Summarize this article with:
AI is becoming a material energy consumer at a moment when European grids are already under strain, and the question is moving from corporate reporting into procurement.
This article looks at where the footprint of a deployed AI system actually sits, which turns out to be inference rather than training, and works through the three levers that
change it, in order of effect. It covers why right-sizing a model beats relocating it, what efficiency-focused labs like Pruna are doing about it, what a company like GreenPT
is proving by publishing its carbon data, and it is explicit about the figures Eden AI cannot give you and why we will not invent them.
The bill is arriving
For a decade, the environmental cost of software was somebody else's problem. Compute was abundant, cheap, and invisible, and the energy question belonged to the datacenter operator rather than to the person writing the code.
AI ended that. Generative models consume orders of magnitude more compute per useful operation than the software they are replacing, and they are being deployed into products used millions of times a day. Grid operators across Europe are now planning capacity around datacenter demand. Regulators are tightening what operators must report and what companies may claim. Procurement teams have started asking for environmental figures alongside latency and price.
None of this is abstract any more. It shows up as a question in a security questionnaire, as a line in an RFP, as a slide your CTO has to present.
And most teams answer it wrong, because they answer the question that was true two years ago.
The question moved
Two years ago, AI and energy meant training. A frontier model is trained once, the number is enormous, and it makes a good headline.
That framing is out of date for almost everyone reading this. You are not training frontier models. You are calling them, thousands of times a day, and that is where your footprint sits.
The research has followed. Work on carbon-efficient inference, such as the SPROUT system published at EMNLP 2024, targets the serving side rather than training, and reports carbon reductions above 40 percent on real workloads without changing the model itself.
The practical consequence is simple. If your AI footprint matters to you, the lever is not where the servers are. It is what you send them.
Here is why. Moving a workload from a grid at 400 g of CO2 per kWh to one at 50 g is a meaningful improvement, and it is roughly a fixed multiplier on whatever you were already consuming. Moving from a frontier model to a small one that handles the task just as well can cut the compute itself by an order of magnitude.
One change divides. The other divides the thing being divided.
Most teams do the second badly, because it requires knowing the alternatives, having them available behind the same interface, and being able to switch without rewriting anything. Which is, as it happens, what a gateway is for.
The reflex that misses
The common first move is to look for a green datacenter. It is a real lever and we will come back to it, but it is the third one, not the first.
The three levers, in order
Right-size the model. A classification task, a short extraction, a reformulation, a moderation check: these do not need a frontier model. They need a model that is good enough, and the gap in compute between the two is not a few percent.
This is the lever nobody pulls, because pulling it means running an evaluation, and running an evaluation means having every candidate model reachable through one call.
When switching model is one string, the evaluation takes an afternoon. When it means three vendor contracts, it never happens.
Use models built to be efficient. This is a real engineering field now, not a marketing claim. Pruna, based in Munich and Paris, maintains an open-source optimization framework with close to fifty algorithms behind a single `smash` function, and ships models built on that work.
Their video model, live on Eden AI, shows the point in a way you can feel in the invoice. Video is billed per second of output, so the real cost was never the clip you publish, it is the ten you deleted first. P-Video-2 has a draft mode for exactly that: you iterate cheap and pay full price only for the take you keep. Fewer wasted seconds of generation is fewer wasted seconds of GPU.
video/generation_async/pruna/p-video-2
Then care about the electricity. Once the workload is the right size, where it runs starts to matter. The European providers on our platform sit on some of the cleanest grids available: Scaleway and OVHcloud in France, IONOS in Germany, Infomaniak in Switzerland. The French and Nordic mixes are structurally low carbon, and that is a property of the grid, not a claim anyone has to certify.
Who is pushing this
The sustainability question is producing actual companies now, not just position papers.
GreenPT, from the Netherlands, is the clearest example. They run on European infrastructure and renewable energy, and they claim up to 40 percent lower CO2 per query than the major clouds. What makes them interesting is not the number, it is that they publish energy and carbon figures back to the user instead of hiding them. That is the part the rest of the industry does not do, ourselves included.
They are not on Eden AI today. We are naming them because the direction they are pushing is the right one, and because a market where buyers can compare carbon the way they compare latency would be a better market for everyone in it.
The lever we can hand you today is choice: every model, one interface, switch in one string, with the region of each one visible. That is not a green label. It is the thing that makes right-sizing possible, and right-sizing is where the energy actually is.
Where to start
Take your highest-volume call, the one that runs thousands of times a day. Not the impressive one, the boring one.
Find the smallest model that still passes your evaluation on it. Measure what happens to cost, because cost is the proxy you do have, and compute and cost move together.
You will usually find you were paying for a lot of capability you were not using. The energy follows the same curve.
![[AUTO-DRAFT] OpenAI Quietly Cut Codex Context Window by 27 Percent: Why Provider-Specific Limits Are Not a Contract](https://cdn.prod.website-files.com/61e7d259b7746e3f63f0b6be/6a74828e12326e8199f45718_6a74828c58976902eb769454_openai-quietly-cut-codex-context-window-by-27-percent-why-provider-specific--openai.png)
