I spent this week at NVIDIA GTC in Berlin, where the appetite for compute felt almost insatiable. Neoclouds are springing up everywhere, demand beyond the major AI labs appears to be growing quarter after quarter, and the entire data center supply chain is expanding at a feverish pace. Watching it unfold, I kept coming back to how quickly a promise about intelligence becomes a demand for land, electricity and water.
In September, Google announced at least €13 billion of investment in Finnish digital infrastructure for 2027–2028.
On 6 October, Finland’s Permit and Supervisory Agency told the company’s local subsidiary to suspend preparatory work that significantly alters the environment at two planned data center sites, in Muhos and Kajaani, until their environmental assessments were complete.
Google confirmed to the Finnish broadcaster Yle that around 330 hectares of forest had already been cleared at Muhos and just under 200 at Kajaani. Google has apologised, acknowledged falling short of its own standards and said it acted according to its understanding of the law. It has also said its landscaping plans include planting trees across 130 hectares at Muhos.
What stayed with me was the regulator’s explanation: large-scale clearing before the assessment is complete can narrow the alternatives that can still be considered. An environmental assessment is supposed to help society choose, but a cleared forest means that part of the choice has already been made.
These are investments intended to serve many products for many years. Their lasting role makes the decisions about their scale and location more consequential.
The IEA expects global data center electricity use to roughly double, from 485 TWh in 2025 to about 950 TWh in 2030. Energy use per AI task is falling fast, while wider adoption and more demanding applications such as reasoning and AI agents drive total consumption higher.
That makes a question about engineering increasingly political: how much computation does a given useful task actually require? The answer depends partly on architectural technical choices. Those choices can change substantially with research. A community considering a new campus is being asked to make a much longer commitment. We owe it a serious examination of how much of the proposed footprint is inherent to the service, and how much we could reduce through better engineering.
At industrial scale, a choice about software becomes a choice about a landscape
There is a commercial tension here. Suppliers have incentives to improve efficiency and expand the volume of computation they sell. Society has to decide what it receives in exchange for the resources committed.
A larger investment is an impressive announcement; its value still depends on what that investment enables. A breakthrough in efficiency changes what an existing supply of power can deliver. That deserves a place in industrial policy alongside the construction of new capacity.
At @Giotto_ai, we built our smallest model to run on a single GPU. We chose that constraint because we wanted to understand how far research into reasoning and architecture could take us within infrastructure an enterprise could own. It makes the computation required for a useful result a research problem in its own right.
For enterprise workloads that smaller models can handle reliably, lower hardware requirements expand the choice of where a system runs and who operates it. More businesses can keep their data and AI operations under their own control.
Efficiency changes who can own AI
This is a material part of sovereignty: the economic ability to own and operate the systems on which you depend. Europe should pursue that ability with the same ambition it brings to attracting investment.
I want Europe to build the models and infrastructure its future requires. I also want every builder, including us, to show what work its systems will perform, at what quality, and with what resource requirements.
Alternatives that use existing sites or less hardware deserve a serious comparison, including their reliability and full environmental cost.
The burden of proof should grow with the footprint. The decisions we make about land will outlast many generations of models.
Our children will inherit those decisions long after this year’s benchmarks have been forgotten.
