AI’s Energy Footprint Is Growing Fast. Here’s How Businesses Can Shrink It.

AI’s appetite for power is growing faster than almost anything else in tech. McKinsey projects global data centre capacity demand will rise from 82 GW in 2025 to 219 GW by 2030, with AI workloads making up around 70% of the total. In Europe, McKinsey expects data centre power consumption to nearly triple, from around 62 TWh today to more than 150 TWh by the end of the decade.

That isn’t a reason to stop using AI. It’s a reason to use it well.

The good news is that you don’t need to run a hyperscale data centre to make a difference. Many of the biggest levers sit in decisions businesses make every day: which model to use, when to use AI at all, which supplier to sign with, and what to measure. None of them require a sustainability team. All of them require someone to ask the question.

Why this is a business problem, not just an environmental one

Energy is cost. Every oversized model, every unnecessary prompt and every failed run shows up on an invoice before it shows up in an emissions report.

Efficient AI and affordable AI tend to be the same AI. The choices that cut your footprint usually cut your cloud bill too. They also make your systems faster and easier to scale.

There’s a reporting angle as well. Customers, investors and procurement teams are asking more questions about digital emissions. “We hadn’t thought about it” is getting harder to say with a straight face.

Five things every business can do now

1. Right-size the model

Not every task needs the biggest model. Researchers at the University of Rhode Island benchmarked over 30 commercial models and found the most energy-hungry ones use more than 70 times the energy per long prompt of the most efficient.

Seventy times is a big gap. Most of that difference buys nothing for routine work.

Classifying a support ticket, extracting fields from a form or summarising a short email rarely needs a large reasoning model. Start with the smallest model that meets your quality bar and only step up when the results prove you need to. For mixed workloads, route simple requests to a lighter model and save the heavyweight for the genuinely hard ones.

The mistake we see most: a team picks the flagship model during the pilot because it’s the safest bet, and nobody revisits that choice once the system goes live. The pilot gets a tiny volume of traffic. In production, that default costs a lot more.

2. Don’t default to gen AI

Generative AI is powerful. It’s also expensive to run. Deloitte estimates a gen AI prompt uses 10 to 100 times more electricity than a typical web search.

So before reaching for a large language model, ask a simple question: does this task actually need language understanding?

If a search, a rule or a simple automation does the job, use that. Matching an invoice to a purchase order is a rule. Looking up a policy is a search. Moving a file when a form is submitted is an automation. These are fast, cheap, predictable and easy to audit.

Save gen AI for the work that needs it, such as messy unstructured text, nuanced judgement and open-ended questions. If you’re weighing up which ideas deserve AI at all, our guide on deciding which AI ideas are worth funding is a good place to start.

3. Ask your providers the hard questions

Most businesses never see the data centre their AI runs in. That doesn’t mean they have no influence over it.

According to Deloitte, servers and cooling each account for around 40% of data centre power use. How a provider handles both makes a real difference to the footprint you inherit. Before you sign, ask:

• Efficiency: What is your power usage effectiveness (PUE), and how has it changed over time?
• Cooling: What cooling approach do you use, and how much water does it consume?
• Renewable supply: What share of your power is renewable? Is that matched to when and where it is used, or covered by certificates bought elsewhere?
• Region choice: Can we run our workloads in regions with a cleaner grid?
• Reporting: Will you give us energy and emissions data for our own usage?

Vague answers are an answer in themselves. A provider that takes efficiency seriously will have the numbers ready. We’ve written more about the tools and practices we use in our own approach to sustainable AI development.

4. Fix the data first

This one surprises people. Data quality is a sustainability issue.

Messy data means more retries, more reprocessing and more wasted compute. Duplicate records get processed twice. Inconsistent formats trigger failed runs. Missing context leads to bloated prompts packed with extra information, just in case. Poor outputs get regenerated until someone is happy with them.

Each of those costs energy, and none of them adds value.

Clean foundations are greener foundations. Data that is structured, governed and easy to find lets AI do the job once, properly. It’s the same principle behind great AI needing great data, and it’s why we usually start client work with the data underneath rather than the model on top.

5. Measure from day one

You can’t shrink a footprint you can’t see.

Track usage, cost and energy at the pilot stage, not after launch. Useful starting metrics include:

• Model calls and tokens per completed task
• Cost per outcome, not just cost per month
• Failure and retry rates
• Estimated energy and emissions, using the carbon reporting tools from your cloud provider

A pilot is the cheapest moment to learn this. Set a baseline while volumes are small. Then, when you scale, you’ll know exactly what you’re scaling, and you’ll spot the expensive habits before they become expensive at volume.

The mistakes to avoid

A few patterns come up again and again.

Treating sustainability as a post-launch report. By the time the system is live, the model, the architecture and the supplier are already locked in. That’s where most of the footprint is decided.

Assuming a supplier’s green claims cover you. A renewable energy commitment at company level says little about the power behind your specific workloads. Ask for your numbers.

Optimising the prompt and forgetting the volume. A model that’s 20% more efficient per request doesn’t help much if usage grows tenfold without anyone asking whether every call is needed.

No single company solves this alone

AI’s energy challenge is bigger than any one organisation. Deloitte expects the technology and power industries to work together to contain AI’s energy impact through cleaner supply, better cooling and more efficient chips.

The rest of us have a part to play too:

• Share what works. If a smaller model or a simple rule did the job, tell your peers.
• Set expectations with suppliers. Make efficiency and transparency part of procurement, not a nice-to-have.
• Build efficiency into the brief, not the post-mortem. The cheapest energy saving is the one you design in from the start.

Every efficient choice adds up

Deloitte projects data centres could account for around 4% of global electricity consumption by 2030. That’s a big number made of millions of small decisions, including which model handles a routine query, whether a task needed AI at all, and whether anyone checked.

Sustainable AI isn’t about doing less with AI. It’s about getting the same value, or more, from less energy, less cost and less waste.

At The Virtual Forge, we work with teams to choose the right tool for each job, build the data foundations that stop compute being wasted, and put measurement in place from the first pilot. If you’re shaping your AI strategy and want efficiency built in from day one, let’s talk.