---
title: "We Could Get the Data Centers"
description: "Europe's AI strategy depends on cheap compute, and we're protesting the thing that makes compute cheap."
url: https://juliussoerensen.com/blog/we-could-get-the-data-centers
updated: 2026-11-19
---

> Profile, projects and all blog posts: https://juliussoerensen.com/llms.txt

# We Could Get the Data Centers

> Europe's AI strategy depends on cheap compute, and we're protesting the thing that makes compute cheap.

Published Nov 19, 2026, updated Nov 19, 2026.

Europe's AI strategy is basically this: we don't have the labs, but we're good at using technology and building on top of it. I actually believe in that. Denmark is one of the most digitalized countries in the world, and my own company is the strategy in practice. We're two people, and in five months we've built something I think would have taken 40 people two years. We don't have the capital for 40 people, so we use AI. We're insanely reliant on it.

But building on top means you're a customer, and a customer lives and dies by the price. Right now the price is super generous. [SemiAnalysis measured](https://www.tomsguide.com/ai/the-math-doesnt-work-why-your-usd200-ai-subscription-is-secretly-worth-thousands) that a $200 Claude subscription gives you around $8,000 worth of tokens at API prices, and the $200 ChatGPT plan up to $14,000. Between us, my co-founder and I run three of each, maxed out. That's roughly $66,000 of tokens a month for $1,200. I'll admit that's the deal my company is built on. We couldn't pay full price.

Today Anthropic goes public. It [raised $81 billion](https://www.cnbc.com/2026/11/18/anthropic-ipo-ai-offering-price.html), and from now on it answers to shareholders. Their S-1 [implies an 85% gross margin](https://fourweekmba.com/ai-anthropic-s-1-85-5-gross-margin-up-from-29-4-a-year-ago/), so those $8,000 of tokens cost them roughly $1,200 to serve. I don't think anyone sells that for $200 forever, and I don't blame them. It just means that building on top only works if the real cost of tokens keeps falling.

The hedge is open models. Anyone can host them, so the price is set by whoever has the cheapest compute. And that's the part that worries me, because that's where Europe is behind. At [Oplema](https://oplema.com/) we serve the same open models from data centers in the US and the EU, and most of them cost more from the EU: 1.5x, 1.9x, and in the worst case over 4x. The closed models cost about the same in both places, because the lab decides. So the one place where Europe could control its own price is a place where we're already behind.

And then twelve days ago people [stood outside a Microsoft data center near Køge](https://via.ritzau.dk/pressemeddelelse/15274275/modstanden-mod-datacentre-vokser/) to protest it. Projects in Bjæverskov and Greve have already been dropped after local protests. I understand the worries about power and noise, but I find it a bit weird. We've picked a strategy that depends on cheap compute, and we're protesting the thing that makes compute cheap.

Meanwhile towns in the US have [started voting to ban data centers](https://www.wfsb.com/2026/10/20/east-lyme-zoning-commission-bans-data-centers/). I think that's an opening for Europe, if we want it. We're not getting the labs. We could get the data centers.

P.S. This is how much they actually use. Data centers used about [1.5% of the world's electricity](https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai) in 2024. And they already help pay for the grid: [an EPRI study](https://www.marketplace.org/story/2026/07/10/data-centers-lowered-electric-bills-in-some-places-for-now) found that data centers lowered household power prices in the US, because the fixed cost of the grid gets spread over more use. Microsoft has [promised to pay the full cost of its power](https://www.geekwire.com/2026/microsoft-responds-to-ai-data-center-revolt-vowing-to-cover-full-power-costs-and-reject-local-tax-breaks/), grid upgrades included. Here in Denmark, Meta's data center in Odense already [heats homes](https://www.ramboll.com/projects/energy/meta-surplus-heat-to-district-heating) with its surplus heat, and Microsoft's will [heat about 6,000 homes](https://investindk.com/cases/microsoft-data-centre-set-to-provide-surplus-heating-for-6-000-danish-homes) in Høje-Taastrup.

As for water, US data centers used [17 billion gallons](https://eta-publications.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report) for cooling in 2023. US golf courses used [about 530 billion](https://gcmonline.com/research/news/from-acre-feet-to-outcomes-interpreting-the-2024-gcep-water-survey). And done right, it can be close to zero: Microsoft's new data centers [use no water for cooling](https://www.microsoft.com/en-us/microsoft-cloud/blog/2024/12/09/sustainable-by-design-next-generation-datacenters-consume-zero-water-for-cooling/).

<figure class="post-figure" role="img" aria-label="Bar chart: US golf courses used about 531 billion gallons of water in 2024; US data centers used about 17 billion gallons for cooling in 2023.">
  <div style="display:grid;grid-template-columns:max-content 1fr;gap:10px 12px;align-items:center;font-size:14px;">
    <span style="color:var(--muted);">Golf courses</span>
    <div style="display:flex;align-items:center;gap:8px;" title="US golf courses, 2024: about 531 billion gallons">
      <div style="height:22px;width:85%;background:var(--strong);border-radius:0 4px 4px 0;"></div>
      <span style="color:var(--ink);white-space:nowrap;">531</span>
    </div>
    <span style="color:var(--muted);">Data centers</span>
    <div style="display:flex;align-items:center;gap:8px;" title="US data centers, cooling, 2023: about 17 billion gallons">
      <div style="height:22px;width:max(2.7%,4px);background:var(--strong);border-radius:0 4px 4px 0;"></div>
      <span style="color:var(--ink);white-space:nowrap;">17</span>
    </div>
  </div>
  <figcaption>US water use, billions of gallons per year. Golf: GCSAA survey, 2024 (1.63 million acre-feet). Data centers: on-site cooling, Lawrence Berkeley National Laboratory, 2023.</figcaption>
</figure>
