Hyperscaler Buyers

Type: Group — the demand side (Microsoft, Google/Alphabet, Meta, Amazon)

The hyperscaler buyers are the four largest US cloud operators whose capital spending is the demand engine for the entire datacenters stack — "capex prints set the tape for everyone above" (the GPU, plant, cooling, and power vendors). In 2026 their combined capex is guided to ~$725B, up from ~$410B in 2025 (+77% YoY), with a 2027 projection of $1 trillion+. They are spending in synchrony because, per Value Add VC, "each company fears being structurally short on compute more than it fears overspending" — see Hyperscaler Capex Cycle for the full cycle treatment.

The per-company guidance for 2026 is Amazon $200B (+100% YoY), Microsoft $190B (+100%), Google/Alphabet $175–185B (+110%), and Meta $115–135B (+80%). The Data Center Knowledge source surfaces the physical buildout behind those numbers — Meta targeting >10 GW total by end-2026 (Louisiana "Hyperion" $27B / 2.5M sq ft / 2→5 GW; Ohio "Prometheus" 1 GW online 2026), Microsoft's Atlanta Fairwater operational Oct 2025 with Wisconsin in early 2026, and AWS international builds in Saudi ($5.3B), Chile (>$4B), and Germany (€7.8B through 2040). On the NVIDIA supply-chain map, Microsoft and Google Cloud also appear as named cloud operators deploying GB300 NVL72-class systems.

Oracle is part of the broader hyperscale demand picture (Oracle/Stargate I in Abilene at 1.2 GW, and named on the NVIDIA map), but the per-company capex figures in the Value Add source cover only the big four — Oracle is not given a 2026 capex number in any ingested source. [Superseded 2026-07-22 — Oracle now has a figure: $50B, +136% YoY. See Oracle.]

Update — 2026-07-22: the aggregate is contested, and Microsoft is why

Confidence on this page is downgraded from medium to low. A third source (Futurum, Feb 2026) puts 2026 capex at $660–690B across five companies including Oracle, implying a big-four figure of $610–640B — $85–115B below the ~$725B this page has been asserting. Reconciling the two sources shows the disagreement is almost entirely one name:

Two further problems. The ~$725B headline exceeds the sum of its own per-company parts ($680–710B) by $15–45B. And the 2025 base year is disputed — ~$410B versus ~$380B — which changes the headline YoY growth rate. The KB's working position is now a band of $610–725B for 2026 big-four capex, not a point estimate, and every figure on this page is guidance or forecast, not measured spending.

Three of the four also report inside a two-week window around this compile, which will move these numbers: Alphabet 2026-07-22 (source-confirmed). Tesla, a non-hyperscaler first-party builder, reports the same day (desk calendar context, not source-confirmed).

Key Facts

• Combined 2026 capex ~$725B (up from ~$410B in 2025, +77% YoY); 2027 projection $1 trillion+. Evidence: moderate (analysis — Value Add VC) (AI Hyperscaler Capex 2026)
• Amazon $200B (+100% YoY) 2026 capex guidance. Evidence: moderate (analysis — Value Add VC) (AI Hyperscaler Capex 2026)
• Microsoft $190B (+100% YoY); Atlanta Fairwater operational Oct 2025, Wisconsin early 2026; named on NVIDIA's cloud map. Evidence: moderate (analysis + technical-report) (AI Hyperscaler Capex 2026); (Hyperscalers in 2026)
• Google/Alphabet $175–185B (~+110% YoY); Google Cloud named on NVIDIA's cloud map. Evidence: moderate (analysis + technical-report) (AI Hyperscaler Capex 2026)
• Meta $115–135B (~+80% YoY) — SUPERSEDED, primary guidance is $125–145B [primary: META 10-Q 2026-04-30]; targeting >10 GW total by end-2026 (Hyperion $27B/2.5M sq ft/2→5 GW; Prometheus 1 GW online 2026). Evidence: moderate (analysis + analysis) (AI Hyperscaler Capex 2026); (Hyperscalers in 2026)
• AWS international builds: Saudi ($5.3B), Chile (>$4B), Germany (€7.8B through 2040). Evidence: moderate (analysis — Data Center Knowledge) (Hyperscalers in 2026)
• Driver: "each company fears being structurally short on compute more than it fears overspending." Evidence: moderate (analysis framing) (AI Hyperscaler Capex 2026)
• Oracle is part of the broader hyperscale picture (Stargate I, Abilene 1.2 GW; named on NVIDIA map) but has no per-company 2026 capex figure in any ingested source. Evidence: moderate (analysis + technical-report); capex figure is a gap (Hyperscalers in 2026) — SUPERSEDED 2026-07-22: $50B, +136% YoY, $523B RPO

Added 2026-07-22

• CONFLICT — Microsoft 2026 capex: ~$190B vs $120B+. A ~$70B single-name gap that accounts for essentially the entire aggregate discrepancy. Unresolved; no Microsoft primary source in this KB. Evidence: n/a (unresolved conflict) (Futurum); (Value Add VC)
• CONFLICT — 2026 big-four aggregate: ~$725B vs $610–640B (derived). Working band is now $610–725B. Evidence: n/a (unresolved conflict) (Futurum); (Value Add VC)
• CONFLICT — 2025 base year: ~$410B vs ~$380B. Evidence: n/a (unresolved conflict) (both sources)
• Microsoft carries ~$80B of unfulfilled Azure backlog attributed to power constraints, and spent $37.5B in its most recent reported quarter. Evidence: weak (single analysis source) (Futurum AI Capex 2026)
• Alphabet's 2026 guidance was re-based ~2.5x — from an initial $71–73B projection to $175–185B across three revisions, then to $180–190B per the July preview. Evidence: weak (analysis + summary-derived figures) (Futurum); (Alphabet preview)
• The cohort now describes itself as supply-constrained, not demand-constrained. Evidence: weak (single analysis source characterizing management commentary) (Futurum AI Capex 2026)
• Their spending 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)
• Their capex no longer converts cleanly into capacity — CoWoS packaging, HBM memory and power are all constrained simultaneously. Evidence: moderate (converging across three sources) (TSMC Q2 2026); (Samsung preliminary); (Hyperscalers in 2026)

Mentioned In

• Hyperscaler Capex Cycle — these four buyers ARE the cycle; per-company splits live here
• Power: The Binding Constraint — their gigawatt-scale sites are what the power bottleneck gates
• CoreWeave — neoclouds are the independent demand layer alongside (and sometimes inside) these buyers
• Contracted Backlog & RPO — the demand-side test of whether this spend is justified
• Data-Center Construction as a Macro Variable — their capex measured in national accounts
• Alphabet · Oracle — split-out per-company dossiers

Outstanding Questions

• Will hyperscaler capex inflect down at any point, and what is the leading indicator? (Frontier debate #4. Nearest test: Alphabet's 2026-07-22 capex guidance revision.)
• Which Microsoft figure is right — ~$190B or $120B+? The largest unresolved number in the topic. Requires a Microsoft primary source.
• What is the realized return on $610–725B/year — is the spend justified by AI revenue or front-running it?
• How much of each buyer's 2026 capex is power/generation vs compute, and does the mix differ by company?
• Should Oracle be split out as its own entity once a capex figure is available? Done 2026-07-22 → Oracle. The same question now applies to Microsoft, which should get its own page once a primary source resolves the capex conflict.
• Does the big-four framing systematically understate total AI-datacenter capex by excluding first-party builders like Tesla and xAI?