On Tuesday, a search company bought half a nuclear power station.
Not the building. The electricity coming out of it. Google signed a twenty-two year agreement with the Finnish utility Fortum. It covers up to half of everything the Loviisa plant generates, from twenty-thirty out to twenty-forty-nine.
I want to tell you what that has to do with the microphone in front of me.
Because the machine I am speaking into has made fifty-nine episodes of this programme. And it has cost me nothing at all.
I am investigating one question today. Not arguing it. Investigating it, because I have been on both sides of this and I am still not certain where the line sits.
When everybody around you is buying scale, how do you work out whether you need any of it?
Start with what was actually bought, because the size of it gets exaggerated in both directions.
According to Google's own announcement, the company is putting thirteen billion euros into Finland over two years. That is about fifteen billion dollars. Data centres at Hamina, Muhos, Vaala and Kajaani. World Nuclear News reports it is Google's first nuclear deal outside the United States, and the company calls it its largest investment in Europe.
Fortum has roughly a billion euros of its own going into Loviisa. The point of that work is to keep the plant running until twenty-fifty. Google's money makes that arithmetic survivable, because a twenty-two year contract is a revenue line a utility can borrow against.
So a Finnish reactor gets two more decades of life. And a search company gets the joules.
Now widen the frame.
CNBC reported in February that the four largest American technology companies are heading toward roughly seven hundred billion dollars of capital spending this year. Other estimates run higher still. Amazon near two hundred billion. Alphabet in the high one hundreds. Meta and Microsoft each above a hundred.
Round it however you like. It is the largest private construction programme in the history of the species, and most of it is sheds full of chips and the power to run them.
Here is the part I want you to hold on to.
Almost none of that money is about making a model that can write a decent sentence. That problem is largely solved, and the solution turns out to be small. The money is about serving that sentence to a billion people at once, in a fifth of a second. And the system can never refuse.
Scale is not intelligence. Scale is availability.
And availability is the expensive part.
Let me tell you how I learned that, because I learned it the humiliating way.
I spent three years writing policy in a corner office. Good policy. Properly researched, correctly footnoted, defensible in any meeting you care to name. None of it worked. Not some of it, and not most of it. None.
You know why? I never met the people it was for.
I had bought the biggest available version of the wrong thing. The most rigorous document, the most complete framework, the most defensible position. And not one hour spent with a person who would have to live inside it.
That is the same error as buying a data centre when what you needed was a laptop. It is not a technical mistake. It is a purpose mistake, and no amount of budget corrects it.
So. What does this show actually run on?
One computer. An Apple M5 Max with a hundred and twenty-eight gigabytes of memory. It sits in a room in Bangkok.
Every word you have ever heard on this programme was written by a model running on that machine. Not one call goes out to a cloud. There is no API key anywhere in this pipeline, which means there is no key to leak and no bill to be surprised by. That was a decision, not an accident.
The voice is mine, cloned from about twenty seconds of me reading a paragraph out loud. Also local.
Since the tenth of August, that one machine has produced fifty-nine episodes across two shows. Twelve point eight hours of finished audio. About a hundred and twenty-nine thousand words, written, spoken, checked, and published.
The marginal cost of the next episode is the electricity to run a laptop overnight.
Now, before you go and cancel anything, I have to tell you what free actually costs. This is the part the people selling you local models leave out, and it is the reason I am not going to pretend this was clever.
This show has broadcast garbage. Real garbage, into real ears.
One episode ended with the words "be well, be good, be kind" repeated for ninety seconds. By the end it had degenerated into single letters. The voice model had been handed more time than the sentence needed, and it filled the silence with confident nonsense. Another episode shipped five separate bursts of invented words inside an otherwise clean read. A third dropped a clause out of the middle of a sentence, and nobody noticed until it was live.
None of those were caught by listening, because nobody listens to twenty minutes of audio before six in the morning. They were caught afterwards, by me, badly, and late.
So the pipeline now checks its own work. Every chunk of audio gets transcribed back by a second model and compared against the words it was given. If the machine did not say what it was told to say, the chunk is thrown away and rendered again. Three failures and the sentence gets cut in half and each half rendered alone.
Yesterday's episode scored ninety-seven per cent coverage. That number is the fraction of the script the machine can prove it actually spoke.
That is what free cost. Not money. About a month of evenings, and a genuinely embarrassing amount of published nonsense along the way.
Which brings me to the objection worth taking seriously, and I only get to do one of these properly, so this is it.
You might say I am comparing two things that have nothing to do with each other. And you would be right that I am not serving a billion people. I am serving a feed. If ten million listeners arrived tomorrow morning, this machine would fall over inside a minute. I would be renting exactly the infrastructure I have spent this episode calling somebody else's problem.
That is true. Hold on to it.
But notice what it concedes. The expensive thing was never the intelligence. It was the audience. In Finland, Google did not buy a cleverer model. It bought certainty. That means the model answers every time for twenty years, and never has to ask the grid for permission.
You are not serving a billion people. So a great deal of what you are being sold is a solution to a problem you do not have.
So here is the useful part. How do you actually tell the difference?
I use four questions. I call them the four Ps, and I ask them out loud, in this order, before I spend anything at all.
Purpose. Who is this for, and what breaks in their day if it does not exist? Not what it could theoretically do. What breaks.
Practicality. Can I build the smallest true version this week, with what is already in the building? Minimum viable product does not mean bad. It means minimal relative to your resources, and buildable upon afterwards.
Proof. What number would tell me I was wrong? Decide it before you start, because afterwards you will negotiate with yourself and you will win.
People. Who has to change their behaviour for this to work, and what is in it for them? People are driven by incentives, not instructions. Write that one on the wall.
Most projects die at the first question and nobody notices, because a large budget makes a bad purpose look like a plan.
Let me give you the version of this that actually mattered.
Sukhothai floods. The province needed to know when the water was coming. A vendor quoted forty million dollars for a system to tell them.
I built it for nothing. Two million data points, and about forty-five minutes of real coding. It gave the province thirty-six hours of warning.
Thirty-six hours is enough time to move a school. It is enough to move a hospital ward, and the animals, and the people who cannot walk to higher ground by themselves.
We cannot prevent floods. We are not God. But we can predict, and prediction is enough to keep people alive.
The forty million dollar system would not have delivered thirty-seven hours. It would have delivered thirty-six hours, a maintenance contract, and a consultant on a plane every quarter.
Now I am going to be honest about the swagger, because I do not want to leave you with the wrong lesson.
Forty-five minutes of coding is true. It is also the tip of about fifteen years. The forty-five minutes is what the fifteen years bought me. For you, that same build might be two days, and two days is a completely reasonable price for something that moves a school.
I will not leave you hanging. But I am not going to lie to you about the ladder either. You still have to climb it. What I am telling you is that the ladder is a great deal shorter than the vendor says, not that there is no ladder.
My cousin is twenty years old. She wanted a camera on her front gate, looked at the prices, and decided they were ridiculous. She built her own with an Arduino. It works. It cost her a weekend and about the price of dinner.
She did not have fifteen years. She had a problem she actually understood, which is the only prerequisite that has ever mattered.
So here is what I would do tomorrow morning, and I mean tomorrow, not next quarter.
Pick the smallest thing that genuinely annoys you and that you already understand. Install a local model on the computer you own right now. Give yourself one evening, not one roadmap. Build the version that works for exactly one person, and let that person be you.
Then show it to a second person. That is the moment you find out whether you built a tool or a hobby. Both are fine. But you should know which one you are holding.
One last thing, and then I will stop.
Loviisa first generated electricity in nineteen seventy-seven. It was built to power a country. For the next two decades, half of it is spoken for by a company that did not exist when the concrete was poured. It will answer questions from people who have not yet thought to ask them.
That is not a scandal. It is what happens when a new appetite meets an old machine.
The only question that was ever yours is whether the appetite is.