Insight
4 min read
"People are trying to estimate ChatGPT's energy consumption," Keus told Change.inc earlier this year. "That simply isn't possible. You really need access to the raw data from the chips to do that."
The turning point came about a year ago, when the Chinese open-source model DeepSeek appeared. "From that point on, open-source models have been getting better and better," says Keus. Good enough to build on. Good enough to self-host. Good enough to actually measure what is happening inside them. So he built GreenPT, a platform that shows you in real time how much energy each AI conversation costs.
A spellcheck on a document? 1.57 watt-hours. An extended conversation about your marketing strategy? More. How much more depends on something most people are unaware of. And that is where things get interesting for anyone building AI products.
Every time you ask a follow-up question in ChatGPT or another AI chat, the model does not only process your new question. It processes the entire conversation again. Every previous question, every previous answer, everything is run through the model once more. Keus puts it bluntly: "If you ask a question about a different topic, there is no need for that and you are better off starting a new conversation."
Consider what that means at scale. A company with 500 employees casually chatting with AI throughout the day, each within a single long conversation window, burns through exponentially more computing power with every follow-up. Not because the questions get harder, but because the context window keeps growing. It is like photocopying your entire filing cabinet every time you want to add a single document to it.
These kinds of insights only become visible when you measure. And measuring is precisely what GreenPT does differently from the rest.
GreenPT started on LeafCloud, a Dutch startup with a concept that sounds almost too elegant. They place servers in the utility rooms of apartment complexes, care homes and swimming pools. The servers have to run regardless. Buildings need heat regardless. LeafCloud captures the residual heat from computing and uses it to warm water. The result is an energy efficiency roughly ten times better than a conventional data centre.
For a founder who wants to build the greenest AI platform in the world, it was a perfect match. Until it wasn't.
LeafCloud could not scale fast enough. "We grew to a point where we needed infrastructure for a thousand paying customers," says Keus. LeafCloud's distributed model, elegant in theory, could not deliver that. He had to make a choice that every founder building on principles eventually faces: stay pure and stay small, or compromise and grow.
He chose Scaleway, a French provider. Air-cooled data centres, largely renewable energy, a share of nuclear energy offset elsewhere. "That is why it is 100 percent sustainable energy on paper," Keus told Change.inc, choosing the words "on paper" deliberately. No heat reuse, no revolutionary infrastructure. Just solid, certified, scalable.
For someone who wanted to build the greenest AI provider in the world, it is the kind of pragmatic decision that does not make for a great headline. But it is honest. And honesty, as it turns out, may be worth more than perfection.
The EU AI Act now requires energy consumption reporting. Dutch data centres already account for more than four percent of national electricity use. Most AI startups here have no idea what their products actually cost in energy terms.
GreenPT's CTO Cas Burggraaf squeezes 20 to 30 percent more efficiency out of their stack through model compression, quantisation and load distribution. They run open-source models from Mistral and OpenAI. The competitive advantage is not in the model. It is in how efficiently you run it.
The platform offers three things: a chat interface, an API for developers, and a tool that allows companies to build their own chatbots. Everything is self-hosted on European infrastructure, with no external API calls. User data never leaves their servers. In a market where American platforms continue to stumble over GDPR, and the EU AI Act now mandates energy reporting, that combination is quietly becoming a sales argument.
Around a thousand users now pay for GreenPT. Government institutions, mission-driven organisations, people who care about where their data resides. Not huge numbers, but enough to demonstrate that a market exists. "We use AI too much for simple things," says Keus. He knows that every new customer enlarges the footprint. The models were not trained sustainably. Not by anyone. But inference, actually running queries, is where the bulk of energy consumption lies and where you can still make a difference.
Google recently published real energy data for the first time. A median Gemini query now costs 0.24 watt-hours. A year earlier, the same query cost 33 times more. That is how fast this is moving, and how little anyone actually knows about it.
So here is the question. When energy reporting becomes mandatory, and it will, can you tell your customers what your product costs to run? GreenPT already can. Can you?
Van de 21 Nederlandse fintech-startups in dit overzicht is de Metropoolregio Amsterdam veruit het sterkst vertegenwoordigd. Historische financiële infrastructuur, internationale talentpools en innovatievriendelijk toezicht verklaren waarom juist daar zoveel bedrijven neerslaan.
Rondom de Rotterdamse haven en de mainport-infrastructuur van de Randstad clusters een opvallend aantal AI-startups in logistiek en mobiliteit. De aanwezigheid van grote verladers, havenbeheerders en logistieke dienstverleners blijkt een concrete aantrekkingskracht te hebben op jong technologiebedrijven.
Finse venture builder Reaktor Ecosystem opent zijn eerste internationale hub in Amsterdam en zoekt ervaren founders om samen AI- en tech-consultancies op te bouwen.
