Skip to content
By NiLu Updated

Data centers — overview of infrastructure, expansion, and limits

How data centers combine computing, power, cooling, territory, and capital—and why their limits emerge in different ways.

In brief

  • Data centers combine IT equipment with power, cooling, connectivity, capital, territory, and authorization.
  • Announced, contracted, installed, energized, usable, and consumed capacity are distinct states.
  • AI increases the density of some configurations, but it neither describes the entire fleet nor explains expansion on its own.
  • Measurement gaps prevent universal rankings, not explanations of the known mechanisms.

In brief

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. [1] [2]

How expansion happens

The sequence begins with demand for digital services. Different workloads create distinct processing, storage, latency, and reliability requirements. These requirements are translated into servers and racks, which require critical power, cooling, and connectivity. For capacity to become usable, the project still requires land, permitting, financing, equipment, a grid connection, and operations. [1] [2]

Therefore, announced, contracted, installed, energized, usable, and actually consumed capacity are distinct states.

Global scale and local concentration

Global estimates help gauge the trend, but they must remain tied to the edition and scenario that produced them. In one edition, the IEA estimated approximately 415 TWh for data centers in 2024; a later update used about 485 TWh in 2025 and a central scenario close to 950 TWh in 2030. These values do not form a single range, and the projection for 2030 is not a future fact. [3] [4]

Impacts appear locally. In Ireland, data centers accounted for 23% of measured electricity consumption in 2025, compared with 5% in 2015. This denominator does not represent all final energy and cannot be generalized globally. [5]

AI, density, and consumption

AI configurations can reach densities far higher than those observed in conventional facilities. An NVIDIA NVL72 configuration is documented at approximately 120 kW per rack. That does not make it a typical rack. In the 2025 Uptime survey, 82% of 709 responding operators reported that the highest density present in their facilities remained below 30 kW per rack. [6] [7]

A component’s TDP does not measure server or site consumption. Power benchmarks measure defined systems, workloads, and scenarios; on their own, they also do not provide annual consumption for a data center. [8] [9] [10]

Benefits, costs, and location

In the case studied by Virginia, most of the modeled economic impact was associated with construction, not permanent operational jobs. In Washington, a tax preference was used only for renovations during the audited period and did not produce new data centers. These cases do not represent all markets, but they demonstrate why investment, employment, tax revenue, and additionality must be assessed separately. [11] [12]

Physical location likewise does not amount to full sovereignty. Jurisdiction, operations, cryptographic keys, portability, software, chips, cables, and continuity must be assessed separately. [13]

What remains unknown

There is no public, comparable global breakdown of capacity by workload. Nor is there a reliable global percentage of installed capacity that is actually usable, a comparable net fiscal return across programs, or a single defensible index of computing sovereignty.

These gaps are part of the editorial result. They prevent rankings and false precision, but they do not prevent explanations of the known mechanisms.

References

  1. U.S. EPA ENERGY STAR — Data center rating FAQ and PUE definition. U.S. EPA. Accessed July 24, 2026.
  2. LBNL — 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory. Accessed July 24, 2026.
  3. IEA — Energy and AI (2025). IEA. Accessed July 22, 2026.
  4. IEA — Key Questions on Energy and AI (2026). IEA. Accessed July 24, 2026.
  5. CSO Ireland — Data Centres Metered Electricity Consumption 2025. Central Statistics Office Ireland. Accessed July 24, 2026.
  6. NVIDIA — DGX GB Rack Scale Systems Hardware Guide. NVIDIA. Accessed July 24, 2026.
  7. Uptime Institute — 2025 Global Data Center Survey. Uptime Institute. Accessed July 24, 2026.
  8. NVIDIA — H100 Tensor Core GPU specifications. NVIDIA. Accessed July 22, 2026.
  9. MLCommons — MLPerf Inference benchmark documentation. MLCommons. Accessed July 24, 2026.
  10. MLCommons — MLPerf Inference power measurement methodology. MLCommons. Accessed July 24, 2026.
  11. JLARC — Data Centers in Virginia (2024). Virginia JLARC. Accessed July 24, 2026.
  12. Washington JLARC — Data Centers in Urban Counties: 2026 Tax Preference Review. Washington JLARC. Accessed July 24, 2026.
  13. OECD — Data localisation trends and challenges. OECD. Accessed July 24, 2026.