expandable to $100B. Acquired Aligned Data Centers in 2025 for ~$40B enterprise value. Microsoft contributes capital as a limited partner; project-level debt sits at the fund and portfolio-company level. Microsoft's filings do not consolidate. (Data Center Knowledge launch coverage).

Microsoft–CoreWeave compute leases. Microsoft's commitments to CoreWeave for Nvidia GPU capacity were disclosed as a master agreement totaling approximately $10B+ over five years. These are accounted for as service / hosted compute arrangements, not on-balance-sheet capex, even though CoreWeave finances the underlying GPUs with junior debt. The capex is on CoreWeave's balance sheet, much of it leveraged at the GPU-secured asset level.

Aggregate "leases not yet commenced" disclosure. This is the highest-leverage off-balance-sheet number in the entire sector. Per Moody's, summarized by Fortune Feb 25 2026 and Bisnow:

  • Top 5 hyperscalers: $662B in signed data-center lease commitments not yet commenced, none on the balance sheet under ASC 842 because the lease period has not begun.
  • MSFT alone: $94.8B in finance leases signed but not commenced as of Mar 31 2025 (FY25 Q3 10-Q).
  • Alphabet: $5.8B near-term + $52.7B long-term future lease payments on data-center leases not yet commenced, commencing 2026-2031, terms primarily 1-25 years (FY25 10-K).
  • Morgan Stanley estimates as much as 20% of the projected $2T 2026-2028 hyperscaler capex will route through finance leases for data center shells, off the parent balance sheet until commencement.

Oracle's lease ramp. Oracle's total lease commitments jumped roughly 150% in calendar 2025 as the Stargate-related capacity comes online; much of this is finance-lease treatment on shells operated by third-party developers (QTS, Vantage, Stack).

5. Stated Capex Guidance: 2025 → 2026 → 2027

Quoting directly from earnings calls and press releases.

Microsoft. Hood, FY25 Q2 (Jan 29 2025): reaffirmed "approximately $80 billion" of FY25 capex. Actual FY25 came in at $88.7B including finance leases (FY25 10-K). For FY26, Hood guided on the Q4 FY25 call (Jul 30 2025) to growth "at a lower rate than FY25" but with quarterly capex of approximately $30B in FY26 Q1 — a 50% YoY increase at the quarterly level despite the "lower growth" framing.

Meta. Q3 2025 call (Oct 29 2025): Susan Li stated "CapEx dollar growth will be notably larger in 2026 than 2025." Q4 2025 release (Jan 28 2026): formal 2026 guidance of $115-135B capex including principal payments on finance leases (Meta Q4 25 release). Q1 2026 call (Apr 29 2026): guidance raised to $125-145B — a $10B mid-point increase one quarter into the year (Fortune, Apr 29 2026).

Alphabet. 2025 actual: $91.4B (guidance was $75B at Q4 2024 call; raised mid-year). 2026 guidance Q4 2025 call: $175-185B. Raised at Q1 2026 call to $180-190B. Ashkenazi (CFO) on Q1 26 call: 2027 capex "will significantly increase" relative to 2026 (Alphabet Q1 26 release).

Amazon. Jassy at Q4 2024 call: $26.3B Q4 figure is "reasonably representative of an annualized rate" → ~$105B 2025 (actual $125B). Q4 2025 call (Feb 2026): **$200B in 2026 capex**, "majority in AWS" (TBPN summary of Jassy 2026 shareholder letter). Olsavsky language on multiple calls: capacity "monetized as fast as we can install it"; references "expected useful life of these assets" as the relevant payback frame.

Oracle. Catz Q1 FY26 (Sep 9 2025): FY26 capex "around $35 billion, could be a little higher" with cloud infrastructure revenue ramping to $18B FY26 → $32B → $73B → $114B → $144B over the following four years (Oracle Q1 FY26 release). Subsequently raised to $50B FY26 capex per November 2025 commentary citing Meta and Nvidia contracts.

6. Debt Issuance to Fund the Build

A defining 2025-26 dynamic: the cash-rich hyperscalers are terming out funding in long-dated debt despite having balance-sheet cash, because the capex schedule is now too large to fund from operating cash alone.

Issuer Date Size Structure Use of proceeds Source
Meta Oct 30 2025 $30.0B (largest IG USD deal of 2025, second-largest ever behind Pfizer's $31B in 2023) 6-tranche: $4B 4.20% 5y (T+50), $4B 4.60% 7y (T+70), $6.5B 4.875% 10y (T+78), $4.5B 5.50% 20y (T+88), $6.5B 5.625% 30y (T+98), $4.5B 5.75% 40y (T+110). Order book peaked at $125B (record). Rated Aa3/AA-. "Term out funding for AI and data-centre capex" Bloomberg, Oct 30 2025; The DESK
Alphabet Apr 2025 $5B inaugural USD IG (4-tranche 5y/10y/30y/40y) First-ever Alphabet public bond AI capex, general corporate
Alphabet Jun 2025 ~$10B incl. debut 100-year tranche (£500M sterling century bond) Multi-currency: USD, EUR, GBP AI capex CNBC Feb 12 2026 retrospective
Alphabet Feb 9 2026 $20B USD 7-tranche (biggest ever Alphabet USD deal) plus follow-on €3B EUR and CHF debut 5/7/10/20/30/40-year USD tranches "Fund $185B 2026 capex program" Bloomberg Feb 10 2026
Oracle Sep 2025 $18B Multi-tranche IG Stargate-related capacity; Oracle is now largest non-financial IG issuer of 2025 per Citi DCD coverage
Oracle Nov-Dec 2025 PIMCO-led $14B private placement / term loan Senior secured private debt AI cloud capacity buildout TradingKey on Oracle/PIMCO
Oracle Late 2025 Up to $38B in syndicated loans being prepped per Octus reporting Bank loans AI buildout Octus
Meta-Hyperion SPV (non-recourse) Oct 2025 $27.3B senior secured notes due 2049 Fully amortizing project-level debt Hyperion campus build PE-Insights
Microsoft n/a Has not issued IG bonds in size since 2021 (had no need given $130B+ OCF) n/a Funds capex from operations and short-term commercial paper
Amazon Nov 2025 $12B IG 6-tranche 2/5/10/30y bonds General corporate; AWS capex AMZN 10-Q footnotes

The aggregate AI-related IG issuance in 2025 hit a record per Mellon Investments. The defining marker: even Microsoft, which has not needed bond markets in nearly five years, was reported in late 2025 to be preparing potential debt issuance for 2026 — a structural shift for a company that had been a net debt repayer for the entirety of the post-COVID period.

7. Free Cash Flow Conversion: Where the Math Breaks

Co. CY/FY25 OCF CY/FY25 Capex (cash, incl. fin. lease where reported) FCF Coverage (OCF/Capex)
MSFT FY25 $130.7B $88.7B (incl. finance lease princ.) ~$42.0B 1.47x
META CY25 $97.3B $72.2B $46.1B (Meta's reported, after fin. lease princ. subtraction) 1.35x
GOOGL CY25 $130.1B $91.4B $38.7B 1.42x
AMZN CY25 $112.4B ~$125B ~ -$12B 0.90x — already negative
ORCL FY25 ~$20B ~$21B ~$0 ~1.0x

Amazon is already past the crossover. Per CNBC's Feb 2026 piece on Big Tech cash, the projection for 2026 has Amazon FCF firmly negative under the $200B capex guide. Meta's cash balance fell from ~$70B in late 2024 to $44B by Q3 2025 (Meta Q3 25 release) — funding the gap before the October $30B bond. By 2026 guidance midpoints, at least three of the five hyperscalers will have capex equal to or exceeding operating cash flow on a TTM basis for the first time in their public-company histories.

The Q3 2025 Meta print is the cleanest tell of where the squeeze hits first: operating cash flow $30.0B, capex $19.2B → quarterly FCF still positive at ~$10.6B, but the YoY trend on FCF is -50%+ even while revenue grew 26%. Meta has not reported a negative-FCF quarter (this remains an outstanding research question for 2026 quarters as guidance escalates).

8. Underexplored Mechanics

Operating-lease to finance-lease shift. Hyperscalers historically treated long-duration data-center shells as operating leases, with the right-of-use asset and lease liability roughly netting. Increasingly, the same shells are being structured as finance leases — usually because the lessee gains effective ownership economics (purchase options, fair-value risk, term covering most of useful life). The accounting effect: depreciation flows through capex/PP&E rather than rent through opex. Per Investing.com summary of Morgan Stanley research, this also lets the lessee fund the build with bank or developer debt that does not consolidate while still capitalizing the asset. Meta's reported "capex including finance-lease principal payments" already runs $10-15B/year above its cash-only capex — i.e. roughly 20-30% of headline 2025 capex is finance-lease draws, not cash outlays.

Announced vs. operational gap. Stargate's headline number ($500B by 2029) overstates actual installed capacity by an order of magnitude through 2026. Oracle's operational OCI capacity in FY26 is guided to $18B revenue against a $455B RPO — implying 25x backlog-to-revenue. The same gap exists at smaller scale for Microsoft (contracted Azure backlog north of $300B against current annualized run-rate ~$140B for Azure & other cloud services per FY26 Q1 release). The relevant question for the wiki: how much of the "$2T 2026-2028 AI capex" is signed-but-not-built vs. powered-and-installed?

GPU resale-implied lives vs. GAAP lives. CNBC Nov 14 2025 covers this directly. Secondary market on H100 clusters showed prices falling 30-40% within 12 months of B100/B200 shipment — implying useful economic life of 2-3 years for the current-generation accelerator inside the host server. GAAP server lives at MSFT (effectively 6y blended), GOOGL (6y), META (5.5y), AMZN (5y post-reversal) are all materially longer than the resale curve implies. Amazon's reversal to 5y is the only formal acknowledgment to date. Michael Burry's Scion 13F-reported short positions in late 2025 cited this gap. If MSFT or META are forced to true up, each year of life-shortening on the AI server fleet would cost approximately $3-5B in incremental annual depreciation (back-of-envelope: $50-80B of AI servers x 1/5 yr - 1/6 yr).

9. Apple and ByteDance — brief notes

Apple. Capex remains structurally low (~$10B/year on PP&E in FY25 per 10-K). Apple Intelligence inference workloads run substantially on partner infrastructure (Microsoft Azure, Google Cloud per multiple reports). Apple's "Private Cloud Compute" Mac-based servers are an internal capital deployment but the dollar magnitude is not separately disclosed and is small relative to peers — Bloomberg estimates "low single-digit billions per year." This is the most capital-efficient AI strategy in the hyperscaler set and a useful counter-anchor.

ByteDance / TikTok. Private; capex figures come from press leakage and supplier filings (notably Nvidia 10-Q customer concentration disclosures, where ByteDance is repeatedly named alongside Microsoft, Meta, Amazon as a top-4 customer). Reported 2025 AI infrastructure spend of $20-25B per FT and Reuters, with much routed through ByteDance's Singapore and U.S. entities; not directly comparable to U.S. hyperscaler disclosures because of differing scope (consumer-app inference dominates vs. external cloud). Worth noting in the wiki only as an Nvidia-demand reference point, not as a comparable on capex disclosure.

10. Key Filings and Documents Index

Last updated: May 2026. Quarterly figures should be refreshed each earnings cycle; bond market data is most useful when paired with current OAS spreads vs. issue date.