ProzessX
Academy ↗DEBook a free consultation
Deep Dive · Green AI

Green AI: the most sustainable AI is also the cheapest.

Between doomsaying and greenwashing lies the uncomfortable truth about AI's energy appetite, and a surprisingly good piece of news for SMEs.

30 minutes, online, free and non-binding.

In short

AI's energy demand depends above all on which model runs for which task: simple text requests need little power, reasoning models a multiple of that. Because ongoing use (inference) accounts for most of it, right-sized models, efficient infrastructure and renewable power cut cost and CO₂ at the same time.

Deep Dive

Energy, cost and choosing the right model

"A single ChatGPT query uses as much electricity as ten Google searches." "A short email makes AI evaporate half a bottle of water." Claims like these have been making the rounds for months, on LinkedIn, on talk shows, in boardrooms. They're catchy, alarming, and, as stated, simply wrong.

On the other side stand the downplayers: "It's not that bad, data centers only account for one and a half percent of electricity use." That's also too short-sighted, because the consumption of AI data centers is growing fast.

The truth lies in between, and it's more interesting for SMEs than either headline: the lever that makes AI more sustainable is the same one that makes it cheaper and more independent. Efficiency, cost and CO₂ all point in the same direction. No alarmism, no greenwashing.

The numbers that hold up, and the ones that don't

According to the International Energy Agency's report ("Energy and AI", 2025), data centers worldwide consumed around 415 terawatt-hours of electricity in 2024, about 1.5 percent of global electricity use. In Germany, demand was around 20 billion kilowatt-hours according to Bitkom and the Borderstep Institute. That's significant, but not the end of the world.

And a single request? The often-cited "3 watt-hours per ChatGPT prompt" comes from a rough 2023 estimate now considered substantially inflated. Google disclosed operational figures in August 2025: an average text request to Gemini uses 0.24 watt-hours of electricity, 0.03 grams of CO₂, and 0.26 milliliters of water. The research group Epoch AI arrives at a similar figure of around 0.3 watt-hours for a typical GPT-4o model.

The "half a water bottle per email" figure rests on a different scope: it includes the entire life cycle, including power generation, and an especially long output. The takeaway: without stating exactly what's being counted, any AI consumption figure is worthless.

Why that's still no reason to relax

These small figures apply to simple text requests. As soon as a model "reasons", plans multiple steps as an agent, or generates images and video, consumption rises sharply. A single answer from a reasoning model can use a multiple of a simple text request.

It's not whether you use AI that determines your footprint, but which AI you use for which task. Model choice is the single biggest lever.

It's not whether we use AI that decides the footprint, but which model we choose for which task.

The real battleground is called inference

Public debate is almost entirely about training AI models. For your company, that's the wrong place to focus. A model is trained once, but then queried millions of times afterward. This ongoing use is called inference, and over the life cycle it accounts for a large share of total energy consumption.

A useful image: training is buying the car, inference is every single drive. The footprint is created by driving, day after day. The good news: inference is exactly the part you can influence. A smaller model, a more efficient serving technique, clever caching all cut energy consumption, CO₂ emissions and your bill at the same time.

Three levers for efficient AI

1. Right-sizing instead of maximizing

For classification, extraction, or answering questions from company knowledge, you don't need a frontier model. A small, specialized model is significantly cheaper to run than a very large one and often delivers equivalent quality for a clearly defined task.

2. Inference engineering

Efficient serving engines like vLLM significantly increase throughput. Quantization (lower compute precision) reduces memory and energy demand. Caching frequent answers saves on repeated computation.

3. Sovereignty and renewable power

A well-utilized local AI setup runs on predictable renewable power, without an opaque cloud footprint and without sensitive data leaving the building. Data sovereignty and sustainability are, in this case, the same architectural decision.

Does this affect your business too?

In a free initial call, we will look together at what matters most for you.

Example scenario: the oversized chatbot

A company rolls out an internal knowledge assistant. The obvious route: connect the largest cloud model via an API. It works, but every simple question ("Where's the vacation request form?") now occupies a high-performance model, and with every query, internal documents travel to someone else's servers.

The sensible redesign: a smaller, task-specific model handles the questions and uses RAG (Retrieval-Augmented Generation) to access your own documents. The data stays in-house, and the frontier model is reserved for the few genuinely complex cases. You scale compute to the task, not to the prestige of the model's name.

AI is both problem and solution

AI is not only a consumer, but also an enabler of sustainability. The IEA estimates that, with broad adoption, AI could avoid roughly 1,400 million tonnes of CO₂ emissions by 2035. The levers are relevant to SMEs too: grid management, predictive maintenance, material optimization, more efficient building technology. The German Federal Environment Ministry's "Green AI Hub Mittelstand" reports savings in material and energy in AI pilot projects.

The honest caveat: the Jevons paradox

When a technology becomes more efficient and cheaper, it gets used more, and absolute consumption can rise despite better efficiency. William Stanley Jevons described this for coal back in 1865. Efficiency is necessary, but not sufficient.

Sustainable AI means: a smaller model, clean power, and deliberate use, together.

What regulators now require

  • EU AI Act: providers of large foundation AI models must document their models' energy consumption. Energy becomes a quantity that must be reported.
  • CSRD / ESRS E1: sustainability reporting requires disclosure of energy consumption and emissions, including along the value chain (Scope 3). Cloud AI use therefore flows into the sustainability report.
  • Energy Efficiency Act (EnEfG): data centers are subject to requirements on PUE value, waste heat use, and renewable power. Make these metrics a selection criterion for your hosting provider.

Some of these rules are still evolving (the EU Omnibus package, the EnEfG revision). The direction, however, is clear: energy and climate transparency for IT is becoming mandatory.

Sustainability becomes a procurement criterion

Through the CSRD, the Supply Chain Due Diligence Act (LkSG) and the EU Corporate Sustainability Due Diligence Directive (CSDDD), large corporations pass sustainability requirements down to suppliers, via questionnaires, self-disclosures and codes of conduct. Even companies not directly covered by these laws are affected indirectly as a supplier. Anyone who can demonstrate that their own AI and IT landscape is energy-efficient, sovereign and documented has an edge in tenders.

The economic truth: cloud or local?

Because ongoing inference accounts for most of the effort, every efficiency measure directly lowers your operating costs. Energy efficiency and cost efficiency are the same thing here. The honest answer to "cloud or local?" is: it depends on utilization.

  • With fluctuating or low usage, cloud APIs are usually cheaper: no purchase, no operating overhead.
  • With steady, high utilization, the math can tip in favor of local hardware. We check that with a sample calculation for your specific case.
  • For many SMEs, a hybrid approach makes sense: local for baseline load and sensitive data, cloud for peaks and experiments.

Seven levers for sustainable AI in your company

  1. Right-size it: don't reflexively choose the largest model.
  2. Optimize inference: efficient serving technology, quantization and high utilization instead of idle time.
  3. Use caching: cache frequent answers instead of recomputing every request.
  4. Tighten prompts: shorter, more precise inputs and capped output lengths reduce compute demand.
  5. Choose renewable-powered data centers: favor providers with a low PUE, a high share of renewable power, and waste heat reuse.
  6. Keep hardware utilized: avoid idle time.
  7. Measure and report: track AI energy consumption, for control, credibility, and CSRD reporting.
Frequently asked questions

Before you reach out

Isn't AI generally a climate problem?

Looked at closely: data centers account for about 1.5 percent of global electricity use according to the IEA, but are growing fast. A simple text request uses little; reasoning, images and video use a multiple of that. What matters is deliberate use and the right model choice, not blanket abstinence.

Do I have to give up powerful AI models?

No. It's about right-sizing: the right model for each task. Smaller, specialized models are enough for routine work; reserve the frontier model for the few genuinely complex cases.

Does local AI pay off ecologically and economically?

With steady, high utilization, generally yes: local hardware is then cost-efficient, predictable with renewable power, and keeps data sovereign. With fluctuating load, the cloud is often cheaper. For many SMEs, a hybrid approach makes sense.

What do the EU AI Act and EnEfG mean for me?

The EU AI Act requires model providers to document energy use. The EnEfG sets efficiency requirements for data centers. Through the CSRD, your AI energy consumption flows into sustainability reporting, including as a supplier via Scope 3.

Isn't "Green AI" just greenwashing?

That risk exists, keyword: Jevons paradox. Efficiency alone can be eaten up by increased use. That's why Green AI, to us, means: a smaller model plus renewable power plus deliberate use. Not a climate promise, but a measurable practice.

Related articles

AI automationIT automation with AIAI and the works councilMarketing automation with AIGuide: AI strategy and organizationKnowledge hub: AI in practice

How efficient is your AI strategy?

Talk to us about a sovereign, energy-efficient AI architecture. We'll get back to you within 24 hours with a proposed time for an initial consultation.

Patrick Sager
Patrick SagerYour contact

By submitting this form you agree to the processing of your data under our privacy policy .