AI Market Growth Statistics Updated November 2026

Human Written & Fact Checked · Updated

100 AI Infrastructure Statistics for 2026

Can you tell the difference between AI infrastructure revenue, cloud revenue and capital spending? The same equipment can appear in a supplier's sales and a customer's investment account. Adding those amounts produces double counting, not a market size.

This page keeps original company disclosures beside energy-system research while preserving their definitions. Fiscal periods, calendar periods, future commitments, management guidance and modelled scenarios remain separate. No broad cloud or consolidated spending total is relabelled as AI-only investment.

Headline figures

  • NVIDIA Data Center revenue, FY2027 Q2: $89.023 billion. (NVIDIA, Aug 2026)
  • NVIDIA supply and capacity commitments: $279 billion. As of 26 July 2026. (NVIDIA, Aug 2026)
  • FY2027 Q1: Microsoft quarterly capital expenditures, including finance leases: $45.2 billion. (Microsoft, Oct 2026)
  • FY2027 Q1: Microsoft reported roughly 20% shorter quarter-over-quarter dock-to-live times for new GB300s in its largest regions. (Microsoft, Oct 2026)
  • Calendar Q3 2026: Alphabet quarterly capital expenditures: $51.4 billion. (Alphabet, Oct 2026)
  • Calendar Q3 2026: Alphabet first-party model APIs processed nearly 32 billion tokens per minute. (Alphabet, Oct 2026)
  • AWS quarterly net sales, Q3 2026: $45.984 billion. (Amazon, Oct 2026)
  • Amazon expected calendar-2026 cash capital expenditures of approximately $220 billion. (Amazon, Oct 2026)
  • IEA estimated global data-centre electricity consumption at 485 TWh in 2025. (International Energy Agency, Apr 2026)
  • IEA estimated 20–25 GW of battery storage could be installed in data centres globally by 2030. (International Energy Agency, Apr 2026)
  • IEA identified 134 policies and regulations addressing data-centre integration into electricity systems. (International Energy Agency, Oct 2026)
  • IEA found nearly 40% of identified AI-dedicated data-centre capacity in areas of high or extremely high baseline water stress. (International Energy Agency, Oct 2026)

What does NVIDIA disclose about the infrastructure buildout?

Supplier revenue and future commitments reveal different parts of the buildout. Keep them separate: buying commitments are not delivered capacity, and chip-supplier sales can also sit inside a cloud customer's capital budget.

How large is NVIDIA's Data Center platform?

  • NVIDIA Data Center revenue, FY2027 Q1: $75.246 billion. (NVIDIA, Aug 2026)
  • NVIDIA Data Center revenue, FY2026 Q2: $41.096 billion. (NVIDIA, Aug 2026)
  • NVIDIA Data Center revenue growth, FY2027 Q2 year over year: 117%. (NVIDIA, Aug 2026)
  • NVIDIA Data Center revenue growth, FY2027 Q2 quarter over quarter: 18%. (NVIDIA, Aug 2026)
  • Hopper shipments to China represented less than 1% of NVIDIA Data Center revenue in FY2027 Q2. (NVIDIA, Aug 2026)

The Data Center platform is the relevant infrastructure sales measure here, instead of the company's total revenue. Its latest result appears in Main Takeaways; the comparisons below retain NVIDIA's fiscal-quarter labels. The latest quarter ended in July, so it should not be relabelled as a calendar third quarter. Revenue measures sales recognised by the supplier, not the operating capacity installed by every customer or the utilisation of the systems sold.

How are hyperscale and other AI-infrastructure customers separated?

  • NVIDIA Hyperscale revenue, FY2027 Q2: $48.710 billion. (NVIDIA, Aug 2026)
  • NVIDIA Hyperscale revenue, FY2027 Q1: $43.050 billion. (NVIDIA, Aug 2026)
  • NVIDIA Hyperscale revenue, FY2026 Q2: $24.168 billion. (NVIDIA, Aug 2026)
  • NVIDIA AI Clouds, Industrial & Enterprise revenue, FY2027 Q2: $40.313 billion. (NVIDIA, Aug 2026)
  • NVIDIA AI Clouds, Industrial & Enterprise revenue, FY2027 Q1: $32.196 billion. (NVIDIA, Aug 2026)
  • NVIDIA AI Clouds, Industrial & Enterprise revenue, FY2026 Q2: $16.928 billion. (NVIDIA, Aug 2026)

NVIDIA reports Hyperscale separately from AI Clouds, Industrial & Enterprise. The latter includes customers with different business models and end markets, so it should not be described simply as enterprise IT spending. The source reclassified a customer and recast its comparative periods. Using those recast figures preserves the basis of comparison within the release, instead of mixing this quarter's categories with previously published classifications.

What future infrastructure commitments has NVIDIA made?

  • NVIDIA cloud-service agreement commitments: $29 billion. (NVIDIA, Aug 2026)
  • NVIDIA uncommenced data-centre lease commitments for its own operations: $25 billion. (NVIDIA, Aug 2026)
  • NVIDIA committed capital expenditures: $8 billion. (NVIDIA, Aug 2026)
  • NVIDIA additional AI-cloud agreement commitments: $36 billion. (NVIDIA, Aug 2026)
  • NVIDIA uncommenced data-centre lease commitments intended for third parties: $20 billion. (NVIDIA, Aug 2026)
  • NVIDIA supply and capacity commitments for the remainder of FY2027: $92 billion. (NVIDIA, Aug 2026)
  • NVIDIA supply and capacity commitments for FY2028: $87 billion. (NVIDIA, Aug 2026)

The commitment schedules separate procurement, cloud services, leases and capital expenditure. Their future payment periods differ, and some arrangements support third parties instead of NVIDIA's own operations. The headline supply commitment appears above. Its fiscal-period allocations below are components, not additional spending to add to the total. These obligations are distinct from recognised quarterly expenses or completed data-centre projects.

How is Microsoft expanding and financing capacity?

Cash expenditure alone misses part of Microsoft's infrastructure investment. Its capex definition includes finance leases at commencement, while its cloud revenue includes services beyond AI.

How do Microsoft's capex and cash payments differ?

  • Microsoft described roughly three quarters of quarterly capex as short-lived assets, primarily CPUs and GPUs. (Microsoft, Oct 2026)
  • Microsoft quarterly finance leases: $7.4 billion. (Microsoft, Oct 2026)
  • Microsoft quarterly cash paid for property, plant and equipment: $39.4 billion. (Microsoft, Oct 2026)
  • Microsoft quarterly company-wide operating cash flow: $58 billion. (Microsoft, Oct 2026)
  • Microsoft quarterly company-wide free cash flow: $18.6 billion. (Microsoft, Oct 2026)

Microsoft includes the full principal value of a finance lease in capex when the lease commences. Cash paid for property, plant and equipment is a separate cash-flow measure. The headline capex result is in Main Takeaways; the remaining amounts below describe the same fiscal quarter. Company-wide cash generation provides financing context, but it does not establish the profitability of AI-specific infrastructure or reconcile every timing difference between cash and commitments.

What do cloud revenue and spending guidance indicate?

  • Microsoft Cloud quarterly revenue: $63.9 billion. (Microsoft, Oct 2026)
  • Microsoft Cloud revenue growth year over year: 30% as reported. (Microsoft, Oct 2026)
  • Azure quarterly revenue growth year over year: 48% as reported. (Microsoft, Oct 2026)
  • Microsoft Cloud quarterly gross margin: 66%. (Microsoft, Oct 2026)
  • Microsoft expected next-quarter capex to exceed $50 billion. (Microsoft, Oct 2026)
  • Microsoft expected calendar-2026 capex of approximately $170 billion. (Microsoft, Oct 2026)

The reported cloud results cover both AI and other services. Azure is also broader than AI infrastructure, so its growth cannot be reported as an AI-only growth rate. Actual quarterly outcomes are separated from management's forward-looking capex expectations below. The next fiscal quarter and the calendar-year spending outlook use different time windows. Neither forecast establishes how much new capacity will be ready for customers by a particular date.

What operational improvements does Microsoft report?

  • Microsoft added 24 data centres during the quarter. (Microsoft, Oct 2026)
  • Microsoft's new quarterly data-centre additions spanned five continents. (Microsoft, Oct 2026)
  • Microsoft reported over 50% higher quarter-over-quarter throughput for Copilot workloads. (Microsoft, Oct 2026)
  • Microsoft said Maia 200 delivered up to 1.7 times more tokens per watt across Teams and Copilot AI experiences. (Microsoft, Oct 2026)
  • Microsoft projected Maia 300 would deliver 40% more tokens per watt than leading GPU-based systems. (Microsoft, Oct 2026)

Microsoft describes both new facilities and improvements to its existing fleet. The dock-to-live result in Main Takeaways applies to specified new GPU deployments in its largest regions, not every site worldwide. The company's throughput and chip-efficiency statements below are management-reported comparisons, not independent benchmark results. The future Maia claim is explicitly forward-looking; it should not be combined with the currently deployed system's reported performance.

What is Alphabet investing in technical infrastructure?

The mix of equipment and buildings counts as much as the spending total. Alphabet's useful-life assumptions also explain why today's cash investment and tomorrow's depreciation follow different schedules.

How is Alphabet's investment divided across asset types?

  • Servers and networking equipment represented approximately 65% of Alphabet's quarterly technical-infrastructure investment. (Alphabet, Oct 2026)
  • Data centres represented approximately 35% of Alphabet's quarterly technical-infrastructure investment. (Alphabet, Oct 2026)
  • Alphabet assigned a general useful life of six years to servers and networking equipment. (Alphabet, Oct 2026)
  • Alphabet reported data-centre useful lives spanning seven to forty years. (Alphabet, Oct 2026)
  • Alphabet maintained calendar-2026 capex guidance of $195–205 billion. (Alphabet, Oct 2026)

Alphabet identifies equipment and data centres as separate components of technical-infrastructure investment. The headline quarterly capex amount appears above. Its composition below is a share of technical-infrastructure investment, not a claim that all company spending is AI-specific. Useful lives are accounting estimates instead of promised hardware replacement dates. The annual guidance remains an expectation, while the quarterly allocation describes investment already reported for the period.

How do commitments, backlog and cash generation differ?

  • Alphabet future purchase commitments at quarter end: $924 billion. (Alphabet, Oct 2026)
  • Alphabet pre-commencement leases and other obligations at quarter end: $152 billion. (Alphabet, Oct 2026)
  • Google Cloud backlog at quarter end: $561 billion. (Alphabet, Oct 2026)
  • Alphabet expected just over 55% of total Cloud backlog to become revenue within the next twenty-four months. (Alphabet, Oct 2026)
  • Alphabet company-wide quarterly operating cash flow: $42.3 billion. (Alphabet, Oct 2026)
  • Alphabet company-wide quarterly free cash flow: negative $9.1 billion. (Alphabet, Oct 2026)

Purchase commitments describe prospective outflows; Cloud backlog describes contracted future revenue. The two are not interchangeable measures of demand, and neither equals cash already collected. Lease obligations also have their own timing. The cash-flow figures below are company-wide results, so they include activity outside Google Cloud and AI. Their purpose is to show financing context, without assigning all cash generation or consumption to one infrastructure programme.

What does Alphabet report about model infrastructure?

  • Google Cloud quarterly revenue grew 103% year over year. (Alphabet, Oct 2026)
  • Alphabet first-party model API token throughput increased 40% from the previous quarter. (Alphabet, Oct 2026)
  • Google reported nearly 100 customers using TPUs. (Alphabet, Oct 2026)
  • Google claimed TPU 8t offered up to 2.7 times the training performance of the previous TPU generation. (Alphabet, Oct 2026)
  • Google claimed TPU 8i offered 80% better inference price-performance than the previous TPU generation. (Alphabet, Oct 2026)

Alphabet's model-throughput and TPU statements describe its own operating platform. The token-throughput level appears in Main Takeaways; its reported sequential change below uses the same company measure. Tokens are not a standard unit of useful work across different models or applications. Product performance comparisons are vendor claims under the stated generation comparison, not universal gains for every customer workload. Google Cloud growth includes core services as well as AI and TPU-system sales.

What do AWS results and Amazon's spending show?

AWS is a broad cloud business, and Amazon's capital accounts are broader still. The figures below retain that distinction instead of turning consolidated investment into an AI-only market estimate.

How did AWS perform in the latest quarter?

  • AWS quarterly net sales, Q3 2025: $33.006 billion. (Amazon, Oct 2026)
  • AWS quarterly operating income, Q3 2026: $17.773 billion. (Amazon, Oct 2026)
  • AWS quarterly operating income, Q3 2025: $11.434 billion. (Amazon, Oct 2026)
  • AWS quarterly operating expenses, Q3 2026: $28.211 billion. (Amazon, Oct 2026)

AWS reported higher sales and operating income than in the corresponding prior-year quarter. The latest sales amount is in Main Takeaways; the accompanying comparison and cost measures below come from the same segment table. These results include the full AWS service portfolio. They do not identify AI-only revenue, standalone accelerator revenue or the number of customers running generative models. Quarterly operating expenses also should not be treated as equivalent to capital investment.

What do the year-to-date AWS accounts show?

  • AWS nine-month net sales: $125.803 billion. (Amazon, Oct 2026)
  • AWS nine-month operating income: $48.555 billion. (Amazon, Oct 2026)
  • AWS nine-month operating expenses: $77.248 billion. (Amazon, Oct 2026)
  • Amazon company-wide technology and infrastructure expense, nine months: $97.539 billion. (Amazon, Oct 2026)

The year-to-date segment accounts aggregate the calendar year through September. They are cumulative results, not annualised run rates. They provide a longer reporting window than the latest quarter, but include that quarter and therefore should not be added to it. Amazon's technology and infrastructure expense is a consolidated operating-expense line, not the AWS segment's capital budget. Keeping the separate labels visible prevents a broad corporate accounting measure from becoming a misleading estimate of AI investment or cloud construction spending.

Which Amazon capital-accounting measures should remain separate?

  • Amazon company-wide quarterly purchases of property and equipment: $77.259 billion. (Amazon, Oct 2026)
  • Amazon company-wide quarterly proceeds from property/equipment sales and incentives: $1.381 billion. (Amazon, Oct 2026)
  • Amazon company-wide quarterly property/equipment acquired through finance leases, net of modifications: $1.472 billion. (Amazon, Oct 2026)
  • Amazon company-wide quarterly finance-lease principal repayments: $433 million. (Amazon, Oct 2026)

Amazon's capital accounts distinguish cash purchases, proceeds, leased additions and lease repayments. These are company-wide amounts that also include activities beyond AWS. The annual cash-capex outlook is in Main Takeaways, while the quarterly actuals below preserve the source definitions. Acquiring an asset under a finance lease differs from repaying its principal later; counting both as new investment would double count parts of the same financing arrangement. Sales proceeds and incentives must also retain their sign and treatment.

How much electricity could data centres require?

Electricity projections are scenarios, not completed infrastructure. The IEA separates overall data-centre demand from AI-focused facilities and tests how efficiency and deployment constraints can alter the path.

What is the IEA's electricity-demand outlook?

The IEA's observed demand estimates and central projections describe different periods. The historical consumption level is in Main Takeaways; growth estimates and forward-looking observations below retain the source's distinctions between all data centres and AI-focused facilities. The projection is not an AI-only electricity total. It also does not establish a fixed power draw for each facility, because annual electricity use depends on utilisation and the mix of computing workloads.

How are power density and onsite-system requirements changing?

  • IEA estimated AI-server power density increased elevenfold between 2020 and 2025. (International Energy Agency, Apr 2026)
  • IEA projected a further fourfold increase in AI-server power density by 2027. (International Energy Agency, Apr 2026)
  • IEA projected a cutting-edge server rack could have peak demand equivalent to 65 households by 2027. (International Energy Agency, Apr 2026)
  • IEA estimated reliable onsite gas supply for critical, variable data-centre loads required generation capacity overbuild of 30–70% relative to demand. (International Energy Agency, Apr 2026)

Power density concerns demand concentrated in equipment, while annual electricity consumption also depends on operating patterns. The IEA's rack comparison is illustrative, not a claim that every installed rack has the same load. Its battery outlook appears in Main Takeaways. The remaining observations distinguish historical changes, anticipated equipment development and the generation overbuild needed in a modelled onsite-gas system. Those figures should not be read as universal design requirements for every data centre.

What do tracked projects and query scenarios establish?

  • IEA estimated 15–27 GW of onsite natural-gas generation could power data centres by 2030. (International Energy Agency, Apr 2026)
  • IEA satellite tracking found about one fifth of identified onsite-gas projects had begun land clearing or construction. (International Energy Agency, Apr 2026)
  • IEA satellite tracking found AI-factory capacity had more than tripled over the preceding eighteen months. (International Energy Agency, Apr 2026)
  • IEA modelled replacing conventional internet searches with simple AI text queries at less than 4 TWh annually. (International Energy Agency, Apr 2026)
  • That illustrative simple-query scenario was equivalent to less than 1% of contemporary data-centre consumption. (International Energy Agency, Apr 2026)

The IEA combines observed project progress with scenario analysis, and the distinction is essential. Satellite evidence of clearing or construction does not prove that a facility is commissioned or connected. The simple-query calculation is an illustrative replacement scenario, not measured electricity consumption by today's AI industry. Its relative share and absolute quantity express different aspects of the same scenario. More demanding workloads and changes in usage can produce different system-level outcomes.

What constrains data-centre grid integration?

A connection request is not an operating data centre. Geography, project maturity, flexibility and environmental exposure determine how much announced capacity can become usable infrastructure.

How extensive is policy activity around grid integration?

The IEA's stocktake records the coverage and timing of identified policy activity. The policy count appears in Main Takeaways. Coverage figures below describe the study's jurisdictional sample, not a measure of compliance or enforcement. A recently announced policy can also differ materially from an implemented rule. The review helps establish where integration is receiving attention, but it does not show that every connection queue has been resolved or that the same requirements apply everywhere.

How do clustering and flexibility affect expansion?

  • IEA estimated around 60% of global data-centre capacity was in geographic clusters in 2025. (International Energy Agency, Oct 2026)
  • IEA estimated the clustered share would exceed 65% if projects under construction were completed. (International Energy Agency, Oct 2026)
  • IEA estimated the clustered share would exceed 70% if all announced projects materialised. (International Energy Agency, Oct 2026)
  • IEA estimated non-firm grid connections could unlock 750–900 GW of additional hosting capacity globally. (International Energy Agency, Oct 2026)
  • IEA estimated annual rental of AI-optimised data-centre capacity could cost $20–30 billion per GW. (International Energy Agency, Oct 2026)
  • IEA estimated AI-optimised compute capacity rental could cost one hundred times equivalent battery capacity. (International Energy Agency, Oct 2026)

The IEA's spatial analysis distinguishes operating capacity from conditional project pipelines. The pipeline shares are not forecasts that every announced facility will be built. Its non-firm connection estimate concerns potential grid hosting capacity, not data-centre capacity already delivered. The rental-cost comparison illustrates why interrupting compute can be expensive; it is not a retail electricity price or a site-specific quotation.

What emissions and water exposure does the IEA identify?

  • IEA associated data-centre electricity consumption with around 150 million tonnes of CO2 emissions at the time of its assessment. (International Energy Agency, Oct 2026)
  • IEA's central scenario projected around 350 million tonnes of data-centre electricity-related CO2 emissions in 2035. (International Energy Agency, Oct 2026)
  • IEA projected this would represent around 1.5% of global energy-related emissions in 2035. (International Energy Agency, Oct 2026)

The emissions estimates concern electricity used by data centres, instead of a complete lifecycle footprint of AI models or semiconductor manufacturing. The future values belong to a central scenario and depend on the assumed electricity supply mix. Water-stress exposure is in Main Takeaways and describes where identified capacity is located, not measured water consumption at each facility. None of these aggregate indicators determines the environmental impact of a particular proposed project without local evidence.

Glossary

TPU
Google's tensor processing unit.
AWS
Amazon Web Services.
GW
Gigawatt, a unit of power capacity.
TWh
Terawatt-hour, a unit of energy consumed.

Methodology and sources

Every figure on this page comes from the organisation that produced it: regulators, statistics offices, public bodies, research institutes, and companies reporting their own data or surveys. Each one links to the original report or dataset and carries its release date. Survey results keep their population and fieldwork period, and financial figures keep the measure the producer used. Figures from different producers are not added together or ranked against each other. The page is reviewed monthly and updated when a producer releases newer data.