Global cloud infrastructure survey: 2025 results and capacity estimates

Summary

AWS remains the scale leader, with $128.7B of FY2025 revenue, $45.6B of operating income, 39 regions and 123 Availability Zones, and about 29% of the market in Q3 2025. Microsoft Azure ($75B in FY2025, growing 34–40%) and Google Cloud ($58.7B, up 36%, with $13.9B of operating income) are growing much faster. According to Synergy Research Group, the Big Three held 63% of a market that reached $106.9B of quarterly revenue in Q3 2025 and $390B on a trailing-twelve-month basis.

By our estimates, AWS ended 2025 with roughly 11 GW of data-center power capacity (range 9–15 GW), Microsoft about 8 GW (7–9.5 GW) and Alphabet's whole fleet about 7.5 GW (6.6–8.7 GW). Oracle, Alibaba and Tencent each sit in the 1–2 GW band, Huawei Cloud below 1.3 GW, and CoreWeave disclosed 850 MW of active power. The estimates rest on company power statements, electricity data from sustainability reports and footprint models, and they reconcile with a top-down check against IEA and Synergy totals.

Infrastructure disclosure is highly uneven. Providers publish region and zone counts and cloud revenue, but almost never total server counts. The comparison that matters has shifted to gigawatts and custom-silicon volumes: AWS says Project Rainier runs nearly 500,000 Trainium2 chips and 1.4 million Trainium2 chips have landed overall, while Alibaba says it had shipped 470,000 AI chips cumulatively by February 2026.

Capex keeps climbing. The four US hyperscalers guided to roughly $725B combined for 2026, up 77% from 2025's record $410B (Amazon ~$200B, Google $175–185B, Microsoft ~$190B for calendar 2026, Meta $115–135B), and Amazon later lifted its plan to about $220B. Oracle (about $50B for FY2026) and the Chinese providers, led by Alibaba's RMB 380B (~$53B) three-year plan, are also ramping hard.

What a gigawatt is New

Every capacity figure in this report is quoted in gigawatts, a unit almost nobody has an intuition for. A gigawatt is a billion watts, and a watt is a rate — the size of the tap, not the amount of water in the bucket. The ladder below climbs from a device most readers own to the unit the rest of this report is written in. Each step is roughly ten times the one before it, which is the only honest way to draw a range this wide.

Read as multiples of a single gigawatt, the same ladder looks like this.

The two comparisons at the bottom are the ones worth carrying. A gigawatt is about one average American nuclear reactor — the US operates 94 of them with a combined 96.9 GW, or roughly 1,030 MW each. It is also the average draw of about 830,000 American homes, since the average home buys about 10,300 kWh a year, an average of roughly 1.2 kW. That second figure compares a data center's installed capacity against households' average consumption, so treat it as a sense of scale rather than an equivalence.

On those terms, AWS's estimated 11 GW is on the order of eleven nuclear reactors, or the average household draw of a city of nine million homes. Microsoft's 8 GW and Alphabet's 7.5 GW are not far behind, and the four US hyperscalers are together planning to spend $725B in 2026 largely on adding more of it.

What a gigawatt buys in servers

Power converts to hardware only through a density assumption, and the report uses the one anchor that is public: Hamilton's AWS design figures of 25–32 MW buildings holding 50,000–80,000 servers.

Treat that number with more suspicion than the rest of this report. Hamilton's figures describe general-purpose fleets of the mid-2010s. An AI hall spends the same gigawatt on far fewer, far hungrier boxes: at 140 kW per rack and roughly 10 kW per AI server, a gigawatt of Fairwater-style capacity is closer to 100,000 accelerated servers than to millions of general-purpose ones. The conversion factor is not a constant, and it has been moving fast in one direction.

Power is not energy

Gigawatts describe capacity at an instant; the sustainability reports this analysis leans on publish terawatt-hours consumed across a year. One gigawatt held flat for a full year would be 8.76 TWh, but no fleet runs flat out, and at the 0.65–0.85 load factor used in the estimation methods a gigawatt of installed capacity produces roughly 5.7–7.4 TWh of annual consumption. That ratio is why Microsoft's disclosed 37 TWh for FY2025 implies something near 8 GW rather than 37, and it is the single step where a wrong assumption moves every estimate in this report.

Market context

The global cloud infrastructure market reached $106.9B of revenue in Q3 2025, up 28% year on year in constant currency. Synergy's John Dinsdale described it as the eighth successive quarter of accelerating growth and the fastest rate in three years. Trailing-twelve-month provider revenue passed $390B and was on course to exceed $400B for the first time.

AWS, Azure and Google Cloud together took 63% of enterprise cloud infrastructure spend in Q3 2025 (AWS 29%, Azure 20%, Google 13%), up from 62% and 61% in the two prior years, so the Big Three are still gaining share while the market grows. Alibaba, Huawei and Tencent dominate China but hold low single-digit shares globally.

Every major provider says it is short of AI compute rather than short of demand. Region counts have become a weaker differentiator than gigawatts of capacity and accelerator counts, which combine NVIDIA GPUs with custom silicon: Graviton and Trainium at AWS, Cobalt and Maia at Microsoft, Axion and TPU at Google, Yitian and PPU at Alibaba, and Kunpeng and Ascend at Huawei.

Data centers: regions, zones and capacity

AWS. The AWS Global Infrastructure page lists 123 Availability Zones in 39 regions, with seven more zones and two more regions (Saudi Arabia and Chile) announced. Each region has at least three isolated zones. Planned 2026 additions include Chile (end-2026, three zones, more than $4B committed), a new Maryland zone for US East (N. Virginia) and the AWS European Sovereign Cloud, whose first region opened on a separate partition in January 2026. AWS does not disclose how many buildings it runs.

Microsoft Azure. Microsoft cites 70+ regions in 33 countries, more than any other provider, and over 400 data centers. It stood up more than 2 GW of new capacity in FY2025 (ended 30 June 2025). New regions opened in Malaysia and Indonesia in May 2025, with India and Taiwan planned for 2026 and a second Malaysian region announced. The Fairwater AI sites in Wisconsin and Atlanta each hold 150,000+ NVIDIA GB200 GPUs alongside Maia 100 accelerators and Cobalt 100 CPUs, run racks at about 140 kW and rows at about 1,360 kW, and are linked by 120,000 miles of dedicated AI-WAN fiber. Fairwater 4 is under construction.

Google Cloud. The locations page (updated 31 August 2026) lists 43 regions and 130 zones, plus 200+ network edge locations. Google notes that Stockholm, Mexico, Osaka and Montreal currently run three zones in one or two physical data centers and are expanding. New regions in Sweden, South Africa and Mexico launched in 2025; Kuwait, Malaysia and Thailand are underway, and a $2B, ten-year commitment to Türkiye was announced.

Alibaba Cloud. At Apsara 2025 Alibaba reported 91 zones across 29 regions, 14 of them in mainland China; its live website shows 106 zones in 31 regions, a gap that is not explained. It announced its first data centers in Brazil, France and the Netherlands and expansions in Mexico, Japan, South Korea, Malaysia and Dubai. The RMB 380B three-year capex plan was described as on track, with CEO Eddie Wu signalling further increases.

Oracle. Oracle's region count aggregates OCI public, dedicated, sovereign and multicloud regions. In March 2025 it cited 101; by its FY2026 Q2 call (December 2025) it reported 147 live customer-facing regions and 64 more planned. Oracle leases rather than owns its buildings, which complicates capacity comparisons, and its delivery figures describe capacity handed over to customers.

CoreWeave. Filings show growth from 32 data centers, 250,000+ GPUs and 360+ MW (December 2024) to 43 active data centers and 850+ MW of active power at end-2025, with about 3.1 GW contracted, including the Core Scientific acquisition.

Servers and AI accelerators

Almost no provider discloses total server counts. The table separates disclosed sub-counts from estimates; a capacity-based translation for AWS appears in the capacity estimates section.

AWS custom silicon. On the Q4 2025 call Amazon said 1.4 million Trainium2 chips had landed and now carry most Amazon Bedrock inference. Project Rainier, activated in October 2025, deployed nearly 500,000 Trainium2 chips for Anthropic across the 1,200-acre St. Joseph County, Indiana site (an $11B investment), with plans to exceed one million. Trainium3 (TSMC 3 nm, 144 GB HBM3e) launched at re:Invent 2025, and Jassy said the entire Trainium3 supply was expected to be committed to workloads by mid-2026.

Google TPU. Ironwood delivers 192 GB of HBM3e per chip; a 9,216-chip pod reaches 42.5 exaFLOPs with 1.77 PB of shared memory. Anthropic committed in October 2025 to up to one million TPUs and over 1 GW of capacity in 2026, rising to 3.5 GW in 2027 per a Broadcom filing. Analysts project about 4.3 million TPU shipments in 2026.

Alibaba (T-Head). CEO Wu Yongming disclosed 470,000 cumulative AI chip shipments on the Q3 FY2026 call, with T-Head chip revenue near an RMB 10B annualized run-rate. The PPU is reportedly comparable to NVIDIA's H20, and Wu acknowledged Alibaba's chips still trail foreign counterparts on performance.

Fleet size. No provider discloses a total server count. Hamilton's AWS design figures (25–32 MW buildings holding 50,000–80,000 servers) remain the best public anchor for converting power into servers; see the conversion factors in the estimation methods section.

Revenue, profit and capex (2025)

AWS. Revenue rose 20% to $128.7B and operating income to $45.6B (from $39.8B). Q4 revenue was $35.6B, up 24% and the fastest growth in 13 quarters. AWS was 18% of Amazon's revenue but about 57% of its operating income. Amazon's capex reached $131.8B in 2025; the 2026 plan was about $200B in February and about $220B by July, after AWS grew 37% to $42.2B in Q2 2026. The AWS backlog stood at $244B, and Amazon's 2025 free cash flow fell 71% to $11.2B.

Microsoft. In FY2025 Azure passed $75B, up 34%. Intelligent Cloud revenue was $106.3B (up 21%) with $44.6B of operating income, and Microsoft Cloud gross margin slipped to 69% as AI infrastructure scaled. Calendar-2026 capex of about $190B sits well above the roughly $152B analysts had expected. Contracted backlog was $368B.

Google Cloud. Revenue rose 36% to $58.7B with $13.9B of operating income. Q4 revenue was $17.7B (up 48%) and Q4 operating income $5.3B. Alphabet capex was about $91B in 2025, guided to $175–185B for 2026, and its backlog reportedly roughly doubled to about $460B on the Anthropic TPU deal.

Alibaba. FY2025 Cloud Intelligence revenue was RMB 118.0B (up 11%) with adjusted EBITA of RMB 10.6B (up 72%). Growth accelerated to 36% in the quarter to December 2025, on revenue of RMB 43.3B and adjusted EBITA of RMB 3.9B. Alibaba targets about $100B of combined external cloud and AI revenue within five years.

Huawei Cloud. External cloud revenue fell 3.5% to RMB 32.16B in 2025, per the annual report as cited by CNBC, a rare decline attributed to slower Chinese AI adoption and export constraints. Huawei group revenue was RMB 880.9B and net profit RMB 68B.

Oracle. FY2025 revenue was $57.4B (up 8%), with OCI at about $10.2B (up 50%) and capex of $21.2B, up from $8.7B. RPO rose to $138B at FY2025 year-end and then surged to $553B (Q3 FY2026) and $664B (Q1 FY2027). OCI revenue reached $7.4B in the quarter to August 2026, up 121%. RPO is contracted revenue not yet recognized.

Tencent. FinTech and Business Services revenue was RMB 229.4B (up 8%), with Business Services, which houses Tencent Cloud, growing close to 20%. Tencent Cloud reached profitability at scale for the first time on a full-year basis. Capex hit a record RMB 79.2B.

CoreWeave. Revenue rose 168% to $5.13B, with adjusted EBITDA of $3.09B, a net loss of $1.17B, capex of $14.9B and a $60.7B backlog. Microsoft accounted for about 67% of revenue, and the company carried about $21.6B of debt.

Capacity estimates New in this edition

This section estimates each provider's data-center power capacity at 31 December 2025. The unit is facility power: the energized power behind a provider's data centers, including cooling and electrical overhead. To convert to IT load, divide by power usage effectiveness (PUE), which the providers report as 1.14 for AWS, 1.09 for Google, 1.19 for Alibaba's self-built sites and 1.246 for Tencent's owned sites. Our estimates are marked in amber and the key disclosed inputs in cyan.

AWS

AWS is the only provider that has stated its capacity as a ratio to an earlier year, which lets us bound it without knowing the base. On the Q3 2025 call Andy Jassy said AWS had added more than 3.8 GW in the preceding 12 months, that its power capacity was now double the 2022 level, and that it was on track to double again by 2027. In Q4 2025 alone Amazon added more than 1.2 GW.

The zone-tier footprint model in the methods section gives 5–16 GW for the same fleet, which is consistent but three times wider. One caution: on the Q4 2025 call the doubling was phrased in a way that could also be read as comparing annual additions, so we rely on the clearer Q3 2025 prepared remarks. "Power capacity" may also refer to utility power secured rather than IT load in service.

Microsoft

Microsoft reported 37 TWh of electricity use in FY2025 (July 2024 to June 2025), up 24%. Its data fact sheet puts total energy at 134.9 million GJ, about 37.5 TWh including fuels. For the first time it also published site-level figures: the Boydton, Virginia campus alone used more than 3 TWh.

A facility count gives an independent check: 400+ data centers at an average of 15–25 MW is 6–10 GW. The Boydton figure shows how concentrated capacity is: 3 TWh is an average draw of about 340 MW, or roughly 400–530 MW installed at a single campus. In October 2025 Microsoft said it would raise total AI capacity by more than 80% that year and roughly double its total data-center footprint over two years, which implies something like 14–18 GW by late 2027. Fairwater Wisconsin alone is planned to scale to 2 GW.

Alphabet (Google)

Google's data centers used 30.8 TWh in 2024, 95.8% of the company's electricity. Its 2026 Environmental Report says total electricity demand rose 37% in 2025, and press coverage of the report puts data-center use at about 42 TWh. Fleet PUE was 1.09.

This covers the whole of Alphabet: Search, YouTube, Gemini and Google Cloud share one fleet and one TPU pool, so a Google Cloud-only figure cannot be separated. The same method applied to 2024 gives an average draw of 3.5 GW, consistent with the growth path.

Oracle

Oracle reports capacity handed over to customers rather than total capacity. It delivered close to 400 MW in the quarter to November 2025, more than 400 MW in the quarter to February 2026, more than 1.2 GW across FY2026 and 850 MW in the quarter to August 2026.

Oracle's buildings are leased or partner-built, so its "delivered" figures may be closer to IT capacity than to facility power. It says it has secured more than 10 GW of power and data-center capacity coming online over the next three years.

Alibaba

Alibaba's ESG data, filed with the SEC, shows group purchased electricity of 6.75 TWh in FY2025 (April 2024 to March 2025), up from 4.97 TWh, and a self-built data-center PUE of 1.190.

The footprint method agrees: 91 zones at 10–20 MW each is 0.9–1.8 GW. The figure includes capacity serving Taobao, Tmall and other internal businesses as well as Alibaba Cloud customers.

Tencent

Tencent's 2025 ESG data shows total energy use of 31.37 million GJ, about 8.7 TWh, up 34.6%. Owned data centers ran at a PUE of 1.246.

Tencent has historically counted energy used in leased data centers as Scope 3, so leased capacity may sit outside this total and the true footprint could be higher. Most of the capacity serves WeChat, games and advertising; Tencent Cloud is a subset.

Huawei Cloud

Huawei does not publish usable energy data, so this estimate starts from announced site designs. Its Gui'an campus in Guizhou is designed for more than one million servers, and the Ulanqab plan calls for about 20,000 cabinets at 10 kW supporting about 300,000 servers.

The revenue method gives only 0.3–0.6 GW for Huawei Cloud's roughly $4.6B of external revenue, but the same campuses also carry Huawei's internal IT and consumer cloud services, so the site-based figure is the better guide.

Top-down reconciliation

The bottom-up estimates can be checked against industry totals.

The bottom-up central sits about 10% above the top of the top-down range, while the bottom-up low end falls inside it. The gap is within the combined uncertainty (the Big Three's share of new capacity probably rose during 2025, and the 2025 growth assumption for global use is conservative), but it is the reason the AWS central estimate is set below its arithmetic midpoint. A second check uses Synergy's count of 1,360 hyperscale data centers: at an average of 25–35 MW each, that is 34–48 GW, in line with step 4.

From capacity to servers: AWS as a worked example

Capacity can be translated into rough fleet sizes with the conversion factors in the methods section. The split between AI and general-purpose load is an assumption, so the output is illustrative.

The accelerator range is consistent with AWS's disclosure of 1.4 million Trainium2 chips landed, plus its NVIDIA fleet. The same steps apply to any provider: IT load = facility power ÷ PUE; servers = general-purpose IT load × servers per MW; accelerators = AI IT load × accelerators per MW.

Estimation methods New in this edition

No single method is reliable on its own. Each estimate above uses a primary method chosen for the hardest available input, at least one independent cross-check, and a top-down reconciliation across providers. The table summarizes the five methods; the subsections give the formulas and parameters.

A sixth approach, summing individual buildings from generator permits, utility filings and satellite imagery, is the most accurate but needs paid or site-by-site data. It is described at the end of this section.

Disclosure anchor

Collect every capacity statement a company has made and turn each into an equation. Additions ("we added 3.8 GW in 12 months") are increments. Ratio claims ("double the 2022 level") link today's capacity to an unknown base: if capacity is k times the base, the base equals the additions since then divided by k − 1. Undisclosed periods are bounded with capex-informed priors, and later deliveries are added on top. This is the strongest method when statements exist, but the wording must be read carefully: "power" can mean utility capacity secured rather than IT load in service, and Oracle's "delivered" figures describe capacity handed to customers.

Energy

Sustainability reports give annual electricity use, which fixes the average power a fleet actually drew. Installed capacity is larger than that average, because operators keep redundancy headroom, new halls take months to fill and load varies through the day.

1. Take annual electricity from the sustainability report and multiply by the data-center share.
2. Divide by 8,760 hours to get the average draw in GW.
3. Divide by the load factor to get installed facility capacity at the midpoint of the reporting period.
4. Roll forward to the target date using the reported growth rate.
5. Divide by PUE if IT load is needed.

The load factor is the most influential assumption: moving it from 0.65 to 0.85 changes the result by about 30%.

Footprint (regions, zones and facilities)

Zone counts are a weak proxy on their own, because an Availability Zone is a logical unit rather than a standard building. Hamilton described AWS zones as separate buildings, sometimes several, while alleging that some competitors marketed separate suites in one building as distinct zones. Google says some of its three-zone regions run in one or two physical data centers. The method therefore splits a fleet into tiers with very different sizes.

Adjust by provider. For Microsoft, the facility count (400+) is a better base than zones: at 15–25 MW each it gives 6–10 GW. Chinese providers' zones are often leased halls of roughly 10–20 MW, which gives Alibaba 0.9–1.8 GW across 91 zones. Dedicated AI campuses such as Project Rainier, Fairwater and Stargate sit outside the public zone map and must be added separately.

Revenue

Revenue divided by revenue per MW-year gives the capacity needed to earn it. CoreWeave is the calibration point because it discloses both: $5.13B of 2025 revenue on active power that rose from about 360 MW to about 850 MW, an average of roughly 600 MW, or about $8.5M per MW-year for a GPU cloud still ramping. Mature mixed clouds earn more per MW from depreciated capacity and higher-margin services, so the method uses a band of $8–16M. The method is only a cross-check: it becomes circular if calibrated on the providers being estimated, and it is sensitive to pricing, utilization and the lag between building capacity and billing for it.

Capex

Capex converts into new capacity once the facility share is separated out. Hyperscale facilities cost about $10–12M per MW in 2025, well up from about $6M per MW several years earlier. In the AI era most capex goes to chips, servers and networking, so we assume a facility share of 25–35%. On that basis Amazon's roughly $200B 2026 plan, if mostly for AWS, funds about 4–7 GW of additions a year, consistent with its goal of doubling capacity by end-2027. Dividing total capex by facility cost alone would overstate capacity several-fold. Leased capacity, which Microsoft and Oracle use heavily, does not appear in capex at all, and buildings typically come online a year or more after spending starts.

Site-level counting

The most accurate approach counts individual buildings. Generator permits reveal building size: AWS filings in Ohio showed campuses of up to five buildings, each backed by 18 generators of 2.5 MW, about 32 MW per building. SemiAnalysis sells a building-by-building model built this way from permits, utility filings and satellite imagery. Microsoft's new site-level data fact sheet allows a partial sum for Azure's key locations.

Combining the methods

1. Choose the primary method with the hardest input, in this order: disclosure, energy, site-level, footprint, revenue, capex.
2. Run at least one independent cross-check. If the ranges do not overlap, look for a definitional mismatch (facility vs IT power, owned vs leased, group vs cloud unit, fiscal vs calendar timing) before averaging.
3. Use the overlap as the range and set the central value from the primary method.
4. Reconcile the sum of providers against top-down totals and adjust central values if the sum sits outside the top-down band.
5. Grade confidence: high for disclosed figures, medium for a hard input with an agreeing cross-check, low for design-based or model-only figures.

Conversion factors

Recommendations

For scale comparisons, use the official infrastructure pages (AWS 39 regions and 123 zones, Azure 70+ regions, Google 43 regions and 130 zones) but normalize for definitions. AWS requires three or more separate zones per region, while some rivals have counted single-building regions, so raw region counts are not comparable. Weight capacity and cloud revenue more heavily than region tallies.

For capacity, treat the estimates here as roughly ±30% for the Big Three and ±50% for the other providers. To tighten them, add site-level data for flagship regions (a SemiAnalysis-type model, generator permits, utility filings), sum Microsoft's site-level fact sheet, and rerun the energy method each year when sustainability reports appear (Google, Amazon and Microsoft in June and July; Tencent in April; Alibaba mid-year).

For AI capability, track gigawatts and accelerator counts rather than regions. AWS (Trainium and Rainier), Microsoft (Fairwater and Maia) and Google (Ironwood and the Anthropic anchor) lead in vertically integrated AI silicon, the decisive axis for 2026–2027. Neoclouds such as CoreWeave matter as overflow capacity but carry concentration and debt risk.

For financial exposure, watch the gap between capex and revenue. Free cash flow is falling at the leaders (Amazon's dropped 71% to $11.2B in 2025). The key signal that spending is monetizable is how fast backlog converts into recognized revenue: AWS $244B, Google about $460B, Microsoft $368B and Oracle $664B of RPO.

Thresholds that would change this assessment. If Azure sustains 39%+ growth for four or more quarters, it could close the revenue gap with AWS faster than expected. If Oracle converts its RPO and its 10 GW power pipeline on schedule, OCI becomes a genuine fourth hyperscaler; its Q1 FY2027 delivery of 850 MW in one quarter is an early sign. If US export controls ease, Alibaba's and Huawei's AI-cloud position would change materially; tighter controls would deepen their reliance on Ascend and PPU. If hyperscaler free cash flow turns negative for several quarters, expect investor pressure to slow the capex ramp.

Caveats and data gaps

Capacity estimates are models. The load factor, the data-center share of electricity and the AI/general split are assumptions, not disclosures, and plausible changes move results by 20–30%. Definitions also differ: AWS's "power capacity", Oracle's "delivered" capacity, Synergy's critical IT load and the facility power used here are not the same quantity.

Scope differs by provider. Alphabet, Alibaba and Tencent figures include internal workloads such as Search, YouTube, Taobao and WeChat. Capacity Microsoft leases from neoclouds is counted at the neocloud, not at Microsoft. Tencent's energy total may exclude leased data centers.

Some energy inputs are secondary. Google's 2025 data-center figure of about 42 TWh and Tencent's 31.37 million GJ come from press and data-aggregator summaries of the companies' reports and should be checked against the original PDFs.

Server counts. No major provider discloses total server or GPU fleet counts. Every fleet figure here is either a disclosed sub-count (1.4 million Trainium2 landed, nearly 500,000 in Project Rainier) or an estimate with wide error bars, such as Next Platform's roughly 600,000 GPUs for CoreWeave against about 250,000 in its December 2024 filing.

Fiscal years. Microsoft's year ends in June, Alibaba's in March and Oracle's in May; AWS, Google, Tencent, Huawei and CoreWeave report calendar years. Cross-provider comparisons are approximate and period-shifted.

Region counts. Alibaba's press release (29 regions, 91 zones) and website (31 regions, 106 zones) disagree. Oracle's region figures aggregate public, dedicated, sovereign and multicloud regions. Huawei does not publish consistent region or data-center counts, and its RMB 32.2B cloud figure covers external customers only.

Market share varies by research firm and scope (IaaS alone or with PaaS and hosted private cloud). Synergy's Q3 2025 split (AWS 29%, Azure 20%, Google 13%) is the reference here; Q4 2025 readings were about 28%, 21% and 14%.

Forward-looking figures, including capex plans, Anthropic's 3.5 GW of TPU capacity for 2027, the 4.3 million TPU shipment projection and the capacity trajectories in this report, are guidance or projections and have been revised repeatedly.

Not covered in depth: IBM Cloud, Baidu AI Cloud and ByteDance's Volcano Engine, whose disclosures are not comparable.

Sources

Capacity, energy and methods (new in this edition)

• DCD: AWS added 3.8 GW in the last year and MLQ: Amazon Q3 2025 call excerpt
• The Register: Amazon Q4 2025 earnings
• MLQ: AWS Q2 2026 and the $220B capex plan
• DCD: Amazon 2025 Sustainability Report
• AWS Global Infrastructure: Regions and AZs
• High Scalability: Hamilton at re:Invent 2014, Network World: Hamilton at re:Invent 2016, DCD: AWS mega-growth
• AIwire: A rare peek into the scale of AWS (2014)
• Data Center Frontier: regional clusters power Amazon's cloud
• MMCG Invest: AWS at scale
• Microsoft FY2026 Q1 earnings call
• Trellis: Microsoft 2026 sustainability report, Microsoft 2026 environmental data fact sheet, Microsoft: energy use at its datacenters
• Google 2026 Environmental Report, Google Data Centers: operating sustainably
• TechCrunch: Google data-center energy use doubled in four years, TechTimes: Google electricity up 37%
• Alibaba FY2025 ESG report (SEC Form 6-K exhibit)
• Tracenable: Tencent energy data from its 2025 ESG report, MarketScreener: Tencent 2025 ESG report
• DCD: Huawei Gui'an data center, HuaweiCentral: Huawei Cloud's five China data centers
• Oracle Q2 FY2026 call transcript, Oracle Q3 FY2026 call transcript, DCD: Oracle 400 MW and 10 GW secured
• ERP Today: Oracle Q4 FY2026, Converge Digest: Oracle Q1 FY2027
• Synergy: hyperscale share of capacity (Apr 2026), Synergy: hyperscale data-center count (Mar 2025)
• Yale Clean Energy Forum: data-center energy consumption (IEA figures)

Scale comparisons (new in this edition)

• EIA: US nuclear fleet, 94 reactors and 96.9 GW of capacity
• EIA: how much electricity an American home uses
• Sunbird: DGX H100 rack power, 10.2 kW per system
• Rack, row and building figures are the Microsoft Fairwater and James Hamilton sources listed above.

Market, footprint and financials (first edition)

• Synergy: Big Three hold 63%
• Futurum: Amazon Q4 FY2025, Global Data Center Hub: Amazon Q4 2025
• Amazon: Project Rainier activated, Technology Magazine: Project Rainier
• Microsoft Azure: cloud infrastructure in Asia, DCD: Azure $75B and 2 GW in FY2025, DCD: Fairwater Atlanta
• Microsoft FY2025 Form 8-K, Microsoft FY2025 Form 10-K
• Google Cloud locations, Data Center Knowledge: hyperscalers in 2026, The Next Web: Ironwood TPU
• Alibaba Cloud: international expansion, DCD: Alibaba ships 470,000 AI chips, TrendForce: T-Head shipments, Caixin: T-Head spinoff, Digital Commerce 360: Alibaba AI plans
• CNBC: Huawei 2025 cloud revenue
• CIO Dive: Oracle passes 100 regions, Futurum: Oracle Q4 FY2025, Futurum: Oracle Q3 FY2025, Oracle FY2026 Q1 results
• TMTPost: Tencent Cloud reaches profitability at scale
• CoreWeave FY2025 Form 10-K summary
• Tom's Hardware: $725B of 2026 hyperscaler capex