Data centers — overview of infrastructure, expansion, and limits
Scope: How digital demand becomes physical infrastructure; which resources and bottlenecks constrain expansion; and how costs, impacts, and power are distributed.
Executive summary
Data centers are not abstract digital infrastructure. They are cyber-physical systems whose expansion depends on the convergence of compute, power, cooling, connectivity, capital, equipment, territory, and authorization.
Complete executive summary
- Data centers are physical systems that combine servers, accelerators, memory, storage, and networks with buildings, power, cooling, connectivity, capital, and authorization. The sector's expansion cannot be explained by a single component. AI accelerates part of the demand; power has become a significant constraint; chips, networks, electrical equipment, territory, and regulation remain interdependent.
- Announced, contracted, installed, energized, usable, and actually consumed capacity are distinct states.
- Measurement gaps prevent rankings and false precision, but they do not prevent explanations of known mechanisms.
System map and relations
Audience: Readers, researchers, investors, operators, and policymakers.
Out of scope
- Technical, environmental, fiscal, or legal details remain in the specialized Dossiers.
- The Overview does not reproduce the complete evidence bases of the four specialized Dossiers.
demand for digital services
↓
workloads + delivery model
↓
servers, accelerators, memory, storage, and networking
↓
racks, critical power, cooling, and building
↓
grid connection + fiber + cables + territory
↓
contracts, financing, incentives, and regulation
↓
installed → energized → usable → actually consumed capacity
↓
economic, environmental, territorial, and strategic effectsSpecialized Dossiers
DATA
Claims with explicit sources, measurement context, limitations, and confidence.
DATA · fact
DC-OV-D01 · A data center combines IT and supporting infrastructure
A data center combines servers, storage, and networking with power supply, conversion, redundancy, cooling, security, and a building.
Limitations
- Implementations vary by facility type and availability requirements.
DATA · fact · taxonomy
DC-OV-D02 · Cloud is not a workload
Cloud is a delivery and operating model; AI, video, databases, web services, analytics, HPC, and storage are applications or workload classes that can share the same infrastructure.
Limitations
- Application categories can overlap in their use of resources, making mutually exclusive shares difficult to determine.
DATA · estimate · prediction
DC-OV-D03 · Global projections are specific to an edition and scenario
The IEA estimated about 415 TWh for data centers in 2024 in the 2025 Energy and AI edition; a later update used approximately 485 TWh in 2025 and a central scenario close to 950 TWh in 2030.
Limitations
- The values belong to different editions, base years, and assumptions; they do not form a single range.
DATA · fact
DC-OV-D04 · Concentration makes local impact disproportionate
Data centers accounted for 23% of Ireland's measured electricity consumption in 2025, compared with 5% in 2015.
Limitations
- This is not a share of all final energy and is not a global measure.
DATA · comparison
DC-OV-D05 · An AI-dense configuration does not represent a typical rack
NVIDIA documentation reports approximately 120 kW for an NVL72 configuration; in the 2025 Uptime survey, 82% of 709 responding operators did not have a rack above 30 kW as the highest-density rack in the facility.
Limitations
- A product specification and an operator survey cover different populations; the survey is not a distribution of all racks.
DATA · comparison
DC-OV-D06 · TDP and benchmarks do not measure annual site consumption
TDP is a component property; MLPerf Power measures a system's AC power during a defined benchmark; neither measure, in isolation, provides a data center's annual consumption.
Limitations
DATA · estimate · comparison
DC-OV-D07 · Water use depends on boundaries and conditions
An LBNL study found a variation of more than four orders of magnitude in water consumption per workload among the evaluated combinations of equipment, utilization, cooling, climate, power grid, and hardware refresh.
Limitations
- This does not represent the observed distribution of global sites and does not justify a universal number 'per query.'
DATA · comparison
DC-OV-D08 · A renewable contract does not prove hourly physical supply
Market-based Scope 2 records contractual instruments; on its own, it does not demonstrate that consumption was supplied by carbon-free generation in every hour and location.
Limitations
- Distinguish the current rule from future proposals for hourly matching and deliverability.
DATA · estimate
DC-OV-D09 · In the Virginia case, modeled economic impact is predominantly construction
The Virginia model estimated 74,000 jobs associated with the sector: 59,000 tied to construction and 15,000 to operations, including direct, indirect, and induced effects.
Limitations
- This is not equivalent to permanent employees of operators and does not represent other jurisdictions.
DATA · fact · estimate
DC-OV-D10 · Use of an incentive does not prove additionality
The Washington audit found fewer than ten beneficiaries, only refurbishments, no new construction under the preference, and an estimated US$42.4 million in savings between 2023–2026.
Limitations
- This does not measure every incremental investment and does not support national generalization.
DATA · estimate
DC-OV-D11 · Regional electricity consumption is not a share of compute
For 2024, the IEA attributed 45%, 25%, and 15% of estimated data-center electricity consumption to the United States, China, and Europe.
Limitations
- This does not measure GPUs, FLOPs, servers, or usable capacity.
DATA · estimate · comparison
DC-OV-D12 · Cables are critical, but cable count does not measure resilience
The ITU estimates that submarine cables carry more than 99% of international data traffic; faults, repairs, capacity, route diversity, and unavailability are distinct metrics.
Limitations
DATA · comparison
DC-OV-D13 · Location is not comprehensive sovereignty
Keeping data and data centers within a territory does not, on its own, determine effective jurisdiction, operational control, keys, portability, technological substitutability, or continuity.
Limitations
INFO
INFO · derived synthesis
DC-OV-I01 · Expansion is the coordination of dependencies
Chips without power create unproductive inventory; power without hardware, connectivity, or demand creates idle capacity; neither becomes operational without permits and contracts.
INFO · derived synthesis
DC-OV-I02 · The bottleneck moves through the system
The constraint can move among chips, HBM, packaging, networking, electrical equipment, grid connection, cooling, construction, and commissioning.
derivedFrom: DC-OV-D01, DC-OV-D05
Provenance note: The canonical derivation also points to the specialized Dossiers on Operations and Energy, without identifying additional claims in the Overview.
INFO · derived synthesis
DC-OV-I03 · Unit efficiency and aggregate consumption can diverge
Lower energy use per task does not determine total consumption; volume, price, utilization, and new uses also change.
INFO · derived synthesis
DC-OV-I04 · Small global impacts can be large locally
A global share of electricity or water does not describe pressure on a grid node, watershed, or municipality.
INFO · derived synthesis
DC-OV-I05 · Benefits and costs appear in different institutions
A state tax expenditure, municipal revenue, utility investment, electricity rates, and temporary employment can coexist; they should not be added together as a net benefit without a counterfactual.
INFO · derived synthesis
DC-OV-I06 · Location is simultaneously an advantage and a dependency
Power, fiber, and regulation attract capacity, but chips, cloud, software, cables, and jurisdiction preserve external dependencies.
INFO · derived synthesis
DC-OV-I07 · A projection or incentive does not demonstrate delivered capacity
Demand projections and incentive programs do not, on their own, demonstrate operational capacity; deployment still depends on financing, equipment, grid connection, permits, construction, and operations.
Uncertainties
UNCERTAINTY · not publicly measured
DC-OV-UNC-01 · Global share by workload
What we know: Adoption and presence are not equivalent to energy, servers, or compute share.
Limitations
Editorial treatment: Do not publish an artificial global breakdown.
Confidence: high confidence in the gap
Last reviewed: 2026-07-25
UNCERTAINTY · not publicly measured
DC-OV-UNC-02 · Installed versus usable capacity
What we know: Power, cooling, space, networks, and reliability can produce stranded capacity.
Limitations
Editorial treatment: Explain the mechanisms without a global percentage.
Last reviewed: 2026-07-25
UNCERTAINTY · incompatible methods
DC-OV-UNC-03 · Comparable environmental impact by site
What we know: PUE, WUE, water, emissions, and embodied carbon depend on boundary, climate, and operations.
Limitations
Editorial treatment: Preserve the site, period, denominator, and method.
Last reviewed: 2026-07-25
UNCERTAINTY · not publicly measured
DC-OV-UNC-04 · Net fiscal return
What we know: Gross revenue and tax expenditures exist at different levels of government.
Limitations
Editorial treatment: Do not produce an incentive ranking.
Last reviewed: 2026-07-25
UNCERTAINTY · imperfect proxy
DC-OV-UNC-05 · Global compute and sovereignty
What we know: Electricity, racks, facilities, and location measure partial dimensions.
Limitations
Editorial treatment: Use a multidimensional framework, not a single index.
Last reviewed: 2026-07-25
Methodology
- DATA
- DATA records verifiable claims with explicit sources, measurement context, limitations, and confidence.
- INFO
- INFO records synthesis or interpretation and always declares the IDs of the claims from which it is derived.
- Uncertainty
- An Uncertainty records a gap as an editorial result; it remains explicit and is not filled with a fragile estimate.
- Update
- A review may correct or update the projection only from the canonical Dossier with editorial authorization; material changes increment the public version and are recorded in the corrections history.
- Public version
- 1.0.0
- Last verified
- 2026-07-29
- Review after
- 2026-08-24
Public corrections
- No public corrections recorded.
References
-
src-epa-energy-star-datacenter-faq-pue· U.S. EPA ENERGY STAR — Data center rating FAQ and PUE definition. U.S. EPA. Accessed . -
src-lbnl-us-data-center-energy-2024· LBNL — 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory. Accessed . -
src-nist-sp800-145-cloud-definition· NIST — The Definition of Cloud Computing (SP 800-145). NIST. Accessed . -
src-iea-energy-and-ai-2025· IEA — Energy and AI (2025). IEA. Accessed . -
src-iea-key-questions-energy-ai-2026· IEA — Key Questions on Energy and AI (2026). IEA. Accessed . -
src-cso-ireland-datacenter-electricity-2025· CSO Ireland — Data Centres Metered Electricity Consumption 2025. Central Statistics Office Ireland. Accessed . -
src-nvidia-dgx-gb-rack-hardware-guide· NVIDIA — DGX GB Rack Scale Systems Hardware Guide. NVIDIA. Accessed . -
src-uptime-global-datacenter-survey-2025· Uptime Institute — 2025 Global Data Center Survey. Uptime Institute. Accessed . -
src-nvidia-h100-specifications· NVIDIA — H100 Tensor Core GPU specifications. NVIDIA. Accessed . -
src-mlcommons-mlperf-inference· MLCommons — MLPerf Inference benchmark documentation. MLCommons. Accessed . -
src-mlcommons-mlperf-power· MLCommons — MLPerf Inference power measurement methodology. MLCommons. Accessed . -
src-lbnl-water-use-workloads-2025· LBNL — The water use of data center workloads (2025). Lawrence Berkeley National Laboratory. Accessed . -
src-ghg-protocol-scope2-guidance-2015· GHG Protocol — Scope 2 Guidance. GHG Protocol / WRI. Accessed . -
src-jlarc-data-centers-virginia-2024· JLARC — Data Centers in Virginia (2024). Virginia JLARC. Accessed . -
src-wa-jlarc-urban-data-centers-tax-preference-2026· Washington JLARC — Data Centers in Urban Counties: 2026 Tax Preference Review. Washington JLARC. Accessed . -
src-itu-submarine-cable-resilience-2026· ITU — International Advisory Body on Submarine Cable Resilience: Working Groups Report 2026. ITU. Accessed . -
src-oecd-data-localisation-2020· OECD — Data localisation trends and challenges. OECD. Accessed .