Hyperscaler Capex Cycle

RECONCILED 2026-07-22 — the $610–725B "conflict" fully dissolves; three axes explain all of it. April-vintage highs: 190 (GOOGL) + 200 (AMZN) + 145 (META) + 190 (MSFT calendar) = $725B. February-vintage lows with Microsoft on a fiscal basis: 175 + 200 + 115 + 120 = $610B. The whole spread is (i) guidance vintage — three of four raised in April, (ii) Microsoft fiscal-vs-calendar (~$70B on one company), (iii) basket membership (four names vs five with Oracle). This page's "the headline exceeds the sum of its own parts" flag was right against the parts it was holding and wrong against current ones — the parts were stale, not the arithmetic. Alphabet is also stale here: $175–185B is the February guide, superseded 2026-04-29 by $180–190B with 2027 to "significantly increase" [primary: GOOGL 8-K Ex-99.1, 2026-06-04]. Current management-guided, calendar-year, lease-inclusive lines: GOOGL $180–190B · AMZN ~$200B · META $125–145B · MSFT $190B.

RESOLVED 2026-07-22 — the Microsoft $190B vs $120B conflict was a DEFINITION ERROR, not a disagreement. Primary, Amy Hood on the FY26 Q3 call 2026-04-29: "For calendar year 2026, we expect to invest roughly $190 billion in capital expenditures" [primary: MSFT earnings call, 2026-04-29] — calendar year, including finance leases, accrual basis (NOT the cash-flow-statement line, which cannot reproduce it). The "$120B+" figure has three candidate origins, all landing at $113–118B and none of them calendar-2026: (1) most likely a six-month residual quoted as a year — an analyst on that same call said "maybe to the tune of $120 billion" for the back half; $190B − $31.9B actual − $40B guided = $118.1B; (2) CY2025 actual $118.0B, which reconciles exactly with the "+61% vs 2025" framing in circulation; (3) fiscal-year AND cash-only ($113B). Note fiscal-vs-calendar alone does not explain it. Desk rule: every capex figure must carry (period basis, lease treatment) or it is not comparable. For a big-four aggregate of management-guided, calendar-year, lease-inclusive capex, Microsoft's line is $190B. Trap: a real Microsoft number sits near $190B that is NOT capex — $196.6B of leases signed but not yet commenced (10-Q 2026-03-31), a five-year commitment stock, not an annual flow.

CORRECTION 2026-07-22 — Meta's figure below is STALE, verified against primary. Every source in this topic quotes Meta at $115–135B. That is the FY2025 10-K number. Meta raised guidance to $125–145B on 2026-04-29, restated verbatim in the Q1-2026 10-Q MD&A: "capital expenditures of approximately $125 billion to $145 billion in 2026" [primary: META 10-Q, 2026-04-30]. Calendar year, includes finance-lease principal. Consequence: the reconciliation below says "the entire big-four gap is Microsoft, because Amazon, Alphabet and Meta are stated identically by both sources." Both sources agree on Meta because both are stale on Meta — agreement between two secondary sources of the same vintage is not corroboration. Every aggregate here understates Meta by $10B at both ends. Meta's run-rate is the live question: Q1-2026 spent $19.84B = 14.7% of the $135B midpoint (a flat year is 25%), so the bottom of its own range needs ~1.77x the Q1 quarterly rate held for three quarters. Prints 2026-07-29.

The hyperscaler capex cycle is the synchronized, debt-and-cash-funded capital spending of the four largest US cloud operators (Amazon, Microsoft, Google/Alphabet, Meta) on AI infrastructure. In 2026 that combined spend is guided to ~$725B, up from ~$410B in 2025 — a +77% YoY jump — with a 2027 projection of $1 trillion+. This is the demand engine that sets the tape for every supplier above it in the stack: GPU vendors, datacenter shells, networking, cooling, and power. The defining behavioral driver, per Value Add VC, is asymmetric fear: "each company fears being structurally short on compute more than it fears overspending."

The spend is not monolithic — it splits roughly in half between compute & silicon (~50%) and everything that houses and powers it. About ~$260B+ goes to Nvidia GPUs and ~$80B+ to custom silicon (the hyperscalers' own accelerators), with the remaining ~50% flowing into datacenter plant (~35%) and power & grid (~15%). That allocation is the key to reading the cycle as an investment: the GPU line is the headline, but the plant and power lines are where the binding constraints — and the supply-chain re-rate — actually live (see Power: The Binding Constraint).

A second source (Data Center Knowledge) independently pegs 2026 hyperscale capex at >$600B. The two figures differ in scope — the ~$725B number is the big-four guidance total, the >$600B is a broader hyperscale framing — but they agree directionally that 2026 is a step-change spending year. The open tradable question the cycle raises is whether and when it inflects down, and what the leading indicator would be.

Every aggregate on this page is a forecast. These are guidance totals and analyst estimates for a year that is barely half over — not measured spending. No source in this KB reports a summed actual 2026 capex figure, and none can until the year closes.

The 2026 Aggregate — Estimate Dispersion (added 2026-07-22)

A third source (Futurum, Feb 2026) puts 2026 capex at $660–690B across five companies (the big four plus Oracle), against a ~$380B 2025 base. Reconciling it against the Value Add VC figures already on this page produces three distinct discrepancies, none of which the sources resolve:

Three things follow. First, the entire big-four gap is Microsoft — Amazon ($200B), Alphabet ($175–185B — STALE, actual $180–190B) and Meta ($115–135B — STALE, actual $125–145B, see correction above) are stated identically by both sources; only Microsoft differs, and it differs by roughly the size of the whole discrepancy. Second, the ~$725B headline exceeds the sum of its own per-company parts: Value Add VC's four figures sum to $680–710B, $15–45B short of its own $725B headline — an internal reconciliation gap in the source this KB has been quoting. Third, the vintage difference (Feb vs Jun 2026) plausibly explains part of the Microsoft gap, since Futurum's "$120B+" is open-ended and predates two quarters of upward revisions — but no source states that reconciliation, so the KB records the conflict rather than picking a number.

Working position: treat $610–725B as the honest band for 2026 big-four capex and stop quoting ~$725B as a settled figure. The band closes only when summed actual quarterly capex is available.

The prints arrive — the demand side has stopped self-funding (added 2026-07-23)

The first Q2 2026 primary prints turn this page's central open question — is the buildout financed from operating cash flow or from the balance sheet? — from hypothesis into an on-print fact for two named buyers. The demand side of the buildout has stopped self-funding. Every capex figure here is stated with its (period basis, lease treatment) per the desk rule.

• Alphabet — Q2 capex $44.9B (calendar-Q2 2026 actual; cash purchases of PP&E from the cash-flow statement — excludes finance-lease additions) outran operating cash flow $39.1B, so quarterly free cash flow went NEGATIVE $(5.9)B. The gap was plugged from capital markets, not operations: a $49.6B June equity raise explicitly earmarked "to scale AI infrastructure," plus $20.3B of senior notes; long-term debt more than doubled to $98.2B in six months. [primary: GOOGL 8-K Ex-99.1, 2026-07-22]
• Tesla — Q2 capex $5.79B (+142% YoY) (calendar-Q2 2026 actual; cash PP&E purchases) outran OCF $4.7B, so FCF went NEGATIVE $(1.1)B — its first negative-FCF quarter since Q1 2024. Tesla did not tap capital markets; it drew its cash pile down $1.2B to $43.5B. Same signal, different funding valve. [primary: TSLA 8-K Ex-99.1, 2026-07-22]

Two things this settles and one it does not. It settles the financing-fragility question (previously open on this page): for the two names that have printed, capex now exceeds internal cash generation, and the buildout is carried by the balance sheet — Alphabet by external capital, Tesla by cash drawdown. It partially breaks this page's "every aggregate is a forecast" caveat: there are now actual quarterly capex figures on the tape for two buyers, even though the summed-year big-four aggregate is still unmeasured. It does not resolve the FY guide — Alphabet's Ex-99.1 carries no FY2026 capex dollar guide and no RPO/backlog line (both are call/10-Q items). The Q2 run-rate ($44.9B) annualizes to ~$180B and H1 ($80.6B) to ~$161B, but the printed release neither confirms nor breaks the pre-print's watched $180–190B — mark it unresolved by the print and do not invent it.

The supply-side mirror (TXN). Texas Instruments — a seller into the buildout, now naming "data center" a growth leader — printed the opposite profile: revenue $5.46B (above the $5.00–5.40B guide top), its own capex falling ($4.94B → $3.31B TTM as its 300mm fab wave crests), and free cash flow +271% YoY to 33.6% of revenue (definition includes CHIPS Act incentive proceeds). The seam this topic tracks, now visible on one tape: the buyers' funding gaps widen while the equipped, subsidised supplier harvests the cash. [primary: TXN 8-K Ex-99, 2026-07-22]

Key Claims

• 2026 combined big-four capex ~$725B, up from ~$410B in 2025 (+77% YoY). Evidence: moderate (analysis — Value Add VC) (AI Hyperscaler Capex 2026)
• 2027 projection: $1 trillion+. Evidence: moderate (analysis projection) (AI Hyperscaler Capex 2026)
• Per-company 2026 guidance: Amazon $200B (+100% YoY) · Microsoft $190B (+100%) · Google/Alphabet $175–185B (+110%) · Meta $115–135B (+80%) — superseded: $125–145B per primary, see correction. Evidence: moderate (analysis — Value Add VC) (AI Hyperscaler Capex 2026)
• Allocation: ~50% compute & silicon (~$260B+ Nvidia GPUs + $80B+ custom silicon), ~35% datacenter plant ($170B shells/land + $85B networking/cooling), ~15% power & grid ($105B generation, PPAs, nuclear contracts). Evidence: moderate (analysis — Value Add VC) (AI Hyperscaler Capex 2026)
• Behavioral driver: "each company fears being structurally short on compute more than it fears overspending." Evidence: moderate (analysis framing) (AI Hyperscaler Capex 2026)
• Cross-check — a second 2026 estimate puts hyperscale capex at >$600B (different scope, directional agreement with the ~$725B big-four figure). Evidence: moderate (multi-source agreement) (Hyperscalers in 2026)
• Hyperscale capacity expected to double in ~12 quarters — the physical correlate of the capex curve. Evidence: moderate (analysis — Data Center Knowledge) (Hyperscalers in 2026)

Added 2026-07-22

• [FORECAST] Third estimate: $660–690B 2026 capex across five companies (big four + Oracle) vs a ~$380B 2025 base. Evidence: moderate (analysis — Futurum; third independent estimate of the same cycle) (Futurum AI Capex 2026)
• CONFLICT — big-four 2026 aggregate: ~$725B vs $610–640B (derived). The lower five-company total implies a big-four figure $85–115B below the Value Add VC number. Unresolved. Evidence: n/a (unresolved conflict between two analysis sources) (Futurum); (Value Add VC)
• CONFLICT — Microsoft 2026 capex: ~$190B vs $120B+. The single largest per-company disagreement and effectively the whole aggregate gap; the other three names are stated identically by both sources. Evidence: n/a (unresolved conflict) (Futurum); (Value Add VC)
• CONFLICT — 2025 base year: ~$410B vs ~$380B, which changes the implied YoY growth rate. Evidence: n/a (unresolved conflict) (Futurum); (Value Add VC)
• [DERIVED] The ~$725B headline exceeds the sum of its own per-company figures ($680–710B) by $15–45B — an internal inconsistency in the source. Evidence: n/a (arithmetic on the source's own stated numbers) (Value Add VC)
• Oracle now has a 2026 capex figure: $50B, +136% YoY, closing a gap this KB previously flagged as open. Evidence: weak (single analysis source) (Futurum AI Capex 2026)
• Alphabet's guidance was revised upward repeatedly — from an initial $71–73B projection to $175–185B across three revisions, per Futurum; the July preview reports a further raise to $180–190B. Evidence: weak (analysis + summary-derived figures in a paywalled preview) (Futurum); (Alphabet preview)
• Management framing has flipped to supply-constrained, not demand-constrained across the cohort. Evidence: weak (single analysis source characterizing commentary) (Futurum AI Capex 2026)
• [FORECAST] Stargate (OpenAI / SoftBank / Oracle JV) targets $500B in AI infrastructure by 2029 — $100B initial deployment, >$400B in commitments inside the first three years. A multi-year target, not committed spend. Evidence: weak (single analysis source, long-horizon projection) (Futurum AI Capex 2026)
• The cycle is now a macro variable — the Fed attributes most of Q1 2026's +11% annualized business fixed investment growth to AI-services infrastructure. Evidence: moderate (technical-report — Federal Reserve, first-party) (Fed MPR July 2026)
• Capex alone no longer converts to capacity — packaging and memory are sold out alongside power, so an approved capex dollar buys less installed compute than it did. Evidence: moderate (converging across three sources) (TSMC Q2 2026); (Samsung preliminary); (Hyperscalers in 2026)

Added 2026-07-23 (first primary Q2 prints)

• [ACTUAL] Alphabet Q2 2026 capex $44.9B (calendar-Q2 actual, cash PP&E purchases) exceeded operating cash flow $39.1B → free cash flow NEGATIVE $(5.9)B, funded by a $49.6B equity raise + $20.3B senior notes (LT debt doubled to $98.2B). Evidence: strong (first-party 8-K) (GOOGL Q2 2026)
• [ACTUAL] Tesla Q2 2026 capex $5.79B (+142% YoY, calendar-Q2 actual) exceeded OCF $4.7B → FCF NEGATIVE $(1.1)B, self-funded by a $1.2B cash drawdown, no capital raise. Evidence: strong (first-party 8-K) (TSLA Q2 2026)
• The demand side has stopped self-funding — for both buyers that have printed, Q2 capex exceeded operating cash flow; the buildout is now carried by the balance sheet rather than operations. Evidence: strong (two first-party 8-Ks) (GOOGL); (TSLA)
• Supply-side mirror: TXN Q2 2026 revenue $5.46B (beat the $5.00–5.40B guide top), capex falling $4.94B→$3.31B TTM, FCF +271% YoY to 33.6% of revenue — the equipped seller harvests as the buyers' funding gaps widen; "data center" named a growth leader. Evidence: strong (first-party 8-K) (TXN Q2 2026)
• Google Cloud accelerated to $24.8B (+82% YoY) at a 35.6% operating margin — the segment funding the capex is itself inflecting up; Cloud now books TPU-system sales as product revenue (own-silicon value capture). Evidence: strong (first-party 8-K) (GOOGL Q2 2026)
• [UNRESOLVED BY THE PRINT] Alphabet's FY2026 capex dollar guide and RPO/cloud backlog are absent from the Ex-99.1 (call/10-Q items) — the watched $180–190B guide and ~$462B backlog can be neither confirmed nor broken from the release; not invented here. Evidence: n/a (absent from primary source) (GOOGL Q2 2026)

Benchmarks & Data

• 2025 combined capex: ~$410B
• 2026 combined capex (big four): ~$725B (+77% YoY)
• 2026 alternative estimate (broader hyperscale): >$600B
• 2027 projection: $1 trillion+
• Per-company 2026: Amazon ~$200B · Microsoft ~$190B · Google $175–185B · Meta $115–135B
• Allocation: ~50% compute & silicon / ~35% plant / ~15% power & grid
• GPU spend: ~$260B+ (Nvidia) + ~$80B+ (custom silicon)
• Capacity doubling: ~12 quarters

Added 2026-07-22 — all figures below are forecasts/guidance, not measured spend:

• 2026 aggregate, third estimate: $660–690B (five companies incl. Oracle) · 2025 base ~$380B
• 2026 big-four implied from that estimate: $610–640B (derived)
• Honest working band for 2026 big-four capex: $610–725B
• Microsoft 2026: ~$190B or $120B+ — conflicting
• Oracle 2026: $50B (+136% YoY); RPO $523B
• Alphabet 2026 guidance trajectory: $71–73B (initial) → $175–185B → $180–190B
• Alphabet Q1 2026 capex: $35.7B (+107% YoY) [summary-derived, unverified]
• Alphabet June 2026 equity raise earmarked for AI infrastructure: $84.75B [summary-derived, unverified]
• Microsoft most-recent-quarter capex: $37.5B
• [FORECAST] Stargate target: $500B by 2029 ($100B initial; >$400B committed in first three years)

Added 2026-07-23 — first ACTUAL quarterly capex prints (calendar-Q2 2026, cash PP&E basis unless noted):

• Alphabet Q2 capex $44.9B (H1 $80.6B; TTM $132.4B) · OCF $39.1B · FCF $(5.9)B · funded by $49.6B equity + $20.3B notes; LT debt → $98.2B. Q2 run-rate annualizes ~$180B; FY dollar guide NOT in the exhibit
• Tesla Q2 capex $5.79B (+142% YoY; H1 $8.28B; TTM $12.9B) · OCF $4.7B · FCF $(1.1)B · self-funded by $1.2B cash drawdown
• TXN (supply-side mirror) Q2 capex $514M (TTM $3.31B, down from $4.94B) · Q2 FCF $2.74B (TTM $6.53B, +271% YoY, 33.6% of revenue) · Q2 revenue $5.46B (beat $5.00–5.40B guide)
• Google Cloud Q2 rev $24.8B (+82% YoY), operating margin 35.6%

Open Questions

• Will hyperscaler capex inflect down at any point, and what is the leading indicator? (Frontier debate #4.)
• Which Microsoft number is right — ~$190B or $120B+? A ~$70B single-name gap is the largest unresolved figure in this topic and drives the entire aggregate dispersion. Resolving it requires Microsoft's own disclosure, which this KB does not have.
• Why does the ~$725B headline exceed the sum of its own per-company parts? Either the headline includes something the four figures do not, or one of them is wrong.
• Reconciling the ~$725B (big-four) vs >$600B (broader hyperscale) vs $660–690B (five-company) figures — what exactly does each scope include?
• Does the growing weight of first-party builders (see Vertically Integrated Compute) mean the big-four aggregate systematically understates total AI-datacenter capex?
• How much of the 2026 spend is debt-financed vs funded from operating cash flow, and does that change the cycle's fragility? (Partially answered 2026-07-23: for the two names that printed, Q2 capex now exceeds OCF — Alphabet funded the gap from capital markets, Tesla from a cash drawdown. The remaining question is how long negative FCF can run before it constrains the buildout, and whether Amazon/Microsoft/Meta show the same when they print.)
• Does the custom-silicon line (~$80B+) erode Nvidia's ~$260B+ share over the cycle?
• What return-on-invested-capital does $725B/year actually earn — is the spend justified by realized AI revenue, or front-running it?

Related Concepts

• Power: The Binding Constraint — the ~15% power-and-grid slice of capex is where the buildout actually gates
• Rack Power Density — the NVL72 density step-change is what makes each capex dollar buy more compute (and more cooling load)
• Contracted Backlog & RPO — the demand-side test of whether this capex is justified or front-running revenue
• Data-Center Construction as a Macro Variable — the same cycle measured in national accounts
• Advanced Packaging Bottleneck — why approved capex cannot convert to installed capacity on demand

Changelog

• 2026-06-24 — Created from 2 sources (ai-hyperscaler-capex-2026, hyperscaler-power-bottleneck-2026)
• 2026-07-22 — Added a third 2026 estimate ($660–690B, Futurum) and recorded three unresolved conflicts (big-four aggregate $725B vs $610–640B derived; Microsoft ~$190B vs $120B+; 2025 base $410B vs $380B), plus an internal inconsistency in the $725B headline vs its own per-company sum. Widened the working band to $610–725B, labelled all aggregates as forecasts, and added Oracle's $50B figure (closing a previously-flagged gap).
• 2026-07-23 — Folded in the first primary Q2 2026 prints (GOOGL, TSLA, TXN 8-Ks). New spine finding: the demand side has stopped self-funding — Alphabet ($44.9B capex > $39.1B OCF → FCF $(5.9)B, funded by $49.6B equity + $20.3B notes) and Tesla ($5.79B capex > $4.7B OCF → FCF $(1.1)B, self-funded by cash drawdown). Added TXN as the supply-side mirror (capex rolling off, FCF +271%). Every capex figure now carries its (period basis, lease treatment). Alphabet's FY capex guide and RPO marked unresolved-by-the-print (absent from the Ex-99.1). Partially answered the debt-vs-cash-flow financing question.