The Hyperscaler Capex Wall: $434B of Buyer-Side AI Spend, the Depreciation Lag, and Why Big Tech Borrows

Four companies bought $433.9 billion of property and equipment in the last four reported quarters. That is an SA estimate only in the narrowest sense — it is the sum of four numbers each company printed in its own SEC cash-flow statement. Nobody disputes the inputs. What gets missed is what the number does next: it becomes depreciation, on a five-to-six-year fuse for servers and a 25-to-40-year fuse for buildings, and the income statements of 2027-2029 will carry it whether or not AI revenue shows up on schedule.

This note works through the buyer side of the AI ledger from primary filings: the quarterly capex record for Microsoft, Alphabet, Amazon, and Meta from 2022 through 1Q26; the widening capex-versus-depreciation gap and the useful-life policies underneath it; capex as a share of operating cash flow; the dated record of who borrowed what, when; and what all of it means for the semiconductor supplier chain. Every reported figure is cited to the filing it came from. Everything we computed is labeled an SA estimate.

A note on definitions: "capex" below means the cash-flow-statement line "purchases of property and equipment" as each company reports it, excluding finance leases unless stated. Microsoft's June fiscal year is mapped to calendar quarters throughout (its "FY26 Q3" is calendar 1Q26). The full mapping and definitional caveats are in the Methodology section at the end.

The one-sentence read: The hyperscalers are spending at nearly 3x the rate their income statements currently recognize, the accounting that made that gap painless has stopped getting more generous, and the borrowing that started in late 2025 is how they intend to keep building through the crossover.

1. The number: $434B in four quarters, and still accelerating

Combined purchases of property and equipment, trailing four quarters through March 2026 (SA estimate: sum of reported figures):

All figures from the companies' 10-Q/10-K cash-flow statements (Microsoft FY26 Q3 10-Q; Alphabet, Amazon, Meta Q1 2026 10-Qs, all filed late April 2026). Two features stand out. First, the level: for calendar 2022 the same four companies spent $151.1B — the run-rate has nearly tripled in three years and now exceeds the GDP of most countries. Second, the slope: combined 1Q26 capex of $129.8B was up 80% from $71.9B a year earlier, and each of the last five quarters set a new combined record.

Combined quarterly capex went from ~$36B (1Q22) to $129.8B (1Q26). The inflection is mid-2024 — and the slope steepened again in late 2025.

Source: Company 10-Q/10-K cash-flow statements via SEC EDGAR, 1Q22-1Q26. Discrete quarters computed from year-to-date figures where companies report YTD (SA computation). Microsoft fiscal quarters mapped to calendar quarters (June fiscal year-end). Meta 2022-2023 values reflect Meta's current restated presentation.

And the forward guidance says the wall gets higher. Company-stated 2026 spending indications, each dated to its earnings call: Alphabet guided capex of $175-185B (Feb 4, 2026); Amazon indicated roughly $200B of capital investments (Feb 5, 2026); Meta guided $115-135B in January and raised it to $125-145B on April 29, 2026; Microsoft indicated roughly $190B for calendar 2026 on the same day, including an estimated ~$25B impact from higher component prices — memory in particular. Summed at guidance midpoints that is roughly $700B of intended 2026 spend (SA aggregation for scale only — the four definitions differ; see Methodology).

For context on what a single quarter of this looks like at the bottom of the supply chain: Amazon's 1Q26 purchases of $44.2B alone exceed the annual capex of every semiconductor company on earth except TSMC and Samsung — the companies our supplier-side capex dataset tracks.

2. Capex vs. depreciation: the lag is the whole game

Depreciation is capex on a delay. A dollar of server spend hits the cash-flow statement today and the income statement in roughly equal slices over the next five to six years; a dollar of data-center shell spreads over 25-40 years. When capex triples in three years, reported depreciation mechanically lags — and reported margins look better than the underlying economics until the wave catches up.

The cleanest quarterly comparison the filings allow is Microsoft and Meta:

In 1Q26 Microsoft spent $30.9B against $10.4B of infrastructure depreciation; Meta spent $19.0B against $6.0B of D&A. Both gaps have widened every quarter for two years.

Source: Microsoft 10-Q/10-K: purchases of PP&E; infra depreciation = PP&E depreciation expense + finance-lease right-of-use asset amortization from the notes (quarterly disclosure begins FY25, i.e. 3Q24 calendar; SA-computed sum). Meta 10-Q/10-K: purchases of PP&E and cash-flow D&A line. SEC EDGAR, filings through Apr 2026.

Both companies are spending roughly 3x their current depreciation. Across all four companies, trailing-4Q reported depreciation sums to roughly $149B against $433.9B of capex (SA estimate; definitions differ by company — Amazon's D&A line is the broadest — see Methodology). The gap, roughly $285B a year at the current run-rate, is spend that has been paid for in cash but not yet recognized as cost.

What the useful-life policies say

The lag is partly a policy artifact, and the policy record is worth reading closely because it stopped moving in one direction:

From 2020 through 2024, every change extended lives — each extension deferring depreciation into later years precisely as capex accelerated. In January 2025 the pattern broke: Amazon shortened the life of a subset of its fleet and retired hardware early, citing "an increased pace of technology development, particularly in the area of artificial intelligence and machine learning" — while Meta extended to 5.5 years in the same month. Two sophisticated operators of similar hardware now disclose opposite directional judgments about how long it stays useful.

The bear case sharpened in November 2025 when Michael Burry publicly argued the accelerator fleets inside these totals have an economic life closer to two to three years than the five-to-six-year schedules in use, and that industry depreciation could be understated by ~$176B over 2026-2028. We do not endorse the specific figure — it is not verifiable from public filings. But the direction of the risk is asymmetric: nobody is arguing GPUs last longer than the schedules assume, and one of the four operators has already conceded the point for part of its fleet with cash consequences (accelerated depreciation on early retirements).

The wave is already visible where disclosure is granular

Meta's 2025 10-K discloses depreciation on servers and network assets specifically: $7.32B (2023) → $11.34B (2024) → $13.36B (2025) — up 83% in two years despite the January 2025 life extension that deferred ~$2.9B of it. Alphabet's PP&E depreciation went $11.9B (2023) → $15.3B (2024) → $21.1B (2025), and annualizes to ~$26B off 1Q26 (SA computation from quarterly disclosures). The wave is not hypothetical; it is compounding at 30-40% a year while capex compounds faster.

SA estimate — the forward D&A run-rate, stated assumptions: take the $433.9B trailing-4Q cohort alone, assume ~85% is depreciable equipment and shorter-life assets on a blended ~6-year straight-line schedule and the rest is 25-40-year shell, and that single year of spend adds roughly $65-75B of annual depreciation once fully in service — on top of the ~$149B already being recognized, and before the larger 2026 cohort implied by ~$700B of guidance. If server lives are economically closer to four years than six (the Amazon-direction scenario), the same cohort implies $95-110B. This is an illustration of operating leverage in reverse, not a forecast of reported figures: actual recognition depends on in-service timing, mix, and future policy changes.

3. Cash flow covers it — except at Amazon

The standard defense of the build-out is that it is funded from operations, not from the balance sheet. For three of the four, the filings support that. Capex as a share of operating cash flow, 1Q26 (SA-computed ratio; inputs from the cash-flow statements cited above):

Amazon spent $1.70 in capex for every $1.00 of operating cash flow in 1Q26. The other three remain comfortably below 80%.

Source: SA-computed ratio from 1Q26 10-Q cash-flow statements (SEC EDGAR): Meta $19.0B/$32.2B; Microsoft $30.9B/$46.7B; Alphabet $35.7B/$45.8B; Amazon $44.2B/$26.0B. Amazon's Q1 operating cash flow is seasonally weak (working-capital timing); its trailing-4Q ratio is 102%.

Single quarters flatter and distort — Amazon's Q1 operating cash flow is always seasonally depressed by working capital. The trailing-four-quarter picture (SA-computed):

Amazon crossing 100% matters. It means every dollar of AWS and retail operating cash generation is going into the ground, plus borrowed money — the first time Amazon's trailing capex has exceeded trailing operating cash flow since 2022, when the overbuild of pandemic-era fulfillment capacity forced a multi-quarter capex retrenchment (Amazon's cash-flow statements, 2022-2023: annual capex fell from $63.6B to $52.7B). That precedent is the strongest public evidence that these companies will cut when the return math stops working — and it is exactly the scenario the supplier chain should track.

The other three still generate $47-77B a year of operating cash flow above their capex. The build is expensive; for Microsoft, Alphabet, and Meta it is not yet strained.

4. Why borrow, then? The dated record — and the logic

If operating cash flow covers the spend for three of the four, the bond market activity needs explaining. The record first, every entry dated and sourced:

~$97B of AI-linked paper hit the market in 4Q25 alone. Two years ago the quarterly figure was zero.

Source: SA compilation of dated issuances: Bloomberg (Meta Aug 2024, Oct 2025; Oracle Sep 2025; Amazon Nov 2025), Alphabet Q2'25 8-K exhibit, Cleary Gottlieb closing notice (Alphabet Nov 2025, EUR converted at ~$1.15), CNBC (Alphabet Feb 2026 multi-currency total), Amazon 424B5 (Jul 2026), Meta press release (Hyperion JV, off-balance-sheet SPV debt as press-reported). Principal issuances; not exhaustive of commercial paper or private placements.

Microsoft is the notable absence: no senior bond issuance in the window. Its equivalent lever is leasing — Microsoft's filings show $26.0B of finance-lease right-of-use assets added over the same trailing four quarters (SA computation from 10-Q/10-K lease notes), capacity that arrives as debt-like lease liabilities rather than bonds.

The SA view: four reasons cash-rich companies borrow

This section is analysis, not reporting. The filings tell you that they borrowed; the why below is our read.

1. Duration matching, at spreads that may not last. A data center is a 25-40-year asset; a 40-year bond against it is textbook asset-liability matching. Meta's Oct 2025 deal drew ~$125B of orders; Alphabet found demand for a 100-year sterling note. Order books like that mean tight spreads — and issuers appear to be taking duration while the market is still eager. The Amazon July 2026 book (peak ~$62B, pared to ~$41B, ~1.6x covered vs. richer coverage on earlier deals) is the first hint that eagerness is normalizing.
2. Protecting buybacks and dividends. All four now return capital (Meta and Alphabet initiated dividends in 2024). Funding 100% of an accelerating build from operating cash flow would force the return-of-capital programs to absorb the volatility. Debt decouples the two — shareholders keep their programs, bondholders fund the concrete.
3. The SPV lever: capacity without the balance sheet. The Meta-Blue Owl Hyperion structure puts ~$27B of project financing in an entity Meta owns 20% of — the debt does not consolidate onto Meta's balance sheet, and the rating agencies treat it differently than senior notes. If the model spreads (and Oracle's obligations around leased capacity suggest the broader industry is already comfortable with off-balance-sheet exposure), reported debt will understate the sector's true fixed commitments. Watch for more of these: they are disclosed in press releases and lease-commitment footnotes, not in the debt tables.
4. Optionality against the depreciation wall. This is the reason that ties the article together. Management teams can see Section 2's math as well as anyone: recognized costs will rise steeply into 2027-2028. Borrowing now, at investment-grade spreads, pre-funds the build without betting the buffer cash on AI revenue arriving exactly on schedule. If revenue lands, the debt is trivial against $124-174B cash-flow streams. If it disappoints, the alternative — having drained the cash pile first — is strictly worse.

5. What it means for semis: the buyer side of the ledger

Every dollar in Section 1 is revenue for someone in the supplier chain — the foundries printing the accelerators, the memory makers feeding them, the packaging houses assembling them, and the power and cooling vendors around them. Hyperscaler capex is the single best forward indicator of demand for advanced-node wafer capacity and the cost stack we model in the AI Chip Cost Bridge.

Two sides of the same ledger, and worth keeping distinct:

• This article is the buyer side — what the customers of the semiconductor industry spend to deploy compute.
• Our semiconductor capex dataset is the supplier side — what foundry, memory, and equipment players spend to build the fabs that serve that demand. The two series answer different questions: buyer-side capex tells you whether demand holds; supplier-side capex tells you whether supply overshoots.

The read-through, labeled as SA view:

• Durability, near term: strong. Guidance summing to ~$700B for 2026 was raised twice this spring (Meta in April; Microsoft's component-price adjustment the same week). Debt pre-funding means the 2026 build is largely financed already — a spender who has already sold the bonds is unlikely to cancel the servers the proceeds were earmarked for. For the memory complex specifically, Microsoft attributing ~$25B of its 2026 number to component-price inflation is a demand signal wearing a cost costume — see our memory capex and qualification coverage.
• Durability, 2027-2028: watch the depreciation crossover. The mechanism from Section 2 is the credible bear case. As recognized D&A compounds 30-40% annually, hyperscaler margins compress mechanically; CFOs facing margin questions historically respond by moderating capex growth (Amazon 2022-2023 is the template, and Amazon is already spending past its cash generation). The supplier chain should treat hyperscaler useful-life disclosures — buried in 10-K estimate-change paragraphs every January and February — as a leading indicator with more signal than any keynote.
• The equipment analogy is exact. Fabs depreciate tools over 5-7 years and the industry has learned, repeatedly, what happens to pricing when a capacity cohort's depreciation tail meets a demand air pocket — see TSMC's 1Q26 capacity-rule break for how the leading foundry manages exactly this. The hyperscalers are now running the largest such experiment in capital-goods history, with shorter asset lives than they assume if the skeptics are right.
• Concentration cuts both ways. Four buyers now control a capex pool (~$434B trailing) larger than the entire semiconductor industry's annual capex. That is extraordinary pricing power for suppliers while the race lasts — and extraordinary cliff risk if even one buyer blinks. The 2022 Amazon retrenchment took roughly $11B of annual capex out of the system; a proportional 2027 move at today's scale would be a $30-40B demand shock (SA illustration).

6. Methodology

Capex definition. "Capex" = the cash-flow-statement line for purchases of property and equipment as reported: Microsoft "additions to property and equipment"; Alphabet "purchases of property and equipment"; Amazon "purchases of property and equipment" (gross, before proceeds from sales/incentives); Meta "purchases of property and equipment" under Meta's current presentation (Meta reclassified this line in later filings; 2022-2023 quarters here reflect the restated values, which differ from the originals by ≤$0.3B/year). Finance leases are excluded from all headline figures. Where we cite Meta's company-defined "$72.2B 2025 capex," that figure includes principal payments on finance leases ($69.7B of PP&E purchases + $2.5B of lease principal — both from the 10-K; the reconciliation is an SA computation that matches Meta's stated total). Microsoft's lease-financed additions are quantified separately in Section 4 as an SA adjustment ($26.0B trailing 4Q of finance-lease right-of-use assets added, from the lease notes).

Microsoft calendar mapping. Microsoft's fiscal year ends June 30. Mapping used for every Microsoft data point: FY22Q3 = 1Q22, FY22Q4 = 2Q22, FY23Q1 = 3Q22, and so on through FY26Q3 = 1Q26. All Microsoft figures labeled with calendar quarters in this article follow this mapping.

Quarterly derivation. Companies report cash-flow statements year-to-date in 10-Qs. Discrete quarters are SA computations: Q(n) = YTD(n) − YTD(n−1), using each figure's original filing (or the latest restated presentation where a company recast prior periods). Full per-value provenance (form, accession number, filing date) was retained in our working dataset.

Depreciation definitions (they differ — do not cross-compare levels). Meta: cash-flow "depreciation and amortization." Microsoft: PP&E depreciation expense plus finance-lease right-of-use asset amortization from the 10-Q/10-K notes (an SA-computed sum; Microsoft's cash-flow line "depreciation, amortization, and other" includes non-D&A items and was recast in FY26 filings, so we do not use it). Alphabet: PP&E depreciation from the notes. Amazon: cash-flow D&A line, which also covers capitalized content and operating-lease amortization — the broadest of the four; it overstates infrastructure depreciation, meaning the true combined capex-to-infrastructure-depreciation gap is wider than our ~$149B figure implies.

Combined and ratio figures. Every multi-company sum, ratio, and the trailing-4Q $433.9B headline are SA estimates (sums/quotients of reported figures). The ~$700B 2026 guidance aggregate mixes company definitions (Meta's includes finance-lease principal; Amazon's "capital investments" is broader than cash PP&E) and is presented for scale, not precision.

What we dropped. Values we could not tie to a primary or clearly reliable source were dropped rather than estimated — including a press claim that Amazon had raised ~$54B of bonds year-to-date in 2026 (inconsistent with the issuance record we could verify) and the precise internal split of the Hyperion SPV financing (press-reported; labeled as such above).

The one-paragraph takeaway

The four hyperscalers put $433.9B into property and equipment in the four quarters through March 2026 (SA sum of reported figures) — 80% more in 1Q26 than a year earlier, with 2026 guidance implying roughly $700B more — while their income statements currently recognize only ~$149B a year of depreciation against it. That gap is a lag, not a subsidy: five-to-six-year server schedules mean 2027-2029 absorbs today's build as cost, the useful-life extensions that softened prior waves have stopped (Amazon has already reversed one), and the borrowing spree — Meta's $30B, Alphabet's ~$56B across three deals, Amazon's $40B across two, plus a ~$27B off-balance-sheet SPV — is best read as management pre-funding the crossover rather than celebrating the boom. For the supplier chain the message is double-edged: the 2026 build is financed and committed, which protects near-term foundry, memory, and packaging demand — but the depreciation wall is now the most important dated, mechanical, public-filings-visible risk to the 2027-2028 order book, and the place it will show first is a one-paragraph estimate-change disclosure in somebody's annual report.

References & Sources