AI uses more energy than you think — and it's costing businesses
New analysis shows AI's real energy and compute footprint is larger and harder to pin down than the tiny per-query figures suggest — and inference (everyday use), not training, is most of it. For a business, wasted compute is wasted money. The fix isn't to avoid AI; it's curated AI — right-sizing models, prompts and deployments to the job. That cuts cost and energy at once, and it is exactly the gap most SMEs can't close alone.
What the numbers actually show
In a May 2025 investigation, MIT Technology Review tried to measure AI's true energy footprint and found the common understanding is full of holes. A few points stand out for businesses:
- A single chatbot reply is genuinely small — on the order of a fraction of a watt-hour — but usage adds up fast as AI is built into everything.
- Inference — everyday use — is roughly 80–90% of AI's compute energy, not the one-off training. How you run AI matters more than how it was built.
- Energy per task varies enormously. A large model can use many times more energy than a small one for the same job, and complex prompts can cost several times more than simple ones.
- Reasoning and agentic models can use roughly an order of magnitude more energy for a simple task than a standard model would.
- Images, and especially video, cost far more than text — a single short AI video can use hundreds of times the energy of a written answer.
- Data centres already draw a meaningful slice of the grid, and that share is projected to climb steeply as AI scales.
The researchers' blunt conclusion was that AI companies do not disclose enough to know the real totals. But the direction is clear: the footprint is growing, and most of it lives in how AI is used day to day.
Small per query, big in aggregate
One prompt is trivial. The problem is that AI is being wired into search, apps, customer service and agents that run continuously — so the trivial cost is multiplied across millions of calls. For a business running AI across its workflows, small inefficiencies don't stay small; they compound on every invoice.
Wasted compute is wasted money
Cloud AI is billed by compute and tokens. Over-provisioning burns budget directly: using a giant model for a job a small one handles, verbose prompts and outputs, unnecessary reasoning steps, re-running because the first answer was wrong, and no caching of repeated work. Energy waste and cost waste are the same waste seen from two angles. Cut one and you cut the other.
Why SMEs are the most exposed
Large companies have platform teams to right-size and monitor their AI. Most small and mid-sized businesses don't — so they default to the biggest, best-known model, wire up tools without measuring outcomes, and adopt every AI feature on offer. The result is paying premium compute rates for mediocre results. The wastage a big firm can absorb is exactly the wastage an SME cannot.
The fix: curated AI
Curation is the whole cure8ai thesis, and it applies directly here — the few things that matter, right-sized:
- Right-size the model. The smallest model that clears your quality bar, not the flashiest one.
- Right-size the prompt and output. Concise, structured instructions instead of sprawling ones.
- Use reasoning and agents only where they earn it — not as the default for simple tasks.
- Ground it in your data. Retrieval keeps answers accurate first time, so you pay for fewer retries.
- Measure against outcomes. Keep what moves the number; cut what is just novelty.
- Automate the genuinely repetitive — and resist AI-ifying everything else.
Done well, this cuts compute cost and energy at the same time. It is efficiency by design, not an afterthought.
How cure8ai helps
cure8 Labs deploys practical AI that is right-sized and grounded in your data, measured against outcomes — not the biggest model wired up and hoped for the best. That is cheaper to run and better at the job. Put rough numbers on your own wasted time with the AI Time-Reclaim Calculator, or start a project.
FAQ
Does using AI harm the environment?
A single query is small; the impact shows up in aggregate and in how AI is deployed. Efficient, curated usage — right-sized models and prompts — reduces both the energy used and the money spent.
How can a small business cut its AI costs?
Right-size the model, tighten prompts and outputs, ground the AI in your own data to avoid retries, use reasoning only where it pays, measure ROI, and automate only the genuinely repetitive. Or get help deploying it efficiently.
Is a bigger AI model always better?
No. A smaller model often handles the task at a fraction of the cost and energy. The skill is matching the model to the job rather than defaulting to the largest one.
Isn't this just about being green?
It is both. Efficiency saves money and energy at once. For most SMEs the cost argument lands first — the sustainability benefit comes along for free.