How much energy does AI actually use?
- Data centers consumed roughly 1.5% of global electricity in 2024 (~415 TWh), and the IEA projects that more than doubling by 2030, with AI the main driver.
- Inference, not training, is now 80–90% of AI compute: the everyday queries, not the one-time model builds (MIT Technology Review, 2025).
- Google's emissions rose ~48% from 2019 to 2023 and Microsoft's ~29% from 2020, both citing AI datacenter buildout in their own reports.
- A local model on an Apple Silicon Mac runs everyday queries on tens of watts, with no datacenter overhead attached.
Every ChatGPT answer is generated in a building full of GPUs that draws power like a small town. One query is trivial. Billions a day are not, and the bill is now showing up in national grid statistics and Big Tech's own emissions reports. Here are the numbers with sources attached, and the part of the story that gets less coverage: most everyday AI work doesn't need a datacenter at all.
The numbers, with their sources
| Claim | Source |
|---|---|
| Data centers used ~415 TWh in 2024, ~1.5% of global electricity; projected to more than double by 2030 with AI as the primary driver | International Energy Agency, Energy and AI (2025) |
| Inference is now 80–90% of AI compute: the daily queries, not training runs | MIT Technology Review (2025) |
| Google greenhouse emissions up ~48% vs 2019; Microsoft up ~29% vs 2020, attributed in their own reporting to AI infrastructure | Google and Microsoft sustainability reports (2024) |
Note what the 80–90% figure means: the energy story isn't the famous training runs. It's the everyday traffic. Which is exactly the part that could run somewhere else.
Why a cloud query costs more than its math
The GPU computing your answer is only part of the draw. Datacenters carry cooling, power conversion, and networking overhead on top of every chip, plus the idle capacity providers keep warm for peak load, plus the transmission to get your prompt there and back. None of that is waste exactly. It's what serving a billion people from centralized buildings requires. But it all rides along on every "summarize this email."
The on-device version of the same query
An Apple Silicon Mac runs local inference inside a power envelope measured in tens of watts; an M1 Ultra under full generation load draws around 100W, a laptop far less. No cooling plant, no conversion losses, no idle fleet, no round-trip. The chip you already own does the work at your desk.
The honest caveat, and it matters: this is not a claim that every local query beats every cloud query. Hyperscale datacenters run newer chips, sometimes on cleaner grids than your wall socket, and a small efficient cloud model can undercut a big local one. Hardware, model size, and your electricity mix all swing the comparison. The defensible claim is directional: everyday inference doesn't require datacenter infrastructure, and moving a share of it to devices people already own removes that overhead entirely.
What this means if you care about the footprint
The leverage isn't guilt about individual prompts; one query is a rounding error. The leverage is architectural: inference is 80–90% of the load, and a meaningful slice of inference is routine work (drafting, summarizing, everyday coding) that mid-size local models handle well. Outlier's bet is that good-enough-and-owned beats rented-and-datacentered for exactly that slice. Use the cloud for what genuinely needs it. Stop shipping your email summaries to a building with a cooling tower.
Frequently asked questions
How much electricity does one AI query use?
Estimates vary widely by model and provider, from fractions of a watt-hour to several. The individual query is small; the aggregate is what moves grid statistics, because inference now makes up 80 to 90 percent of AI compute and runs billions of times daily.
Is training or inference the bigger energy cost?
Inference, now. Training runs are huge one-time costs, but serving queries day after day has overtaken them: MIT Technology Review put inference at 80-90% of AI compute in 2025. That's why the growth projections track usage, not new model announcements.
Does running AI locally actually use less energy?
Often for everyday tasks, because there's no datacenter overhead, idle capacity, or transmission attached, and Apple Silicon is efficient. Not always: grid mix, hardware, and model size matter, and we'd rather say that plainly than print a green badge.
Try Outlier free
Free Nano + Lite — local, private, no account. Pro $20/mo or $149/yr adds everything (all 7 model tiers incl. Plus 397B). Lifetime Pro from $99 (Founding 200, first 200 seats) or $200 (Founders 500). Apple Silicon only.
Download for Mac