DEAR EUROPE - WE HAVE A COAL PLANT THAT WOULD LIKE TO HELP YOU CONSUME LESS COAL

A FEBRUARY ARTICLE IN FINANCE RUINED MY COFFEE, part 2

A 600 MW unit for hire, gently used, specialises in powering electricity-saving analytics. Apply within.

In the previous part of this electricity-meter trilogy we talked about bytes - about how many of them you need to store a meter reading properly, and how that number swings between almost nothing and almost everything depending on how seriously you take the word "properly." We arrived at a figure of roughly 150 bytes per reading, and from there we sketched the rough size of the national meter-data backend.

This part is about what happens when you also turn on the data centre. 

Spoiler: by the end of the post we will be writing a small classified advertisement for a coal plant. Try to enjoy the journey.

A NUMBER ON ITS OWN IS A RUMOUR

If your meter sends a reading that says only "two hundred and thirty," you can do almost nothing useful with it. Two hundred and thirty what, from where, on which phase, in what direction, at what moment, with what accuracy. The number is a rumour, and to turn it into information you have to wrap it - you add the unit, you add the timestamp, you add the meter identifier, you add the location in the grid, you add a flag that tells whoever reads it next whether the value is a real measurement or an interpolated guess. None of those wrappings are optional if you ever intend to do anything serious with the data.

This is the whole reason a thing called Smart Data Models exists, and the whole reason network operators eventually agree on shared vocabularies. It is not technical religion. It is the simple acknowledgement that if every utility describes a meter reading in its own private dialect, then the country ends up paying somebody to translate between them for the next forty years, and that somebody usually charges by the hour.

The shared vocabulary is a bargain. It says: I will write this reading down in a form that is verbose and a little cumbersome, but anyone who reads it will know exactly what it means without having to call me. That bargain is worth the verbosity. The verbosity, however, is real, and somebody will eventually have to pay for it.

THE SAME READING, WRITTEN SEVEN WAYS

A meter reading can be written as a tiny SMS-style shorthand that fits comfortably inside something the size of this sentence, or it can be written as a small novel with chapter titles and footnotes and a polite preamble explaining what every chapter title means. Both versions contain exactly the same physical information, but they look very different on disk - same readings, same meters, same moment in time, but roughly a thousand times the verbosity between the cheapest and the most expensive way of writing them down.

For a long time nobody had to worry about which version they were paying for, because the meters were sending data once every fifteen minutes or once a day, and nobody was sampling them at any cadence that mattered. The wire was empty, the storage was cheap, and the bill was invisible. 

That world is quietly ending. The customer ports already installed inside houses are pushing readings every few seconds today, once per second on the Dutch DSMR standard, every couple of seconds on the Norwegian and French equivalents, somewhere between one and ten seconds on the Slovenian P1 and I1 interfaces. 

The European Commission has introduced a polite regulatory term for the next phase, "near real-time," and has carefully avoided putting an actual number next to it, which is the kind of phrasing regulators use when they know it is going to be seconds and would prefer not to be the first to say so out loud.

AND THEN SOMEBODY WANTS TO PUT AI ON TOP

Storage by itself is not the interesting part of the bill. Disks are cheap and efficient, and even the largest version of a national meter dataset fits inside a server rack that does not need its own power station. The interesting part of the bill lives one floor above the storage, in everything that happens to the data after it has been written down - indexing it, replicating it, serving it, and, increasingly, running models over it.

Everybody now wants to put AI on top of their meter data, and the uncomfortable thing is that the interesting AI use cases are not the silly ones. 

Nobody serious is going to ask a chatbot to write a poem about a consumption curve. 

The interesting cases are entirely sensible - forecasting where the grid will be in two hours so the operator can pre-position resources, detecting that something has gone wrong on a feeder before any customer picks up the phone, reasoning across many simultaneous signals to figure out which one is the real anomaly and which one is just a warm Tuesday afternoon, helping a customer understand why their bill suddenly grew an extra leg. Every one of those is a sensible thing to want, and every one of them costs electricity. Not catastrophic amounts on its own, but the multipliers add up quickly, because a country has a lot of meters, a grid has a lot of feeders, a day has a lot of seconds, and an operator might quite reasonably want to run several of these models several times an hour.

THE QUIET MATH OF THE WHOLE THING

We can put a rough number on this. Not a precise number, because precise numbers belong in a design document that nobody reads, but a number that is honest to within a factor of two or three either way. Take TEŠ6, the only remaining active unit of the Šoštanj Thermal Power Plant, rated at six hundred megawatts of electrical output. Use that as the ruler. Now lay the meter-data backend next to it.

The "modest path" assumes we keep more or less the current cadence and dashboards, with no AI worth mentioning. The "realistic path" assumes the cadence the wire is already delivering, a properly modelled storage layer, and sensible AI working on aggregated data at substation, feeder and regional level. The "ambitious path" pushes the cadence toward genuine seconds and adds AI at every meaningful aggregation level, including agentic reasoning where it pays off. The numbers include typical data-centre overhead. They will move with implementation choices, but the orders of magnitude are stable.

The headline reading is straightforward. At Slovenian scale, one ageing coal unit is roughly a hundred to five hundred times bigger than the data backend in question. At European scale, the same data backend reaches the size of one entire TEŠ6, give or take.

THE UNCOMFORTABLE COINCIDENCE

The same TEŠ that is currently being closed - quite reasonably, after decades of service, with the official phase-out set for 2033 and the practical closure coming around 2027 - has, on its nameplate, the size of the European meter-data backend that the EU is in the process of building. The thing we are politely retiring, after a long and useful life, happens to have approximately the dimensions of the thing we are politely building in its place. Not in Slovenia, where we are oversized by a factor of one or two hundred, but across the whole EU, the arithmetic settles where it settles.

This is not a serious proposal to keep TEŠ open. The plant is being closed for very good reasons - climate, economics, the fact that lignite is not, on close inspection, an ideal twenty-first-century fuel. The workers in Velenje and the wider Šaleška valley will need a careful and respectful transition, and the country has rightly committed to providing one. None of that is up for discussion in this post.

But the coincidence is hard to look away from, because the data centre that is supposed to help us optimise our electricity consumption - to monitor the renewable transition, to balance the variable supply, to read every meter every few seconds, to run the AI that turns all of that into intelligence - is going to be powered by something. And that something has to come from somewhere. And the irony of building one approximately to power the other, in almost equal sizes, is uncomfortable enough to be worth saying out loud.

Especially since this is not an exclusively Slovenian habit. Microsoft has signed a contract to restart the Three Mile Island nuclear plant specifically to power its data centres. AWS has bought a data centre directly next to a different American nuclear plant. The pattern is global. Plant that nobody wanted, meet workload that nobody budgeted for. Have a coffee. Get to know each other.

THE SMALL CLASSIFIED ADVERTISEMENT WE THREATENED EARLIER

Dear Europe. We have a slightly used 600-megawatt thermal power plant, gently retired, located in central Slovenia, with excellent transport links to most of the union. It was originally designed to provide approximately one third of a small country's electricity needs and is now surplus to our requirements. Suitable for powering, at full European scale, one continuously running data centre dedicated to reading every electricity meter in the union once every few seconds, applying serious analytics on top, running AI models for forecasting and demand response, and producing the kind of intelligence that should, in time, help everybody consume less electricity. Some assembly required. One careful previous owner. Will throw in a fully equipped coal mine, lightly used, for free.

The pitch is absurd, of course, which is the point. We are not actually offering anybody a coal plant. We are pointing out that the European data infrastructure being built to optimise electricity consumption is, at full scale, approximately the size of the coal plant Europe is currently retiring in order to consume less electricity. That is not a useful place to land. It is, however, a useful place to look at honestly, before we land somewhere even worse.

A QUIET WORD TO THE PEOPLE WRITING THE RULES

Brussels is famously good at writing energy directives. It is less famously good at calling an engineer first to check whether the directive, as drafted, can be implemented in a way that does not require a new power plant. This is not a complaint. It is a polite suggestion that the phone could occasionally ring in the other direction.

Common sense - the kind that asks "and how much electricity will this thing actually require to operate, in continuous megawatts, once we have built it?" - is a perfectly respectable engineering input to a policy document, especially for policy documents that are themselves about electricity. The "near real-time" right enshrined in EU Directive 2019/944 is a beautiful idea and an excellent piece of consumer protection. It will also, depending on the cadence Brussels eventually settles on and the storage architecture each member state happens to deploy, cost somewhere between a small server rack and a large coal plant per country. Knowing which of those it is going to be, in advance, is the kind of thing that benefits from a phone call.

There are excellent engineers in every member state who would happily pick up the phone. They are also, on average, easier to find than one would expect, because a good number of them have already written about exactly this question on their own websites.

CONFESSION TIME

I did the arithmetic for this post with the help of Claude, an AI model, sitting at my kitchen table during my morning coffee time. I made calculations by myself and then asked AI for its own calculations. More or less they fit.

I am not a data-centre engineer. I am not a hyperscaler. I am a person who uses context brokers, implements them where they are actually needed, and pushes Context DataHub and NGSI-LD into the places where they solve a real problem rather than where they happen to look good on a slide. I assume I made mistakes along the way - probably several of them, with some figures off by a factor of two and others off by a factor of five. Somebody who actually runs a serious meter data centre will eventually correct me in the comments, which is precisely what I would like.

Even if the real bill turns out to be half what I am suggesting, or twice, or five times, the shape of the conclusion does not move. 

Data costs money and energy, processing the data costs more again, AI on top costs significantly more even when used for entirely sensible purposes, and collecting all of this data without first deciding what you intend to learn from it is, in the most literal sense, ecologically dumb. 

There is also significant room for optimisation in the picture I have sketched - smaller models, smarter storage formats, the simple discipline of looking at less data more carefully. The picture above is the trajectory, not the destination. Paying attention is the cheapest line item on the entire account.

THE QUIET PARETO OF THIS WHOLE STORY

In a slightly modified Pareto principle for this case, roughly one percent of the storage-and-processing budget gets you roughly one hundred percent of the interoperability value, as long as you stop confusing the wire format with the storage format and the storage format with the analytics format. The shared vocabulary belongs at the boundaries, the dense format belongs at rest, and the analytics belong on a layer that knows exactly what question is being asked, so it can fetch only what it needs, run only what is necessary, and go quiet again.

On a zero-to-one-hundred scale of where the industry sits today - where zero is "a folder of CSV files on somebody's laptop" and one hundred is "a fully layered, fully governed, AI-aware data architecture humming away in the background" - most national meter backends sit around twenty to thirty %. A properly designed context broker with a streaming ingestion layer, a columnar temporal store, a policy gateway and a thin shared-vocabulary view exposed outward sits comfortably above seventy. 

Nobody is operating in that upper bracket yet at national scale. There is, however, the possibility of getting there without quietly rescuing a six-hundred-megawatt coal plant to power the analytics. We happen to be working on exactly that kind of architecture, which is the polite way of saying that our technology is already slightly ahead of most of our customers' organisations - still a very comfortable place to be.

DISAGREE, COMMENT, OR WISH TO KNOW MORE?​

Curious how a properly layered context broker, the right transformation engine, and a thoughtful policy gateway could let you keep all the interoperability value of a smart-meter dataset without quietly volunteering an entire coal plant to make the analytics work?

BUSINESS AND LEGAL STUFF

For the avoidance of doubt, should the European Union, acting on the strength of this analysis, choose to acquire TEŠ for the purposes described above, the author respectfully reserves the right to discuss a modest advisory fee of 2,5 % of the agreed sale price, payable in any major European currency or, where mutually agreeable, in kilowatt-hours.