AI data centers are often described as buildings filled with advanced chips. That is true, but incomplete.
Before a GPU can train a model, electricity must cross a regional grid. It then passes through a substation, transformers, backup systems, and power distribution equipment. Finally, it reaches the server rack and leaves the building as heat.
The chip sits near the end of this long physical chain.
The best data-center site is not simply where electricity is cheapest. It is where enough reliable power can arrive on time, remain affordable, and be cooled efficiently.
AI Is Becoming a Large Industrial Load
Data centers already use large amounts of electricity. AI is now pushing both their total demand and their power density higher.
The International Energy Agency estimates that global data-center electricity use reached about 485 terawatt-hours in 2025. Its central projection rises to about 950 terawatt-hours in 2030. Electricity use from AI-focused data centers could triple over the same period. The IEA explains the updated outlook here.
The change is also happening inside each building. The IEA reports that the power density of AI servers increased elevenfold between 2020 and 2025 and could rise another fourfold by 2027.
From the grid's point of view, a large AI campus is not another office building. It is a concentrated industrial load that may need new substations, transformers, transmission capacity, and cooling infrastructure.
Follow the Electricity from the Grid to the Chip
The system becomes easier to understand when we follow the electricity.
Regional grid → substation → transformer → switchgear and UPS → server rack → heat → cooling system
The grid delivers high-voltage electricity to the site. Transformers change the voltage. Switchgear controls the flow. An uninterruptible power supply, or UPS, protects equipment from short interruptions and unstable power.
The electricity then reaches servers, storage, and network equipment. Almost all of that energy becomes heat. Cooling systems must remove the heat continuously. The building therefore needs more electricity than the servers alone consume.
This is why a data center is not just a collection of chips. It is a connected system of computing, electricity, cooling, and backup equipment.
The wider grid may move more slowly than the data-center project itself. The earlier Contexta article The Hidden Bottleneck of the Electric Age explains why transmission lines, transformers, permits, and grid connections often take years to deliver.
Start with MW, MWh, and PUE
Three terms are enough to build a first cost model.
- Megawatts, or MW, measure power. They show how much electricity a facility needs at one moment.
- Megawatt-hours, or MWh, measure energy. They show how much electricity the facility uses over time.
- Power Usage Effectiveness, or PUE, measures facility overhead. It compares total facility energy with the energy used by IT equipment.
PUE = Total Facility Energy ÷ IT Equipment Energy
A PUE of 1.30 means the facility uses 1.30 units of electricity for every unit used by IT equipment. A value closer to 1.0 means less electricity is being used by cooling, pumps, power conversion, lighting, and other support systems.
The U.S. Department of Energy uses the same definition in its Best Practices Guide for Energy-Efficient Data Center Design .
PUE does not measure the value of the computing work. It also does not include land, chips, financing, or grid construction. It answers a narrower question: how much extra electricity does the building need to support its IT equipment?
A Simple 100 MW Example
Consider an illustrative AI data center with the following assumptions.
| Input | Assumption |
|---|---|
| Maximum IT capacity | 100 MW |
| Average IT load factor | 85% |
| PUE | 1.30 |
| Electricity price | $70/MWh |
| Operating time | 8,760 hours/year |
The average IT load is 85 MW. After multiplying by a PUE of 1.30, the average facility load becomes 110.5 MW.
Average IT load = 100 MW × 0.85 = 85 MW
Average facility load = 85 MW × 1.30 = 110.5 MW
Annual electricity use = 110.5 MW × 8,760 h = 967,980 MWh
Annual electricity cost = 967,980 MWh × $70/MWh = $67.8 million
This is an illustrative model, not a project quotation. Real contracts may add demand charges, grid fees, taxes, time-of-use prices, and other costs.
Still, the example shows the scale of the decision. If PUE rises from 1.30 to 1.50, estimated annual electricity cost rises from about $67.8 million to $78.2 million. The difference is roughly $10.4 million every year.
Cheap Electricity Can Still Be Expensive
A low electricity price is attractive. It is not enough by itself.
A site may need a new substation, transmission upgrade, or large transformer. The IEA notes that a data center can become operational in two to three years, while the wider energy system often needs longer planning and construction periods.
| Clock | What Moves on This Clock? |
|---|---|
| Technology clock | Chips, servers, software, and buildings can change quickly. |
| Infrastructure clock | Generation, substations, transformers, and grids usually move more slowly. |
A data center cannot open until both clocks reach the same point. The cheapest site on a spreadsheet may therefore become expensive in practice.
Grid Delay Is a Business Cost
Waiting for power does not appear on the electricity bill, but it still affects the project.
- land and buildings remain unused
- borrowed capital continues to earn interest
- purchased equipment loses value
- newer chips enter the market
- customers move to other providers
- expected revenue begins later
Compare two possible locations.
| Factor | Site A | Site B |
|---|---|---|
| Electricity price | $50/MWh | $75/MWh |
| Grid connection | 4 years | 18 months |
| Expansion capacity | Uncertain | Available |
| Reliability | Moderate | High |
Site A offers cheaper electricity. Site B offers faster and more certain access to power. There is no universal winner.
A long, stable workload may justify waiting for lower operating costs. A company racing to sell scarce AI capacity may value an earlier opening much more.
The right question is not, "Which site has the lowest electricity price?" It is, "Which site produces the best total outcome after power cost, connection time, reliability, cooling, financing, and expansion are included?"
Cooling and Reliability Change the Answer
Electricity enters a server rack as useful power and leaves as heat. As rack density rises, more heat must leave a smaller space.
A cooler climate may reduce cooling energy during part of the year. A hot or humid location may need more mechanical cooling. Water availability can change the design. Direct liquid cooling can move heat efficiently, but it adds pumps, pipes, heat exchangers, controls, and maintenance.
Reliability also has a price. Operators may use two grid connections, UPS batteries, backup generators, redundant transformers, spare cooling equipment, and multiple network links. Each layer improves resilience, but each layer adds capital cost, maintenance, and energy losses.
The correct level depends on the workload. Some AI jobs can pause or move. Other services cannot. Flexibility changes the value of reliability.
Five Questions to Ask Before Choosing a Site
| Question | Why It Matters |
|---|---|
| How many megawatts are available? | The grid must support both IT load and facility overhead. |
| On what date is power guaranteed? | A firm commitment is more valuable than a broad estimate. |
| What is the full tariff? | Energy prices may exclude demand charges and grid fees. |
| What PUE can the design achieve here? | Climate and cooling design change annual electricity use. |
| Can the site expand later? | The first phase may be smaller than the final campus. |
A sixth question should follow: what happens when one assumption is wrong? Electricity prices may rise. The connection date may slip. Utilization may be lower than expected. Cooling performance may miss its target.
A good site should remain acceptable under more than one scenario.
Power Access May Decide the AI Winners
The AI race begins with chips, but it does not end there.
Advanced GPUs need power electronics, transformers, cooling systems, backup equipment, land, and grid capacity. These physical systems cannot scale at the same speed as software.
Lawrence Berkeley National Laboratory estimates that U.S. data centers could account for 11.8% of national electricity use in 2030. Its scenario range runs from 9.5% to 15.3%. The 2025 update explains the assumptions and range.
Demand will not spread evenly. Data centers tend to gather in a few regions, which puts intense pressure on local grids. The companies and regions that connect chips, available electricity, efficient facilities, and fast grid development may gain an advantage.
Advanced chips + available electricity + efficient facilities + fast grid development
Cheap generation helps. A strong grid makes it usable.
What to Watch Next
- Firm grid-connection dates for large projects
- Transformer and substation delivery times
- Contracted electricity prices, not only wholesale prices
- Expected and measured PUE
- The gap between project announcements and actual energization
The critical date is not when a company announces a data center. It is when enough electricity reaches the site and useful computing begins.
Conclusion
An AI data center is a physical chain. Electricity moves from the grid to a substation, through power equipment, into server racks, and out again as heat. Every part of the chain affects cost and timing.
Chips determine how much computing a facility can perform. Grid access determines whether the facility can exist. Electricity prices and PUE shape operating cost. Lead time determines when revenue can begin.
The future of AI will therefore depend on more than semiconductor supply. It will also depend on where power is available, how efficiently it is used, and how quickly the grid can deliver it.
Key Vocabulary & Phrases
Grid access
The ability to receive enough electricity from the power network. Grid access can determine whether a data-center project starts on time.
Power density
The amount of electrical power used within a given space. AI servers are increasing power density inside each rack.
Power Usage Effectiveness, or PUE
A measure that compares total facility energy with IT equipment energy. A lower PUE reduces the electricity used by cooling and support systems.
Lead time
The time required before equipment or infrastructure becomes available. Transformer lead time can delay a new grid connection.
Redundancy
Extra equipment or capacity kept available in case another system fails. Redundancy improves reliability but increases project cost.
Next in This Series
- The Hidden Bottleneck of the Electric Age: Why Power Grids Take So Long to Build
- The Electric Age Explained: From Falling Energy Costs to AI and Global Power
- Who Will Lead the Electric Age? Power, Grids, Chips, and Intelligent Machines
References
- International Energy Agency, Key Questions on Energy and AI , 2026.
- International Energy Agency, Energy Demand from AI , 2025.
- Lawrence Berkeley National Laboratory, United States Data Center Energy Usage Report: 2025 Update , 2026.
- U.S. Department of Energy, Best Practices Guide for Energy-Efficient Data Center Design .
- Bloomberg Originals, How the Grid Is Being Rebuilt for AI .
Published: July 2026 · Data verified through: July 2026 · Update trigger: a major IEA data-center outlook revision, new U.S. electricity-use estimate, or significant change in grid-connection lead times.