The giants consume as much as a whole country
One of the biggest cloud giants now consumes as much electricity as an entire European country, Denmark. Since 2020, its electricity consumption has risen by 244 per cent and its emissions by 80 per cent: a 25 per cent increase in the last year alone. Another giant has increased its electricity consumption by 37 per cent in twelve months, and its emissions by 18 per cent. This is the biggest increase in its history.
These are the figures from their sustainability reports, published this summer. And there is one detail that strikes me more than the absolute figures. One of the two companies covers 100 per cent of its energy consumption through the purchase of renewable energy, yet its emissions are still rising. Around 80 per cent of its carbon footprint stems from the supply chain: the construction of data centres, and the production of chips and servers. No green certificates apply there.
The overall picture is no different. According to the International Energy Agency, by 2025 electricity consumption by data centres had risen by 17 per cent, whilst that of centres dedicated to artificial intelligence had risen by 50 per cent. By 2030, it will double, reaching around 950 TWh. The climate targets of the tech giants remain unchanged, set for 2030 and 2040. They seem increasingly out of reach.
The problem isn’t AI. It’s the model.
I don’t think the giants are the villains of the story. I think their model has a structural flaw: always building more and more, always bigger and bigger, and promising to make up for it later. It’s a race in which efficiency always takes second place.
In our sector, there is a widespread misconception: that being modern means always having the latest hardware, the greatest capacity and the most features. I see it the other way round:
«Being at the cutting edge doesn’t mean changing servers more often. It means doing more work whilst using less energy.»
Sustainability isn’t about setting a target for fifteen years’ time. It is measured today using figures that anyone can verify. And to measure it, you need an infrastructure where you know exactly what it does, how much energy it consumes and why it is switched on.
We know nothing about what came before. We know everything about what came after.
Whenever a customer moves to Artera from a major public cloud, the same thing always happens. We try to work out how much they were using before, but we can’t. Nobody knows: neither we nor, often, the customer themselves. As for afterwards, however, we know everything, in detail and automatically.
In my view, this difference is the real crux of digital sustainability.
In a pay-as-you-go public cloud, the system encourages waste. You spin up an extra instance “just to be on the safe side”, services multiply, and you forget about resources that have been running for months. The bill fluctuates, and nobody really has the full picture.
Artera offers another service: managed private clouds, which become part of the client’s organisation. In practice, this means:
- fewer services, each with a reason for existing;
- more centralised, and therefore easier to optimise;
- Flat-rate and transparent charges, not based on usage: the customer knows how much they’re spending, and we know what’s going on;
- stability, because a bespoke infrastructure does not become overloaded at every peak.
A cloud you know is a cloud you can make efficient. One you don’t know, you can only pay for.
Our cards, face up
If I criticise others, I have to lay my cards on the table. All of them, even the less flattering ones.
Around 300 servers. Not thousands of warehouses: an infrastructure we know inside out, server by server.
The average PUE across the DHH Group’s data centres is around 1.3. This means that for every kWh that reaches the servers, we spend around 0.3 on cooling and everything else. The ideal figure is 1.0: we’re working on it, but it’s a reasonable figure.
Lugano, Breganzona. The Swisscom data centre has been purchasing renewable electricity since 2010. Let me explain this properly, because this is where a lot of greenwashing stems from: the non-renewable portion of the grid is covered by certified guarantees of origin. This does not mean that every electron comes from a hydroelectric dam. It means that this renewable generation is tracked and is not counted twice.
Zurich, Rümlang. The NTT data centre uses free cooling, that is, outside air when it is cold enough, and recovers heat using heat pumps to heat offices and systems. It holds ISO 50001 certification for energy management. Achieving 100 per cent renewable energy here is a stated target for 2030, not yet a reality. I’d rather say so myself.
Inert gas fire suppression. In Zurich, the system uses argon or nitrogen, gases already present in the air, instead of chemical agents that have an impact on the climate.
Redundancy, yes. Generator sets and UPS systems consume resources, and we’re not about to do without them: ensuring uninterrupted service is what we do. That is precisely why every other choice must be efficient.
Hardware lifecycle. In my view, this is the most underestimated aspect. Much of a server’s environmental impact occurs when it is manufactured, not when it is switched on. The longer it operates reliably, the lower its environmental impact. Our approach:
- to install a server sized to meet actual requirements, without any unnecessary capacity;
- make the most of it by stabilising loads and monitoring consumption;
- to maintain and update it for as long as it remains safe, reliable and efficient: on average, between 3 and 5 years;
- replace it only when the actual benefit justifies the cost and impact of the new one;
- Erase data in a verified manner, then reuse and recondition; recycle only what is no longer usable.
AI in the right place
Artificial intelligence consumes far more energy than traditional technologies. The tech giants are right about that. Where they go wrong, in my view, is in the idea that every query must go through the largest model, in the largest data centre, on the other side of the world.
Most of a company’s real-world use cases (summarising documents, categorising requests, responding to customers) do not require a massive model. They need the smallest model that does the job well.
At DHH, we have developed Regolo.ai, a European inference service using open-source models. It runs on GPU servers in European data centres powered entirely by renewable energy, cooled without using water, and certified to ISO 14001. Most importantly, it shows how much CO₂ each model you use produces. This means that choosing a model also becomes an environmental choice, based on the figures.
It’s the same principle as with a private cloud: if you can measure it, you can reduce it.
For agencies: “Is your hosting sustainable?”
If you’re a web agency, sooner or later a client will ask you this. Don’t just say, “Yes, it’s green.” Back it up with figures. Here are the questions to ask your provider, and our answers.
| Question to ask your provider | Artera’s response |
|---|---|
| What is the PUE of your data centres? | An average of around 1.3 across the group’s farms |
| Where does the energy come from, and how is it certified? | Lugano: renewable since 2010, with guarantees of origin. Zurich: free cooling and heat recovery, aiming for 100 per cent renewable energy by 2030 |
| How long does a server remain in service? | On average 3–5 years, with targeted updates rather than scheduled replacements |
| What happens to decommissioned servers and hard drives? | Data erasure verified, followed by reuse and refurbishment; recycling only if the items are unusable |
| Pay-as-you-go? Do you know how much energy my infrastructure uses? | Flat-rate charges; consumption is measured automatically |
| What about artificial intelligence? | Inference on Regolo.ai: renewable energy, no water consumption, CO₂ visible per model |
If your provider can’t answer the first question, you already have the answer to yours.
A heater with an IP address
The giants will carry on building. Fair enough: the world needs computing power, and AI is here to stay. But every company can choose which side to be on. It can choose infrastructure that is measured rather than promised, scaled rather than inflated, and kept running as long as it works well.
Whenever I look at a rack, I ask myself just one question. Is this machine doing any useful work?
Because if it doesn’t do any useful work, it isn’t a server: it’s just a heater with an IP address.
