The power number is solid. The emissions number isn’t in it.
US data centers used 4.4% of the country’s electricity in 2023 and are headed for somewhere between 6.7% and 12% by 2028. What that does to the atmosphere depends on something the figure doesn’t contain.
11 min read · Answers the first two of the seven questions
The number that holds up.
Of everything in this debate, the electricity demand figure is the most solid — federal, methodologically transparent, and conservative in its framing. It is also a range, and the range is the finding rather than a hedge.
of all US electricity used by data centers in 2023 — up from 58 TWh in 2014 to 176 TWh in 2023
LBNL / DOE (Dec 2024)
projected share of US electricity by 2028, or 325–580 TWh. The spread is nearly a factor of two
LBNL / DOE (Dec 2024)
projected share of world electricity from data centres by 2030, roughly doubling from ~485 to ~950 TWh
IEA, Energy and AI (2025)
share of the growth in data centre demand met by new renewables, which are growing about 22% a year
IEA, Energy and AI (2025)
Read the badges. Every figure above is a National total or Global aggregate. Not one of them is a Grid regionclaim — which is exactly why none of them answers the question most people actually have, which is what the facility near them burns. LBNL’s own limitations section says the report says nothing about where the load lands. That is not a weakness in the research. It is a different question.
One discipline this page holds to throughout: the projection is 325 to 580 TWh, and it is cited that way every time. Rounding the high scenario up and quoting it alone is the most common way this figure gets misused, and the spread exists because the answer depends on hardware efficiency, utilization, and growth in the rest of the economy — not because anyone is being evasive.
Electricity use is not emissions.
This is the guide's whole reason for fusing two questions onto one page. Separating them would let a reader finish the first and assume they had answered the second.
A kilowatt-hour is a quantity of energy. It is not a quantity of carbon. What a kilowatt-hour emits depends entirely on which generating plants actually ran to serve it — at what hour, in which balancing authority, against what else was on the system.
Two identical facilities on two different grids have different carbon footprints and identical electricity bills.
That is why the federal demand report, excellent as it is, cannot resolve the emissions question, and why anyone who moves directly from “data centers used 4.4% of US electricity” to a claim about the atmosphere has skipped the only step that mattered. The demand figure is an input to the carbon question, not an answer to it.
The renewables half of the story.
Leaving this out would be the same selective citation this series teaches against, just pointed in the more comfortable direction.
Renewables are the fastest-growing supply source for data centres, growing roughly 22% a year through 2030 and meeting nearly half of the growth in data centre demand. The supply side is responding, at scale, and any account of this topic that presents demand growth without it is telling half a story.
That finding settles less than it appears to. It is a statement about growth, globally, in aggregate — not about what serves a particular facility, in a particular hour, on a particular grid.
What nobody has settled.
“Half of growth met by new renewables” leaves three questions open, and they are the three that determine the actual emissions consequence:
- What clears at the margin. When new load appears at 7pm in July, the plant that ramps up to serve it is usually not the one that was added to the portfolio that year.
- Whether fossil retirements get delayed. A coal plant that stays open five years longer than planned to serve new load is an emissions event that appears nowhere in a demand statistic.
- Local grid strain, and who funds the upgrade. A transmission build is both an emissions question and a cost question, and Guide 4 takes up the second half.
CPAI does not assert the marginal-emissions question in either direction.
We have not done that research, and the honest position is that the question is answerable and largely unanswered. Anyone telling you confidently that AI’s electricity growth is or is not driving emissions at the margin is ahead of the evidence — in whichever direction they are arguing.
Working on this? We’d like to hear from you →One more caution on the projections themselves. LBNL published in December 2024, before the 2025–26 capital expenditure surge — hyperscaler capex passed $400 billion in 2025 with a further sharp rise expected. The report may prove conservative on the upside, or wrong about efficiency in either direction. Projections are scenarios, not forecasts, and a scenario published before a surge is a scenario that has not seen it.
A grid question and a planet question.
This is the cleanest worked example of the framework in the whole series: one activity, two questions, two scales, and no single number that answers both.
“AI's electricity use is out of control.”
The version that goes too far
Quotes the top of a two-scenario range as though it were a measurement, treats a projection published in 2024 as a current fact, and skips the supply-side response entirely — including the finding that renewables are meeting nearly half of demand growth.
The version that waves it away
Answers with “it's only about 3% of global electricity.” A real figure, at the wrong scale, for someone whose county just received a multi-gigawatt interconnection queue. A global fraction is not a response to a local load.
What the evidence supports
Demand more than tripled between 2014 and 2023 and is projected to roughly double or triple again by 2028, to 325–580 TWh. The supply side is responding faster than most coverage admits. And the emissions consequence at the margin is unresolved — which is a different thing from being small.
Sources for this split: lbnl2024 · ieaEnergyAI — full citations below.
Notice that neither of the first two panels contains a Grid region claim. Both argue about the atmosphere using national and global aggregates, which is why they can both be factually correct and still talk past each other indefinitely.
Action for every level of influence.
For yourself
- Find out which balancing authority serves you. It is public, most people have never heard the term, and it is the unit that actually determines what burns when you turn something on.
- Look up your grid's fuel mix on the EIA's public dashboards. The same AI query has a different carbon footprint in different places — and that is a fact about your grid, not about the model.
- Notice when an article gives you a national or global percentage and then draws a local conclusion from it. That move is the single most common error in this topic.
For an organization
- If you procure cloud or AI services, ask the vendor for region-level carbon intensity rather than a corporate annual average. The average is a portfolio statement; the region is where your workload actually ran.
- Ask what clears at the margin, not what the portfolio says. A company can be fully renewable on paper on an annual basis while its incremental load is served by whatever plant ramps up at 7pm in July.
- Distinguish procurement claims from physics. Both matter, and conflating them is how a sustainability report stops being informative.
For a community
- When a facility is proposed, the question is not "how much power." It is "what generation is being built to serve it, and on what timeline" — and the answer lives in your utility commission's docket, not the press release.
- Ask whether any planned fossil retirement is being delayed to serve new load. That is the question with the largest emissions consequence and it is rarely asked in public.
For policy
- Require marginal-emissions accounting rather than annual averages in large-load proceedings. An annual average cannot answer what an incremental megawatt does.
- Require that new-generation commitments tied to a large customer be disclosed in the docket, so the supply-side promise is auditable against what gets built.
- Fund the research. The marginal question is answerable and largely unanswered, and public proceedings are currently deciding billions of dollars without it.
Related
What Are We Arguing About?
The seven separate questions hiding inside “AI and the environment” — and the geographic scale at which each is true. Start here.
The Water Question
The most-quoted statistic in this debate has been retired by its own authors. What replaced it is more useful — and it points at your watershed, not the planet.
Who Pays
The health and cost burden of the AI buildout lands on specific counties and specific ratepayers — and the algorithms optimizing for aggregate efficiency make that worse, not better.
Where this leads
Reading is one thing. Practicing it is another.
The Applied AI Certification builds practical AI fluency across all six domains — the working competence that advances toward proficiency, with structured practice, feedback, and a cohort on the same problems.
Research & further reading.
Want CPAI to teach this in your community?
We deliver this material as workshops and sessions for schools, libraries, local government, and community organizations.