AI Data Center Power Demand in 2026

A new data center can be built in roughly one to three years. The major grid work needed to power it can take five to fifteen.

Watch it play outServers arrive years before the grid can power them

A developer picks a site and starts building. The grid is already there; the connection to this site isn't.

A new data center can be built in roughly one to three years.

The major grid work needed to power it can take five to fifteen years. By year five, the line is only part-built.

At the long end, power arrives in year fifteen. That gap, not a shortage of electricity, is the real AI power story.

Replay

That gap is the real AI power story. The world is not simply running out of electricity: the constraint is delivering enough power to a specific site, at the right voltage, by the date the servers arrive. Investors should look for who controls that scarce connection, who pays for the upgrades, and which projects will actually switch on. Further down, we turn a hypothetical 100 MW campus into an annual power bill, step by step.

Check the claimThree claims to test as you read

ClaimThe world is running out of electricity for AI.Tap to checkMythThe constraint is delivering enough power to a specific site, at the right voltage, by the date the servers arrive.Flip backClaimA multi-gigawatt campus announcement is capacity on the way.Tap to checkMythAn announcement is not operating capacity. Energized megawatts are what count.Flip backClaimEvery utility is an automatic AI winner.Tap to checkMythRegulation, contracts and customer concentration decide who keeps the money.Flip back

The short version

• The International Energy Agency (IEA) expects global data-center electricity use to rise from about 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030.
• Gartner separately estimates about 565 TWh in 2026, with AI-optimized servers making up 31% of the total.

Watch it play outThe IEA sees data-center power use nearly doubling by 2030

In 2025 the world's data centers used about 485 terawatt-hours of electricity, the IEA estimates.

By 2030 it expects about 950 terawatt-hours. The outer circle is drawn to scale by area.

That is nearly double, roughly 14% a year. Forecasts differ, so treat this as a scenario.

Replay

• A multi-gigawatt campus announcement is not operating capacity. Energized megawatts are what count.
• Not every utility, power producer or cooling company is an automatic AI winner. Regulation, contracts and customer concentration decide who keeps the money.

How much electricity do AI data centers use?

No global meter tracks AI-only electricity. Operators usually report total data-center demand without splitting out each AI workload.

The forecasts use different methods, but they point the same way: AI servers are raising data-center electricity demand, and dense racks add cooling and power-distribution load on top. Treat the numbers as scenarios, not a precise forecast of utility revenue or equipment shipments.

AI-optimized servers: about 31% of 2026 data-center power

Gartner 2026 estimate565 TWh

• 31%175 TWhAI-optimized servers
• 69%390 TWhEverything else

Why AI servers change the math

Ordinary data centers already use power for processors, storage, networking, cooling and power conversion. AI clusters add three pressures:

1. Higher rack density: more electrical load packed into less floor space.
2. Sustained use: big training and inference jobs can keep chips busy for long stretches. (Inference means running a trained model to answer requests.)
3. More supporting gear: dense systems need extra cooling, power distribution, backup and networking.

So AI affects more than total generation. It changes how much power must reach one site, how reliably, and by when.

The bottleneck is the connection

A country can generate plenty of electricity over a year and still lack room at the exact substation or transmission line where a data center wants to plug in. The IEA's build-time mismatch, one to three years for the data center against five to fifteen for major grid work, creates a queue.

See the IEA's Electricity 2026 grid analysis and its energy and AI outlook.

A typical project moves through six steps:

1. A developer secures land and asks for a grid connection.
2. The utility studies local generation, transmission, substations and reliability.
3. The project may need new transformers, lines, switchgear or generation.
4. Regulators decide who pays.
5. Construction goes ahead only if equipment, permits, financing and power line up.
6. The campus is switched on in phases.

Every campus waits in the same queue

1. Request (Land and a connection ask) to Utility study (Lines, substations, reliability)
2. Utility study (Lines, substations, reliability) to Upgrades (Regulators decide who pays)
3. Upgrades (Regulators decide who pays) to Switch on (In phases)

Four chokepoints inside the power stack

• Generation. Data centers need power around the clock, not just annual renewable-energy credits. The mix can include grid power, gas, nuclear contracts, renewables with storage and on-site ("behind-the-meter") systems. Ask about delivery date, capacity factor (how much of the time a plant actually produces), fuel and price exposure, and whether the contract shifts risk to the customer or to the utility's other ratepayers.
• Transmission and substations. New generation is useless if power can't reach the site. Large transformers, substations, lines and connection studies often set the schedule. Separate equipment backlog backed by funded projects from nonbinding plans that may never get connected.
• On-site electrical systems. Switchgear, uninterruptible power supplies, busways, backup and power-management software become part of the compute bill. Every point of electrical loss becomes heat and shrinks the power left for computing.
• Cooling. Accelerator heat has to leave the rack, and denser systems increasingly use liquid cooling instead of air alone. Cooling interacts with water supply, local climate, building design, chip thermal limits and efficiency.

Four chokepoints between the power plant and the chip

1. GenerationAround-the-clock power, not annual credits
2. Transmission and substationsOften set the schedule
3. On-site electrical systemsEvery loss becomes heat
4. CoolingLiquid replaces air in dense racks

Power stack

Who could benefit, and what could go wrong

Rate base is the pool of investment a regulated utility is allowed to earn a return on. Hyperscalers are the giant cloud platforms, and capex is the money they spend on long-lived assets like data centers. The hyperscaler AI capex tracker shows the spending envelope behind this demand.

Numbers that beat a project announcement

• Contracted vs. energized megawatts. Contracted capacity can be years away; energized capacity can run today. Ask how much of an announced pipeline has a credible delivery schedule.
• Time to power. The cheapest land isn't always the most valuable. A pricier site with an earlier, reliable connection earns revenue sooner.
• Power-usage effectiveness (PUE). Total facility energy divided by the energy reaching computing equipment. Lower generally means less overhead, but PUE says nothing about whether the computing is useful or well utilized.
• Customer concentration. A supplier can grow extraordinarily while depending on a handful of hyperscalers. Check how much revenue, backlog and receivables tie back to the same buyers.
• Capital recovery. A new substation or transmission line can be paid for by the data-center customer, utility shareholders, government or other ratepayers. Regulation can change project economics materially.

Energized megawatts beat announced ones

The announcement

• Multi-gigawatt headline
• Contracted capacity years away
• Nonbinding plans that may never connect

vs

The evidence

• Energized megawatts running today
• A credible delivery schedule
• An earlier, reliable connection
• Revenue tied to more than a few buyers

Member research: Capital recovery is the crux: someone has to finance the substation, the campus and the equipment. Issue 02 follows how the AI boom is being financed and which bottlenecks keep their pricing power.

Worked example: from forecasts to a power bill

Start with the forecasts. What do they mean per year?

A compound annual growth rate (CAGR) is the steady yearly rate that gets you from the start figure to the end figure. The two forecasts use different methods, so read each row on its own.

Now shrink it to one site. Hypothetical example: a campus with 100 MW of computing equipment, a PUE of 1.25, running at an average 75% of capacity, paying $60 per megawatt-hour.

Make a guess

A campus has 100 MW of computing equipment, a PUE of 1.25, runs at 75% on average and pays $60 per MWh. What's its yearly electricity bill?

Tap your guess

$0–25MYour guess$25–50MYour guess$50–75MYour guessOver $75MYour guess

$01234567890123456789.0123456789MActual

Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was $49.3M.

100 MW × 1.25 = 125 MW, × 75% = 93.75 MW of average draw, × 8,760 hours × $60.

Two lessons:

• Overhead is real money. Moving from a PUE of 1.25 to 1.5 adds about $9.9 million a year here, with no extra computing.
• Delay is lost revenue for someone. Until the site is energized, none of that $49 million flows to a utility or power producer. That is why energized megawatts matter more than announced ones.

Overhead and price both move the bill

Pick a PUE and an electricity price

PUE

1.251.5

Electricity price

$60/MWh$120/MWh

Yearly electricity bill$49.3M$98.6M$59.1M$118.3M

$0.0M+$49.3M+$9.8M+$69.0Min line with Base caseabove Base caseabove Base caseabove Base case

Base case $49.3M

$49.3M$118.3M

What could shrink the forecast

Consumption could come in below expectations if:

• Model designs need less computation
• Quantization (storing numbers with fewer bits) and specialized chips cut energy per token
• Utilization stays below plan
• Local opposition or connection limits delay projects
• AI demand grows more slowly
• Workloads move to regions with spare capacity

Efficiency cuts the energy for one unit of work, yet total demand can still rise if far more units get used. Track both energy per task and total usage.

Efficiency can still raise total demand

1. Energy per task falls (Better chips, quantization) to Far more tasks get used
2. Far more tasks get used to Total demand can still rise

Try it yourself

A 10-minute exercise with the free tools:

1. Open the data center PUE and electricity cost calculator. Enter 100 MW, a PUE of 1.25, a 75% load factor and $60 per MWh, and check you get roughly the $49 million above.
2. Raise the PUE to 1.5, then double the electricity price. Note which change moves the bill more.
3. Open the AI power exposure model. Enter a campus size, then lower the completion probability and add a delay. Watch how probability-adjusted load and supplier exposure shrink.
4. Read the IEA's grid analysis and list the constraints it names for connecting new load.

You should now be able to answer: for an announced campus, how many megawatts are likely to be energized, when, and who pays for the connection?

Our read

The strongest signal is not a giant gigawatt announcement. It is a funded project with land, permits, equipment, a connection date, a creditworthy customer and economics that survive regulatory review.

What we'd listen for in the next earnings cycle:

We would see the bottleneck easing if energized megawatts start catching up with contracted megawatts and connection timelines shorten. If sites switch on and utilization stays below plan, the problem shifts from too little power to too little demand. For how power fits alongside memory, packaging and networking, see the AI infrastructure bottleneck tracker and the AI infrastructure investing map.

Questions readers ask

Will AI data centers raise household electricity bills?

It depends on who pays for new substations, lines and generation. Regulators decide how those costs are split between data-center customers, utility shareholders, government and other ratepayers. That is also why regulated utilities may face limits on passing costs to households.

Why don't data centers just build their own power plants?

Some do use on-site, behind-the-meter systems such as gas generation. Those projects still need fuel, permits, equipment and financing, and generation is only one of the four chokepoints. The power still has to be converted, distributed and cooled on site.

What counts as a good PUE?

Lower is better, because it means less power spent on overhead. A PUE of 1.0 would mean every watt reaches the computers. In the worked example, 1.5 instead of 1.25 cost about $9.9 million more a year, but PUE says nothing about whether the computing is useful.

If AI gets more efficient, won't power demand fall?

Not necessarily. Efficiency cuts the energy for each unit of work, but total demand can still rise if people use far more of it. Track energy per task and total usage side by side.

Go deeper with member research

Power demand becomes an investment case only when a company can deliver it profitably. Two member case studies test exactly that:

• GE Vernova and the value of delivering dependable power: Gas-power and grid economics as of January 29, 2026, and why it was already a recognized trade by then.
• Comfort Systems and the scarce capacity behind AI construction: Whether the backlog and cash conversion looked investable on February 20, 2026, and what the later gains really rewarded.

Limits of this data

• Electricity forecasts are scenarios. The IEA and Gartner figures use different methods and shouldn't be combined.
• Project economics vary by region, regulation, contract and technology.