Why Do Data Center Power Forecasts Disagree? – How to Read Them

One report implies data centers will need a few dozen more gigawatts. Another points toward a hundred. A utility says its own territory alone could add more than that, and a grid operator's large load queue says something bigger still. For anyone planning a project or pricing a power deal, the spread is maddening, as if every data center power forecast came from a different planet. Why do these forecasts disagree so much, and how do you read one without getting fooled?

Data center power forecasts disagree mainly because they measure different things: different regions, different years, energy in terawatt-hours versus peak demand in gigawatts, and different assumptions about chips, efficiency and how many proposed projects will actually be built. Forecasts built from chip shipments usually come in lower than those built from utility load requests, which include speculative and duplicate projects, and not every forecast models supply limits on turbines, transformers and transmission. To read one without getting fooled, check its scope, units, method and scenario before comparing it to anything else.

What Are the Main Data Center Power Forecasts?

A handful of public forecasts get quoted again and again. They do not share a common format, which is the first clue to why they disagree.

Look at the units column in your head. Two are in terawatt-hours (TWh), one is a percentage and one is in gigawatts (GW). Two cover only data centers; one covers all US load growth, data centers included. None of these can be lined up side by side without some translation.

Why Do Units Make Forecasts Look So Different?

Because energy and power are different quantities. A TWh forecast tells you how much electricity data centers will use over a year. A GW forecast usually tells you how much peak capacity the grid must be ready to serve, or how much load has asked to connect.

You can convert between them, but only with an assumption about how hard the load runs. Here is the math with Berkeley Lab's 2028 range:

• Low case: 325 TWh ÷ 8,760 hours ≈ 37 GW of average load.
• High case: 580 TWh ÷ 8,760 hours ≈ 66 GW of average load.
• Peak load is higher than average. If data centers run at about 80 percent of their peak on average, the high case implies roughly 66 ÷ 0.8 ≈ 83 GW of peak demand.

So a "37 GW" headline and an "83 GW" headline can both come from the same report, depending on which number the writer picked. That alone explains a lot of the noise.

Why Do Bottom-Up and Top-Down Forecasts Disagree?

They start from opposite ends of the problem. Bottom-up forecasts, like Berkeley Lab's, estimate how many servers and accelerators will ship, how much power each draws and how efficient the buildings will be. Top-down forecasts start from what utilities and grid operators see: signed contracts, load studies and interconnection requests.

Utility requests tend to run higher, and there is a reason. Developers often ask several utilities about the same project while they shop for the fastest connection. A single 500 MW campus can appear in three queues at once, and only one will ever be built. Some requests are early placeholders from developers who have land but no tenant.

Texas shows the scale of the effect. The Texas Reliability Entity reported that large load demand under study in ERCOT rose from about 72 GW to more than 200 GW in roughly 11 months during 2025, with data centers and crypto mining making up over 80 percent. No one expects all of that to connect. On the generation side, Berkeley Lab's Queued Up study found that only about 13 percent of capacity requested from 2000 to 2019 ever reached commercial operation. Large loads follow different rules, but the same caution applies.

Why Do Regional Forecasts Run Hotter Than National Ones?

Because regional forecasts are built closer to the request pile, and data centers cluster in a few regions. When a grid operator publishes its load outlook, it usually leans on what utilities and developers have formally asked for, and in hot markets that number grows fast.

Texas is the extreme case. Data Center Dynamics reported in December 2024 that ERCOT's own forecast showed about 125 GW of new load by 2029, including about 82 GW from data centers. A year later, Utility Dive reported roughly 233 GW of large load requests in ERCOT, more than 70 percent of them data centers, against an interconnection process designed to handle 40 to 50 large loads at a time. For scale, ERCOT's record peak demand has stayed under 100 GW.

PJM, the grid operator for much of the mid-Atlantic and Midwest, has seen the same pattern in a more measured way. It has repeatedly raised its peak load forecasts, and it has said most of the recent increase comes from data centers. Those forecast changes flow straight into capacity auction prices.

None of this means regional operators are wrong. Their job is to plan for what might connect, and running short is worse than overbuilding a little. It does mean a regional request number and a national consumption estimate should never be added together or compared directly.

How Much Do Efficiency and Chip Assumptions Matter?

A great deal, because small changes compound. A forecast has to guess how many accelerators will ship, how much power each new generation will draw, how busy those chips will be and how efficient the cooling around them will be. Shift any one of those a few percent per year and the answer five years out moves by tens of terawatt-hours.

The International Energy Agency makes this explicit with scenarios. Its Energy and AI analysis sets out cases for 2035 ranging from about 700 TWh in a "Headwinds" scenario to more than 1,700 TWh in a "Lift-Off" scenario, around a base case of about 945 TWh by 2030. The chart below shows how wide that fan is.

Notice that the low 2035 scenario sits below the 2030 base case. That is not a typo. It reflects a world where efficiency gains, supply constraints or weaker demand slow growth enough that use in 2035 ends up lower than the central path expected for 2030.

Do Supply Limits Cap the Forecasts?

They should, but not every forecast models them. Demand only becomes consumption once someone builds the power plant, the substation and the transmission line to serve it. Turbine makers report multiyear backlogs, large transformers have long lead times, and new transmission can take most of a decade.

A forecast that counts every requested megawatt without asking whether equipment can be delivered will run high. One that assumes today's constraints never ease may run low. The most useful forecasts say plainly which constraints they include.

Business shifts matter too. New kinds of buyers can appear quickly, such as the GPU cloud providers often called a neocloud, which lease capacity on short contracts and can grow or pull back faster than a traditional cloud company. And big buyers adjust: in early 2025, analysts at TD Cowen reported that Microsoft had let some data center lease agreements lapse, even as the company said its overall spending plan was unchanged.

How Should You Read a Data Center Power Forecast?

Run every forecast through the same short checklist before you quote it or use it:

1. Scope: global, national, regional or one utility? Data centers only, or all load?
2. Units: annual energy (TWh), average load (GW) or peak demand (GW)?
3. Method: bottom-up from equipment, or top-down from requests and contracts?
4. Filters: does it discount duplicate or speculative requests?
5. Constraints: does it limit growth by equipment, fuel and transmission?
6. Scenario: is the headline the base case, the high case or a range?
7. Date: when was it published, and how much has changed since?

Once you answer those, most "contradictions" turn out to be different questions with different answers. The grid operator's queue number and the national lab's TWh number can both be right at the same time.

Why Forecasts Still Matter, Even When They Disagree

Forecasts drive real decisions. Utilities use them to justify new plants and transmission to regulators. Grid operators use them to set capacity requirements. Developers use them to pick markets, and investors use them to size bets. When forecasts run high, ratepayers can end up paying for capacity nobody uses. When they run low, the grid comes up short.

Grid Strategies has highlighted how quickly expectations moved: utilities' aggregate five-year load growth forecast reached 166 GW by 2030 in its 2025 report, about six times the 2022 figure. Some of that growth will arrive, and some will fall away. Watching how forecasts change from year to year is often more revealing than any single number.

Reading a Data Center Power Forecast Like a Grid Planner

Most of the disagreement between data center power forecasts is translation, not contradiction. Put two forecasts on the same scope, units and scenario and the gap usually shrinks, though real uncertainty about chips and supply remains.

Keep the checklist above handy. When a new forecast lands, run it through those seven questions, convert it to both TWh and GW, and compare it only with forecasts that cover the same ground. You will spot the headline-grabbers quickly, and you will make better calls about where power is truly tight.

← All guides