Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Most European data centres still won't publish water and power figures, study finds

Less than a quarter of the Netherlands' larger data centres publish how much electricity and drinking water they use, a year-long study by Lighthouse Report found. It was published on 30 September.

AI & modelsAnalysisRachel NwosuPublished: 30 September 20265 min readSources 12
Most European data centres still won't publish water and power figures, study finds
Less than a quarter of the Netherlands' larger data centres publish how much electricity and drinking water th

The researchers spent a year trying to get numbers out of data centre operators across Europe. They used freedom of information laws in several countries. According to NL Times, which worked with Lighthouse Report, the attempts yielded no results.

The legal hook already exists. The European Energy Efficiency Directive obliges data centres with an installed capacity of at least 500 kilowatts to report energy and water consumption. It has been in force for three years. Most operators still do not report.

The Dutch numbers show the gap plainly. The Dutch Datacenter Association counted 186 commercial data centres at or above 500 kW in the Netherlands at the end of last year. The Netherlands Enterprise Agency (RVO), which collects the data for the EU, has records for only 104 of them. Public figures exist for the electricity use of 44 centres and the water use of 47. That is less than a quarter of the larger sites. RVO told Trouw it holds more data on some centres but did not hand over the figures.

The stakes are larger than the reporting gap suggests. Statistics Netherlands (CBS) put Dutch data centre electricity use at 5.1 billion kilowatt-hours in 2024. That is 4.6 percent of national consumption, close to double the level five years earlier. Grid operator TenneT projects 10 to 15 percent by 2030. Grid capacity shortages already leave businesses and housing projects on waiting lists.

One company, one percent of a national grid

Where figures do surface, they are striking. RVO data released this summer showed that Microsoft's largest data centre in the Netherlands alone accounts for 1 percent of national electricity consumption. Google runs two large data centres in the country. It does not disclose the equivalent figure, according to NL Times. Water disclosure is thinner still. The Dutch estimate that data centres take about 0.1 percent of total water consumption is just that, an estimate.

That small share is the industry's usual defence. On water it is not obviously wrong. But a national average hides the local picture, and planning fights and drought plans turn on the local picture. The same week the Lighthouse findings ran, news coverage across Europe tracked moratoriums, abandoned projects and new transparency rules. The headline list includes a US Senate floor speech calling for water and energy guardrails. It also includes an EU plan to grade large data centres from A to G on energy and water. This piece relies on none of those beyond noting the direction of travel.

The reporting picture has a counter-example. In Lincoln, Nebraska, a local station reported that an improper redaction revealed the water and electricity usage of a Google data centre. The numbers exist, and someone decided they should not be read. The Lighthouse method was to ask formally, on paper, under law. It got less than the redaction error did.

The models behind the meters

The measurement problem sits on top of a fast-moving technical layer. This week's releases show how quickly the layer changes. Fulcrum said on 30 September that it had trained Echo, a writing model for style imitation, for less than $5,000, per its launch page. Fermion Research released Phonon-2, a speech recognition model that fits in a 164 MB download. It says the model averages 5.21 percent word error across seven English sets on the Open ASR Leaderboard. Cohere launched Embed 5 in Pro and Fast tiers, priced at $0.12 and $0.08 per million tokens respectively, according to the company's blog.

Smaller, cheaper, more specialised models change the arithmetic of a data centre. They do not always change it in the direction the headline suggests. NVIDIA published Kumo Tabular, an open foundation model for tabular prediction. It says the model ranks first on four benchmarks and comes in three sizes from 28M to 215M parameters, under the OpenMDW-1.1 licence. Bespoke Labs' Nimble, a 9B decision model, now runs locally through Ollama's new /v1/systemone endpoint. Ollama says it averaged 91ms per decision on an M5 Max. A classification that used to be a network round trip to a hosted model becomes a local forward pass.

Individual developers are pushing the same way. A writeup published on 30 September described running Qwen3.5 at 4-bit across 20 rented GPUs of 16 GB each, connected through a relay. It reached 0.6 tokens per second on the first run. The author measured 21 milliseconds of compute per token across the whole fleet and about 1.5 seconds of network latency, 98.6 percent of the total. Efficiency work at that level is about wires, not watts.

None of this resolves the disclosure question. Some of it sharpens the question. A model that runs on a laptop draws no data centre water at all. A model that runs on a rented cluster draws someone's water, in a jurisdiction that may or may not require the operator to say how much.

Why the silence holds

The obvious explanation is competitive. Water and power draw are cost data, and cost data is commercially sensitive. The less obvious one is that nobody is enforcing the directive hard enough to make silence expensive. Three years in, with a 500 kW threshold, the Netherlands can produce public electricity numbers for 44 of 186 qualifying sites. The RVO told Trouw it has more. It did not publish them.

Two things would change the picture quickly. The first is the EU rating scheme for large data centres, which would turn a reporting obligation into a visible grade. The second is local pressure, which is already arriving. The news list has a township moratorium, a developer abandoning a A$1.2 billion Sydney project, and county-level fights in the US. Operators who will not say how much water they use will find it harder to argue that the amount is small.

Until then, the best public figures on European data centre water come from freedom of information requests that failed. They also come from a redaction error in Nebraska and a survey of 186 Dutch sites where three quarters of the answers are missing.

Comments 0

Sources

12
  1. 01Most data centers refusing to say how much water, electricity they useEN
  2. 02Codex PricingEN
  3. 03Echo – A SOTA writing model for style imitationEN
  4. 04Phonon-2: most accurate open speech recognition model in a 164 MB downloadEN
  5. 05Cohere: Embed 5 – A New Family of Frontier Embedding ModelsEN
  6. 06Nvidia Kumo Tabular: Open Foundation Model for Tabular PredictionEN
  7. 07Ollama now supports Jev-style decision modelsEN
  8. 08I ran a 397B-parameter model on 20 GPUs with 16 GB eachEN
  9. 09AI models keep posting screenshots showing sensitive data from tech companiesEN
  10. 10Disrupting a coordinated model-distillation campaignEN
  11. 11Coop: A small language model pretrained by volunteersEN
  12. 12Ollama now supports Jev-like decision models all locally in 0.35EN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Rachel Nwosu

Rachel Nwosu

AI, models and technology

Rachel Nwosu covers AI, models and technology for FLASH24, working from public model documentation, benchmark releases and repository histories rather than press summaries, and she skips announcements that arrive without reproducible numbers. She checks training-data claims against dataset cards and reruns reported metrics where code is available. She spends much of her week interviewing researchers and engineers, tracking model launch calendars, and comparing vendor benchmarks with independent evaluations. Outside the desk she runs 3D printers, restores old computers, and tests how models learn from internet junk. She does not publish benchmark figures she cannot trace to a source.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.