04Buildout· as of 2026-10-15
AI datacenter buildout
The gigawatt land-grab. Financial disclosures reviewed 2026-10-15; cluster ledger reviewed 2026-10-14. Companywide capex includes spending beyond AI.
Big-4 capex guidance · 2026
$732.5B
≈1.8× 2025 actuals · MSFT + GOOG + META + AMZN midpoints
NVIDIA DC revenue · FY27 Q2
$89.0B
~$360B/yr run-rate
Tracked frontier pipeline
~37 GW
Sum of planned capacity, clusters below
Largest single campus
Hyperion
Meta · up to 5 GW
Big-tech capital expenditure
US$B per calendar year (2026 = guidance midpoints)
Guidance reviewed Oct 2026; older actuals retained. Company definitions differ: Microsoft includes finance leases; Oracle’s FY27 estimate is NET cash capex and is not comparable with prior-year gross capex. See source notes below.
Exact values for Big-tech capital expenditure
| Company | 2023 | 2024 | 2025 | 2026E |
|---|---|---|---|---|
| Amazon | $52.7B | $83.0B | $131.8B | $220.0B |
| Microsoft | $41.2B | $75.6B | $118.0B | $175.0B |
| Alphabet | $32.3B | $52.5B | $91.4B | $200.0B |
| Meta | $28.1B | $39.2B | $72.2B | $137.5B |
| Oracle | $6.9B | $21.2B | $55.7B | $70.0B |
NVIDIA data-center revenue
US$B by fiscal quarter (FY ends late January)
NVIDIA quarterly releases. Next-quarter guidance ~$108B total revenue, assuming zero China DC sales.
Exact values for NVIDIA data-center revenue
| Fiscal quarter | Data-center revenue |
|---|---|
| FY24 Q1 | $4.28B |
| FY24 Q2 | $10.32B |
| FY24 Q3 | $14.51B |
| FY24 Q4 | $18.40B |
| FY25 Q1 | $22.56B |
| FY25 Q2 | $26.27B |
| FY25 Q3 | $30.77B |
| FY25 Q4 | $35.58B |
| FY26 Q1 | $39.11B |
| FY26 Q2 | $41.10B |
| FY26 Q3 | $51.20B |
| FY26 Q4 | $62.30B |
| FY27 Q1 | $75.20B |
| FY27 Q2 | $89.00B |
Guidance sources and accounting basis
- Amazon · CY2026 cash capex ~$220B, raised from ~$200B as memory costs increased (Amazon’s Q2 call recording; Amazon publishes no transcript). Cash capex is net of proceeds and incentives, while 2023–2025 are gross purchases of property and equipment (net: $48.1B / $77.7B / $128.3B). Companywide spending, not AI-only. Company source (2026-07-30) Review: 2026-10-14.
- Microsoft · CY2026 ~$175B including finance leases; reduced from $190B due to operating-lease reclassification, with underlying investment expectations unchanged. Company source (2026-07-29) Review: 2026-10-14.
- Alphabet · 2026 guidance $195–205B midpoint, raised from $180–190B as capacity delivery accelerated. Company source (2026-07-22) Review: 2026-10-14.
- Meta · 2026 guidance $130–145B midpoint (narrowed from $125–145B), including principal payments on finance leases. Company source (2026-07-29) Review: 2026-10-14.
- Oracle · Oracle’s Q4 FY26 earnings slides (June 10, slide 15): ~$70B FY27 net cash outlay for capex, Oracle’s non-GAAP measure net of capex-related financing and customer prepayments. The ~$90–95B gross figure rests on secondary reports of the June 10 call. Prior years are gross capex (Oracle’s own net cash outlay: FY25 $19.8B, FY26 $47.7B). Fiscal years June–May mapped to prior calendar year. Company source (2026-06-10) Review: 2026-10-15.
- NVIDIA latest quarter and guidance (2026-08-26)
Frontier-scale clusters & campuses
26 sites · 9 operating · ~37 GW planned
Provenanced site ledger, surveyed Oct 2026. Power figures are IT load where trackers publish it and campus power where only the operator’s headline exists — open Provenance for each figure’s dated source.
Anthropic–Google TPU fleetest — estimated / secondary-source figure Provenance for Anthropic–Google TPU fleetCapacity program, not a single campus: >1 GW in 2026 plus 5 GW of next-gen TPU capacity (Google/Broadcom) from 2027. Anthropic calls that deal an expansion of its $50B US commitment, so it may overlap the Fluidstack rows, which carry no MW.
| Anthropic · Google Cloud | Multiple US sites | — → 6 GW | Up to 1M TPUs (v7 Ironwood era); next-gen TPUs via Broadcom from 2027 | Building |
Hyperionest — estimated / secondary-source figure Provenance for HyperionPlanned figure is Meta’s 5 GW target, which the tracker says lacks concrete evidence at this site alone; it models ≈1.6 GW IT and first power in 2028.
| Meta | Richland Parish, Louisiana | — → 5 GW | NVIDIA Rubin-era planned (~469K chips modeled) | Building |
Jamnagar AI campusest — estimated / secondary-source figure Provenance for Jamnagar AI campusReported 3 GW target — would be the world’s largest by capacity; $20–30B reported cost. Reliance itself says “multi-gigawatt”; construction had begun by Feb 2026 and the first 120 MW is due to be commissioned by end-2026.
| Reliance Industries | Jamnagar, Gujarat, India | — → 3 GW | NVIDIA GB300 (initial fleet; >75K H100-eq for inference) | Building |
Stargate Doña Ana County (Project Jupiter)est — estimated / secondary-source figure Provenance for Stargate Doña Ana County (Project Jupiter)Power plan switched to an on-site Bloom fuel-cell microgrid (Apr 2026); planned figure is facility power (tracker-modeled IT ≈ 1.75 GW).
| OpenAI · Oracle · STACK | New Mexico | — → 2.5 GW | NVIDIA (≈5.1M H100-equivalents projected) | Building |
Fairwater Wisconsinest — estimated / secondary-source figure Provenance for Fairwater Wisconsin“World’s most powerful AI datacenter” claim; first facility complete Jun 2026, campus path to ~2 GW by 2028.
| Microsoft | Mount Pleasant, Wisconsin | 369 MW → 2.3 GW | Designed to interconnect hundreds of thousands of NVIDIA GPUs (GB200 servers) in one cluster (~176K B200 live, tracker estimate) | Operating |
Project Rainierest — estimated / secondary-source figure Provenance for Project RainierFlagship of the Amazon–Anthropic expansion (up to 5 GW, >$100B AWS commitment over 10 years). Live figure is tracker IT load; planned figure is Amazon’s 2.2 GW campus, which the tracker models as ≈1.9 GW IT by Q1 2028.
| Anthropic · AWS | New Carlisle, Indiana | 910 MW → 2.2 GW | ~1.05M Trainium2 live (tracker estimate); ~1 GW Trainium2+3 across AWS by end-2026 | Operating |
Colossus 2est — estimated / secondary-source figure Provenance for Colossus 2Spans Memphis, TN and Southaven, MS; a third building (projected ~Feb 2027) takes the site toward 2 GW of facility power. Anthropic also contracted GPU capacity here in 2026.
| SpaceXAI (SpaceX, formerly xAI) | Memphis, Tennessee | 946 MW → 2 GW | ~550K GB200/GB300 operating (Sep 2026); +660K GB300 targeted by late Dec 2026 | Operating |
Stargate Shackelford Countyest — estimated / secondary-source figure Provenance for Stargate Shackelford CountyVantage “Frontier” campus; Vantage targets first delivery in H2 2026, while the tracker projects first IT load in 2027.
| OpenAI · Oracle · Vantage | Texas | — → 2 GW | NVIDIA Rubin-era (≈4.6M H100-equivalents projected) | Building |
Stargate Saline Townshipest — estimated / secondary-source figure Provenance for Stargate Saline TownshipRelated Digital build southwest of Ann Arbor; announced Oct 2025, construction celebrated Jun 1, 2026 (first of three buildings then nearing completion).
| OpenAI · Oracle · Related Digital | Michigan | — → 1.4 GW | GB300-era NVIDIA (≈2.9M H100-equivalents projected) | Building |
Stargate Port Washington (Lighthouse)est — estimated / secondary-source figure Provenance for Stargate Port Washington (Lighthouse)
| OpenAI · Oracle · Vantage | Port Washington, Wisconsin | — → 1.3 GW | GB300-era NVIDIA (≈2.6M H100-equivalents projected) | Building |
Stargate Abileneest — estimated / secondary-source figure Provenance for Stargate AbileneFlagship of the Stargate program — 6 of 8 buildings delivered by Sep 2026 (Oracle); planned figure is campus facility power (tracker-modeled IT ≈ 0.84 GW).
| OpenAI · Oracle · Crusoe | Abilene, Texas | 618 MW → 1.2 GW | ~403K B200/B300 projected at full buildout (tracker estimate) | Operating |
Stargate Milam Countyest — estimated / secondary-source figure Provenance for Stargate Milam CountySB Energy (SoftBank) builds and operates the site; its first building’s construction completion is filed for Oct 15, 2026 (Texas TDLR), with no energization announced yet.
| OpenAI · SB Energy | Texas | — → 1.2 GW | GB300-era NVIDIA (≈2.5M H100-equivalents projected) | Building |
Prometheusest — estimated / secondary-source figure Provenance for PrometheusMeta’s ~1 GW cluster target; nuclear supply deals (TerraPower, Oklo, Vistra) signed Jan 2026. The Oct 2026 tracker re-model (B300 chips) lowered the live estimate; the site did not shrink.
| Meta | New Albany, Ohio | 471 MW → 1 GW | ~166K B300 (→ ~363K by Q3 2028, tracker estimate) | Operating |
El Paso campusest — estimated / secondary-source figure Provenance for El Paso campus | Meta · BlackRock | El Paso, Texas | — → 1 GW | AI-optimized campus (chips TBD) | Building |
Stargate UAEest — estimated / secondary-source figure Provenance for Stargate UAE1 GW cluster anchoring the 5 GW UAE–US AI campus. The first 200 MW was due in Q3 2026 (Mubadala, Dec 2025); no go-live had been announced by mid-Oct 2026, and Sep 2026 press reports describe a hardened redesign.
| G42 · OpenAI · Oracle | Abu Dhabi, UAE | — → 1 GW | NVIDIA GB300 systems | Building |
Visakhapatnam AI hubest — estimated / secondary-source figure Provenance for Visakhapatnam AI hub$15B for 2026–2030, Google’s largest investment in India to date; ground broken Apr 28, 2026 and operations expected in 2028. Petitions against its environmental clearances are pending (Oct 2026 press reports).
| Google · AdaniConneX · Airtel (Nxtra) | Visakhapatnam, Andhra Pradesh, India | — → 1 GW | Google AI infrastructure (accelerators undisclosed) | Building |
Fairwater Atlantaest — estimated / secondary-source figure Provenance for Fairwater AtlantaQTS-built, grid-only design (no UPS/gen-sets); linked with Wisconsin into one “AI superfactory” WAN.
| Microsoft | Fayetteville, Georgia | 636 MW | ~304K B200 (tracker estimate) | Operating |
HUMAIN AI factoriesest — estimated / secondary-source figure Provenance for HUMAIN AI factoriesPIF-backed sovereign AI buildout; capacity target “up to 500 MW” across factories in the Kingdom. A Sep 2026 secondary report of a live NVIDIA B300 cloud awaits a HUMAIN or NVIDIA primary source.
| HUMAIN · NVIDIA | Riyadh, Riyadh Province, Saudi Arabia | — → 500 MW | Up to 600K NVIDIA GPUs over 3 years (KSA + US, incl. GB300); first cluster ~18K GB300 | Building |
Stargate Argentinaest — estimated / secondary-source figure Provenance for Stargate ArgentinaLetter of intent only (up to $25B), with OpenAI as a prospective offtaker of a Sur Energy joint venture; the project had not been presented to the government or its RIGI investment regime as of late May 2026.
| Sur Energy · OpenAI | Patagonia, Argentina | — → 500 MW | Next-generation AI computing (chips TBD) | Announced |
Polaris Forge 1 Provenance for Polaris Forge 1Applied Digital campus fully leased to CoreWeave across three buildings (~$11B contracted revenue); Buildings 1–2 (250 MW) live by Oct 2026, third 150 MW building under construction.
| CoreWeave · Applied Digital | Ellendale, North Dakota | 250 MW → 400 MW | CoreWeave GPU cloud capacity (NVIDIA, models undisclosed) | Operating |
Omaha TPU campusest — estimated / secondary-source figure Provenance for Omaha TPU campusGoogle’s flagship disclosed-scale TPU training campus; expansion projected through Q3 2028.
| Omaha, Nebraska | 237 MW → 395 MW | ~382K TPUs (v5e/v5p/v6e/v7, tracker estimate) | Operating | |
Colossus 1est — estimated / secondary-source figure Provenance for Colossus 1First ~100K H100 built in 122 days in 2024; a grid link upgraded toward 150 MW plus on-site gas turbines and Megapacks. Since May 2026 all of its compute capacity is contracted to Anthropic.
| SpaceXAI (SpaceX, formerly xAI) | Memphis, Tennessee | 300 MW | ~230K GPUs (150K H100, 50K H200, 30K GB200) | Operating |
Stargate Lordstownest — estimated / secondary-source figure Provenance for Stargate LordstownSoftBank/Foxconn advanced data-center design; ground broken 2025 and targeted operational 2026 (Sep 2025), while the tracker now projects Q1 2027.
| OpenAI · SoftBank · Foxconn | Lordstown, Ohio | — → 300 MW | NVIDIA Rubin (≈0.55M H100-equivalents projected) | Building |
Nscale Narvik (announced as Stargate Norway)est — estimated / secondary-source figure Provenance for Nscale Narvik (announced as Stargate Norway)Hydropower-fed campus, now run solely by Nscale (the Aker joint venture was folded in, Mar 2026) with Microsoft as offtaker; OpenAI reportedly left lease talks (Apr 2026). Ambition to add a further 290 MW beyond the initial 230 MW.
| Nscale (Microsoft offtake) | Narvik, Nordland, Norway | — → 230 MW | ≈52K NVIDIA GB300 (Microsoft, multi-year) + 30K+ Rubin from 2027 | Building |
Fluidstack build (Texas)est — estimated / secondary-source figure Provenance for Fluidstack build (Texas)Part of Anthropic’s $50B US datacenter program with Fluidstack (Texas, New York and more sites to come); no per-site MW published. | Anthropic · Fluidstack | Texas | — | Custom Anthropic build (chips undisclosed) | Building |
Fluidstack build (New York)est — estimated / secondary-source figure Provenance for Fluidstack build (New York)Part of Anthropic’s $50B US datacenter program with Fluidstack (Texas, New York and more sites to come); no per-site MW published. | Anthropic · Fluidstack | New York | — | Custom Anthropic build (chips undisclosed) | Building |
The contracts that pay for these campuses, from cloud capacity to chips, leases and power, are in the compute deals ledger.