Big AI is an industrial system, not just software. Training and serving large models depend on hyperscale data centres, accelerator chips, high-capacity networks, electricity, cooling water, land and very large pools of capital. Cloud companies increasingly secure power, design chips and cooling systems, and negotiate for sites in ways that resemble industrial firms and utilities.
The immediate constraint is physical capacity. The International Energy Agency (IEA) estimates that data centres used about 415 terawatt-hours (TWh) of electricity in 2024—around 1.5% of global electricity consumption—and projects roughly 945 TWh by 2030 in its base case, with AI the leading growth driver. That is a scenario, not a certainty: the IEA’s estimates for 2035 range from about 700 to 1,700 TWh depending on adoption, efficiency and infrastructure bottlenecks.
How much electricity do AI data centres use?
Published statistics generally cover data centres as a whole rather than isolating AI workloads. They include conventional cloud services, storage and enterprise computing alongside model training and inference. AI is nevertheless the fastest-growing source of demand inside the sector.
| Measure | Value | Qualification |
|---|---|---|
| Global data-centre electricity use | About 415 TWh in 2024 | IEA estimate; approximately 1.5% of worldwide electricity use |
| IEA base-case demand | About 945 TWh in 2030 | Projection; AI is the most important growth driver |
| Accelerated-server electricity growth | About 30% per year | IEA base case; mainly AI-driven servers |
| 2035 scenario range | Roughly 700–1,700 TWh | IEA scenarios reflecting uncertainty in adoption, efficiency and infrastructure |
| Global data-centre investment | About half a trillion US dollars in 2024 | IEA estimate for worldwide investment |
Scale is also local. The IEA says a typical AI-focused data centre can consume as much electricity as 100,000 households. The largest facilities under construction can use about 20 times that amount. Nearly half of US data-centre capacity is concentrated in five regional clusters, so a national percentage can understate the difficulty of connecting a new facility in a particular grid area.
#1 Best Overall
Why does AI depend on the cloud?
Training and inference need concentrated hardware
Large-model training uses thousands of accelerators linked by very fast networking. Serving a popular model (inference) also requires fleets of accelerators that can respond with low latency and remain available around the clock. Concentrating this equipment in a hyperscale facility makes it possible to share storage, networking, monitoring and power systems, but it also creates large, location-bound electrical loads.
Public cloud platforms turn that concentration into a service: developers rent computation instead of building a data centre, buying a cluster of accelerators or operating a private power system. The trade-off is dependence on a small number of platforms and on the supply chains those platforms use. A cloud outage, accelerator shortage, delayed grid connection or regional water restriction can affect many customers at once.
Software demand is arriving faster than energy infrastructure
According to the IEA, a data centre can become operational in two to three years, while generation, transmission and grid-connection projects usually require longer planning and construction periods. The agency estimates that around 20% of planned data-centre projects could face delays if grid risks are not addressed. This timing mismatch explains why a company can have a site, customers and financing yet still lack deliverable power.
What resources do AI data centres consume?
Electricity and firm grid capacity
Energy is the largest operating cost identified by the OECD: it runs the computing equipment and the cooling systems that keep it within safe temperatures. The relevant question is not only how much electricity exists nationally, but whether a site can receive reliable, high-capacity power when its cluster is ready. Transmission availability, connection queues, backup generation and the emissions profile of the local grid all affect the economics and environmental impact.
Water and cooling equipment
High-density accelerator racks produce intense heat. Facilities may use air cooling, chilled-water systems, direct liquid cooling or other designs, with water requirements varying by climate and technology. The OECD cites a French competition-authority study reporting that water-based cooling at OVHcloud and Scaleway can save up to 40% of energy compared with conventional air conditioning. “Up to” is a site-specific potential, not a universal result for every liquid-cooling installation.
Cooling can shift pressure from electricity systems to water systems. A facility in a water-stressed region may face limits even when its power supply is adequate, while a design that minimizes water use may require more electricity or different capital equipment. Microsoft says that, in its fiscal-year 2025 report, it replenished more than 14.2 million cubic metres of water and matched 100% of its annual electricity consumption with renewable energy. Those are the company’s reported global results for that reporting year, not a guarantee that every individual data centre has the same local balance.
Accelerators, networking and minerals
AI facilities require GPUs or other accelerators, high-bandwidth memory, advanced packaging, servers, optical and electrical networking, storage and replacement parts. Their production depends on semiconductor fabs, specialist manufacturing equipment and minerals extracted through global supply chains. Concentration at any one layer can create a bottleneck even when other components are available.
The IMF describes the chain plainly: “Behind every chatbot or image generator lie servers that draw electricity, cooling systems that consume water, chips that rely on fragile supply chains, and minerals dug from the earth.”
Land, buildings and capital
Data centres need suitable land, substations, fibre routes, construction capacity and access to finance. The IEA’s estimate of roughly half a trillion dollars of global data-centre investment in 2024 illustrates the capital intensity before a model generates revenue. Utilisation matters: an expensive cluster that sits idle wastes embodied materials and fixed costs, while a heavily used cluster increases power and cooling demand.
How hyperscalers are becoming infrastructure companies
Cloud providers increasingly make decisions across several industrial layers instead of simply leasing servers. They contract for generation, sign power-purchase agreements, reserve transmission capacity, build or acquire facilities, and redesign cooling for denser racks. Microsoft reports a power-purchase agreement supporting the restart of the Crane Clean Energy Center and says it is developing data-centre cooling that uses less water.
They are also designing custom application-specific integrated circuits (ASICs). Google’s and Amazon’s in-house accelerators are intended to improve performance per watt and reduce reliance on general-purpose GPUs. Custom silicon can lower energy or operating cost for suitable workloads, but it adds design expense, manufacturing commitments and software-porting requirements. It does not remove dependence on semiconductor foundries, advanced packaging, memory or electricity.
Who controls the infrastructure behind big AI?
No single company controls every layer. Dependence is concentrated across linked markets, and the strongest position can differ by workload and region.
Free tools Windows power users keep installed
One-click scans. No signup required.
| Layer | Typical decision-makers | What creates dependence |
|---|---|---|
| Cloud and data-centre capacity | Hyperscalers, colocation operators and specialist data-centre owners | Available racks, contracts, network reach and facility location |
| Accelerators and custom chips | GPU suppliers, ASIC designers, foundries and memory manufacturers | Fabrication capacity, packaging, software ecosystems and performance per watt |
| Electricity | Utilities, independent generators, grid operators and large buyers | Generation, transmission, connection queues and firm-power contracts |
| Cooling and water | Data-centre designers, equipment vendors and local water authorities | Climate, water availability, cooling technology and operating permits |
| Upstream materials | Mining, refining and component supply chains | Geographic concentration, logistics and exposure to disruption |
| Customers and model developers | Businesses, governments and AI labs | Switching costs, application interfaces and access to scarce capacity |
This structure gives hyperscalers significant bargaining power as large buyers, but it does not make them utilities, chip foundries or mineral producers. A disruption at any linked layer can limit the whole system.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How should a new AI facility or workload be evaluated?
Comparing locations or deployment strategies requires more than checking a headline electricity price. Use these axes together:
- Delivered power: firm capacity, connection date, redundancy and the grid’s emissions mix.
- Transmission and queue risk: the lead time for substations and lines, not merely the existence of a nearby plant.
- Water and cooling: local water stress, seasonal restrictions and whether air, liquid or hybrid cooling fits the site.
- Accelerator supply: availability, memory and networking, plus measured performance per watt for the intended models.
- Capital and utilisation: construction cost, financing, expected load factor and the risk of stranded equipment.
- Supply-chain resilience: dependence on one chip, foundry, network component or cloud region.
- Community effects: land use, jobs, electricity prices, water competition, noise and tax revenue.
A site with cheap renewable generation can still be unsuitable if it lacks transmission or water permits. Conversely, a constrained grid can sometimes be managed through phased expansion, flexible workloads, storage or on-site generation, provided those measures meet reliability and environmental rules.
What could change the energy outlook?
The IEA’s wide 2035 range shows why the base case should not be treated as destiny. Several variables can move demand up or down:
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errors- Model and hardware efficiency: better algorithms, lower-precision computation, memory improvements and custom silicon can reduce energy per task.
- Workload design: smaller models, caching, batching and moving non-urgent inference to less-constrained periods can improve utilisation.
- Siting and grid planning: locating facilities where power and transmission can arrive, and coordinating projects with utilities, can reduce delays.
- Cooling innovation: liquid and hybrid systems may lower energy use, but their water, maintenance and capital requirements must be counted locally.
- Policy and market rules: connection standards, reporting, water permits, emissions limits and tariffs influence which projects proceed.
- Adoption: if AI services grow faster than expected, demand rises; if applications consolidate or remain uneconomic, build-out slows.
What “industrialized AI” means in practice
Industrialization means that AI’s limiting inputs are increasingly physical and coordinated: electricity contracts, grid interconnections, cooling systems, semiconductor capacity, buildings, logistics and capital. The IEA calls AI a general-purpose technology and states that “there is no AI without energy.” Countries that can deliver affordable, reliable and sustainable electricity at speed and scale are therefore better positioned to capture AI’s economic benefits.
For users, the implication is straightforward: a model’s capability depends not only on code and data but also on where its servers are, who supplies its accelerators, how its facility is cooled and whether the local grid can support continued growth. For businesses and policymakers, resilience requires treating cloud capacity, power, water and supply chains as one connected infrastructure problem rather than as separate procurement decisions.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
