Imagine two places for the same 100 MW AI data center.
Site A offers electricity at $50 per MWh.
But the utility says full power may not arrive for four years.
Site B costs $75 per MWh.
Its grid connection can be ready in 18 months, with room to expand.
Which site is cheaper?
At first, Site A looks obvious.
Then another question appears:
What is cheap electricity worth if the data center cannot turn on?
This is why the next AI bottleneck may not be the chip itself.
A company can have GPUs, land, financing and customers—and still wait for the one input that makes all of them useful: deliverable power.
The GPU is only one component. Useful compute begins after enough electricity reaches the site, passes through power equipment, and can be cooled continuously.
The Answer Direction: AI Is Becoming a Power-Delivery Problem
The chip story still matters.
Advanced accelerators remain expensive, difficult to manufacture and central to AI performance.
But a GPU that cannot be energized produces no useful compute.
The International Energy Agency’s 2026 update projects global data-center electricity use rising from roughly 485 TWh in 2025 to about 950 TWh in 2030. AI-focused data centers grow much faster than the total, roughly tripling their electricity use over the same period.[1]
The physical system must scale with that demand.
Generation has to exist.
Transmission has to carry it.
Substations and transformers have to step it down.
Switchgear and UPS systems have to distribute and protect it.
Cooling has to remove the heat.
The AI boom is therefore becoming an infrastructure problem as much as a semiconductor problem.
The Contexta Grid-to-Chip Chain
Follow one unit of electricity from the grid to useful AI compute.
Regional Grid → Substation → Transformer → Switchgear / UPS → Server Rack → Heat → Cooling
This chain matters because every link has its own capacity, lead time and failure mode.
| Layer | What it must do | What can delay the project |
|---|---|---|
| Grid | Provide enough firm capacity to the area | Congestion, generation shortfall, interconnection queue |
| Substation | Connect the campus to high-voltage infrastructure | Permitting, engineering, construction capacity |
| Transformer | Change voltage to usable levels | Manufacturing lead time and supply-chain constraints |
| Switchgear / UPS | Control, distribute and stabilize power | Equipment availability, commissioning, redundancy design |
| Rack | Turn electricity into compute | Rack power density and power-delivery limits |
| Cooling | Remove nearly all of that electrical energy as heat | Water, pumps, chillers, liquid-cooling equipment, climate limits |
The important point is simple:
The chip sits near the end of the chain.
Why 2026 Makes the Grid Problem Harder to Ignore
The IEA now describes grid capacity as a critical bottleneck for connecting new supply, storage and large loads.
Its 2026 electricity outlook says more than 2,500 GW of projects—including renewables, storage and large-load projects such as data centers—remain stalled in grid-connection queues worldwide.[2]
For data centers specifically, the IEA estimates that grid constraints could delay around 20% of global capacity planned for construction by 2030.[3]
This does not mean one in five announced projects will disappear.
It means connection timing has become a material uncertainty in the buildout.
That is very different from a simple electricity-price problem.
The Two Clocks of an AI Data Center
An AI campus lives on two clocks.
| Clock | What moves on it? | Typical problem |
|---|---|---|
| Compute clock | GPUs, servers, software, model releases, building fit-out | Technology changes before infrastructure is ready |
| Grid clock | Generation, transmission, substations, transformers, permits | Multi-year planning, construction and equipment lead times |
The IEA notes that a data center can be built and become operational in roughly two to three years, while broader energy infrastructure often takes longer. New transmission lines can take four to eight years in advanced economies, and wait times for critical components such as transformers and cables have increased sharply.[4]
The project opens only when both clocks meet.
You can accelerate the building and still lose the race to the substation.
Start With Three Numbers: MW, MWh and PUE
You do not need a full engineering model to understand the first-order economics.
Three quantities are enough to begin.
- MW measures power. It tells us how much electricity the facility needs at a moment in time.
- MWh measures energy. It tells us how much electricity is consumed over time.
- 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 needs 1.30 units of total electricity for each unit used by the IT equipment.
The extra 0.30 supports cooling, power conversion, pumps, lighting and other building systems.
PUE does not tell us whether the AI workload is valuable.
It tells us how much electrical overhead the building adds around the compute.
A 100 MW Example Shows Why Efficiency Matters
Take an illustrative 100 MW AI data center.
| Input | Assumption |
|---|---|
| Maximum IT capacity | 100 MW |
| Average IT load factor | 85% |
| PUE | 1.30 |
| Electricity price | $70/MWh |
| Operating time | 8,760 h/year |
The average IT load is 85 MW.
With a PUE of 1.30, average facility load becomes 110.5 MW.
85 MW × 1.30 = 110.5 MW
Annual electricity use becomes about 967,980 MWh.
At $70/MWh, annual electricity cost is about $67.8 million.
If PUE rises from 1.30 to 1.50, annual electricity cost rises to roughly $78.2 million.
That is a difference of about $10.4 million every year in this simplified example.
The model is illustrative. Real contracts can include demand charges, grid fees, taxes and time-of-use pricing.
But it shows why cooling and facility efficiency belong in the power discussion.
Time-to-Power Can Matter More Than the Cheapest Tariff
Now return to Site A and Site B.
| Factor | Site A | Site B |
|---|---|---|
| Electricity price | $50/MWh | $75/MWh |
| Grid connection | 4 years | 18 months |
| Expansion capacity | Uncertain | Available |
| Reliability | Moderate | High |
There is no universal winner.
A long-lived workload may justify waiting for lower electricity costs.
A company racing to bring scarce AI capacity online may value 30 months of earlier operation much more than the tariff difference.
Grid delay creates costs that do not appear on the electricity bill:
- idle land and buildings
- financing costs
- equipment depreciation before use
- newer chips arriving while older equipment waits
- customers choosing another provider
- revenue starting later
This is why time-to-power is an economic variable, not just an engineering schedule.
A 2026 Project Shows the Risk Is No Longer Theoretical
In September 2026, Reuters reported that Oracle issued a force-majeure notice connected with potential power delays at Project Jupiter, a large AI data-center project in New Mexico. Oracle said the project remained on schedule, but the contractual step showed how power availability can affect rent timing, financing and project obligations even before servers begin operating.[5]
Developers are also looking for ways around long grid schedules.
Reuters reported in late September that demand for smaller behind-the-meter gas turbines is rising as data-center developers seek faster onsite power while waiting for larger grid or generation projects.[6]
This does not mean onsite generation will replace the grid.
It shows how valuable earlier energization has become.
Cooling Turns Electricity Into a Site-Design Problem
Electricity does not disappear after it enters a GPU.
Nearly all of it becomes heat.
As rack density rises, more heat must be removed from less space.
That can change:
- cooling-system design
- water requirements
- pump and fan energy
- heat-exchanger capacity
- maintenance
- PUE
A cooler climate can help some designs.
Direct liquid cooling can move heat efficiently at high rack densities.
But every cooling solution adds equipment, cost and operating constraints.
Power availability without a workable heat-removal system is not usable AI capacity.
Reliability Has a Price Too
An AI data center does not only need electricity.
It may need electricity that is extremely reliable.
That can mean:
- multiple grid feeds
- UPS batteries
- backup generators or onsite generation
- redundant transformers
- spare cooling capacity
- multiple network routes
Each layer improves resilience.
Each also adds capital cost, maintenance and sometimes energy losses.
The right answer depends on the workload.
Some training jobs can pause or migrate.
Other services cannot tolerate interruption.
The Contexta Time-to-Power Test
Before calling a site “AI-ready,” ask six questions.
- How many megawatts are actually deliverable?
Not theoretical regional generation—firm capacity at the site. - On what date is that power committed?
A credible energization date is more useful than a broad queue estimate. - What is the full delivered electricity cost?
Include tariffs, demand charges, grid fees and other material charges. - What PUE can the site realistically achieve?
Climate and cooling design change facility electricity use. - How reliable must the workload be?
Redundancy requirements change both cost and power demand. - Can the site expand?
A 100 MW first phase may eventually want several times more capacity.
Then add one more:
What happens if the connection date slips by a year?
A good site should still make sense under more than one scenario.
Power Access May Decide Which AI Capacity Actually Exists
Lawrence Berkeley National Laboratory’s 2025 Update estimates that U.S. data centers could account for about 11.8% of total U.S. electricity use in 2030, with a scenario range of 9.5% to 15.3%.[7]
The IEA expects data centers to account for roughly half of U.S. electricity-demand growth through 2030.[8]
But demand is geographically concentrated.
A country can have enough generation in aggregate while a specific data-center cluster still lacks transmission, substation capacity or a firm connection date.
That distinction matters.
Enough electricity somewhere is not the same as enough electricity at this site, on this date.
What Should You Watch Next?
- Firm energization dates: When will announced campuses actually receive full power?
- Grid-connection queues: Are utilities reducing the backlog or only adding requests?
- Transformer and substation lead times: Are equipment bottlenecks easing?
- Behind-the-meter generation: How often are developers building temporary or permanent onsite power?
- PUE and rack density: Are efficiency gains keeping pace with higher AI power density?
- Expansion headroom: Can a site add another 100 MW after the first phase?
The Main Idea
Return to Site A and Site B.
Site A has cheaper electricity.
Site B can turn the machines on much sooner.
The better site cannot be chosen from the electricity price alone.
You need the whole physical and economic chain:
Power Capacity → Connection Date → Reliability → PUE → Cooling → Expansion → Useful Compute
Chips determine what a data center can compute.
Power delivery determines whether that compute can exist at all.
The critical date is not when an AI data center is announced. It is when enough reliable electricity reaches the site and useful compute begins.
Continue Reading
- The Hidden Bottleneck of the Electric Age: Why Power Grids Take So Long to Build — Why can grid infrastructure take much longer than the technology it serves?
- The Electric Age Explained: From Falling Energy Costs to AI and Global Power — How does cheaper electricity connect AI, industry and infrastructure?
- Who Will Lead the Electric Age? Power, Grids, Chips, and Intelligent Machines — Which capabilities matter when digital and physical systems scale together?
Key Terms
- grid access: the ability of a site to receive enough electricity from the power network when the project needs it
- time-to-power: the time between choosing or developing a site and receiving the electrical capacity required for operation
- MW: megawatt, a measure of instantaneous power capacity
- MWh: megawatt-hour, a measure of electrical energy used over time
- PUE: Power Usage Effectiveness, the ratio of total facility energy to IT-equipment energy
- power density: the amount of electrical power concentrated in a rack, room or other physical area
- load factor: average power use divided by maximum or rated capacity over a period
- energization: the point when electrical infrastructure is connected and can begin supplying power
- behind-the-meter power: generation or storage located on the customer side of the utility meter and used directly by the site
- redundancy: extra equipment or capacity kept available so service can continue if another component fails
- lead time: the time between ordering, approving or starting infrastructure and having it ready for use
Sources
- IEA — Key Questions on Energy and AI, Executive Summary — 2026 data-center electricity outlook.
- IEA — Electricity 2026, Executive Summary — global grid-connection queues and investment needs.
- IEA — AI and Energy Security — data-center capacity at risk of grid-connection delay.
- IEA — Energy and AI, Executive Summary — data-center construction timelines, transmission and equipment lead times.
- Reuters — Oracle data-center project and potential power delays, Sep. 24, 2026.
- Reuters — Data-center demand for smaller behind-the-meter gas turbines, Sep. 29, 2026.
- Lawrence Berkeley National Laboratory — United States Data Center Energy Usage Report: 2025 Update.
- IEA — Electricity 2026: Demand — U.S. electricity-demand growth and data-center contribution.
- U.S. Department of Energy — Best Practices Guide for Energy-Efficient Data Center Design — PUE definition and facility-efficiency context.
Status checked September 30, 2026. Electricity-demand forecasts, project schedules, grid queues and equipment lead times can change. The Grid-to-Chip Chain, Two Clocks, and Time-to-Power Test are The Contexta analytical frameworks. The 100 MW cost example is illustrative rather than a project quotation. This article explains AI infrastructure and power-system constraints.