> ## Content Index
> Fetch the complete content index at: https://julie.beliao.fr/llms.txt
> Use this file to discover other available public pages before exploring further.

# Who owns the meter
- URL: https://julie.beliao.fr/ai-public-utility-who-owns-the-meter/
- Published: 2026-08-12T13:25:08.000Z
- Updated: 2026-08-12T13:25:08.000Z
- Description: Sam Altman says intelligence will be a utility, bought from his company on a meter. But a utility is a set of obligations, not a business model, and the history of water shows what the difference costs.
- Author: Julie Valérie Belião
- Tags: Public AI, AI Governance, News, Public Sector

For some time now I have been collaborating with non profit organisations, with community groups and with people inside governmental organisations, all of them trying to figure out what to do with machine intelligence for the people they serve. What emerges from these collaborations, every time, is the same conclusion: what the field needs is a public utility. A real one, with the interest of people at its core, and the civil benefit as the goal and not as a byproduct.

This year, two announcements showed that the question is not just theoretical.

In July, South Korea announced[ AI for Everyone](https://www.upi.com/Top%5FNews/World-News/2026/07/13/ai-for-everyone-public-services/9121783997023/?ref=julie.beliao.fr), a free general purpose assistant for every citizen, run by the Ministry of Science and ICT, with a rule that at least half of the system must run on domestic models, and with public money and public GPUs behind it. You can argue about the design, and people do, but a government decided that this is infrastructure and started to fund it like infrastructure.

In May, Sam Altman told[ an audience in Washington:](https://www.tomsguide.com/ai/people-will-buy-intelligence-from-us-on-a-meter-chatgpts-ceo-sam-altman-has-critics-worried-with-his-ai-vision?ref=julie.beliao.fr) "We see a future where intelligence is a utility, like electricity or water, and people buy it from us on a meter."

So the word utility is already being used, and not always to mean the same thing. Mozilla uses it seriously in their[ Public AI paper](https://assets.mofoprod.net/network/documents/Public%5FAI%5FMozilla.pdf?ref=julie.beliao.fr). In Altman's sentence it means one private company selling intelligence to everybody, on a meter that the company owns, and for me that is not a utility at all. I wrote this to reflect on what a utility really is, and what it would take to build a real one for machine intelligence.

## A utility is a set of obligations, not a business model

The easy version of my argument would be to say that a utility is public and what Altman sells is private, and I want to get rid of this version myself, because it is wrong. Thames Water is private, and Veolia and Suez, the two French groups that run the water for thousands of cities in France and elsewhere, are private and listed on the stock market, and all of them are utilities. What makes them utilities is not who owns them, it is what they are obliged to do.

A utility has a duty to connect you, which means serving everybody in its area, including the addresses that lose money. Its price is regulated, because people cannot shop around for another pipe. Its quality is measured by somebody who does not work for it, and the measurements are published. And when it fails there is a public accounting, which is why you have heard of Thames Water and why the fights about it are loud and boring and happen in public. For me this last one is the most important, because a utility is an institution you are allowed to be angry at.

Now looking at Altman's sentence and checking it against this list: 

- Duty to connect, no.
- Regulated price, no.
- Independent measurement, no.
- Public accounting, not really.

The only property of a utility in his sentence is that you get billed every month. He is allowed to sell what he sells, bottled water is a real business and it will not go away. What I am not ok with is taking the word, because if utility starts to only mean that we bill you monthly, then we have no word left for the thing that actually needs to be built, and the people who need it the most will be told they already have it.

## Bottled water is not neutral

Here is where I separate from the polite version of this argument, which says that the premium products and the public one can coexist happily. 

Sometimes yes, but not automatically.

Bottled water is not simply tap water with a nicer label. It is mostly packaged in plastic, transported by truck, sold at a markup of several hundred times, and depending on the study, its energy cost is between hundreds and two thousand times the cost of tap water. And it is a product that had to create its own need, because it grew by making people distrust their own tap. So the story of bottled water is not a nice coexistence story, it is the story of one product that profits from the neglect of the other.

I think the parallel holds, when you look at the third most common use of machine intelligence last year, on the [Harvard Business Review list ](https://hbr.org/2026/06/how-people-are-really-using-ai-in-2026?ref=julie.beliao.fr)I use below, was: fun and nonsense. Not judging anybody, I generated silly images of me with strange haircuts too. But each of these images has an energy and water cost that we never really get to see, and I think we would think about it for one second if we saw it. We do not see it because the product is built that way, since friction is bad for engagement and engagement is the business. 

Same for what Marc Zao-Sanders calls [thinkslop](https://marczaosanders.substack.com/p/thinkslop), which is when you hand over your own thinking to something that does it faster and worse. It’s probably too late to tell students to stop using it, this battle is lost and it would be a bit ridiculous coming from somebody who uses it all day. My problem is not with the student, it is with products where the easiest path is always the one where the machine thinks instead of you, and that is a choice, because a product that asks you to do the work gets used less.

And this is what I mean when I say people should have a choice, and I do not mean the choice of opening the app or not. Almost nobody chooses which system they use, a school installs one, an employer buys a licence, a ministry signs a framework agreement, and at the end the person has one option that somebody else chose, for reasons they will never see. The one who decides and the one who uses are never the same person here, and a system designed only for the second one never gets installed. So if the only thing in the tap is the bottled stuff, saying that drinking is voluntary does not mean much.

## The tap does not want you thirsty

There is a second difference between the bottling plant and the mains, and it is not the ownership, it is what each one wants from you.

A public authority decides what the pipe carries. The tap gives water and not Coca-Cola, and this is not a technical limitation, it is a decision about what a shared system is for. South Korea deciding that half of the system must run on domestic models is the same kind of decision. So is India with [Bhashini](https://bhashini.gov.in/?ref=julie.beliao.fr), a public language stack that runs across more than five hundred government sites and is built to work offline and in Indian languages, and not only in fluent English on a good connection. A public authority saying what goes into a public pipe. You can disagree with their answers and still think the question is the right one.

Look also at the interface. Water and electricity are behind a switch, it is there when you press it and it is silent the rest of the time. It does not send notifications and it does not miss you. You go to a utility when you need it, and it does not come to you. Products built on engagement work the other way, always on, always suggesting one more thing, because your attention is what they sell. A tap has no reason to make you thirsty.

I want to be careful here, because the easy explanation is that governments care about their citizens and companies do not, and I do not believe that. Governments are not reliably good and we should not build an argument on their virtue. The difference is more about arithmetic. A government that pays for a service out of a budget sees your consumption as a cost, so it wants you to get what you need and not more, and this is why water utilities run campaigns every summer asking people to use less of what they provide. A company that bills you for consumption sees your consumption as the goal. Neither of the two is virtuous, they are just pointed in different directions, and one of these directions happens to pass through what is good for you.

This is also, for me, the answer to the homework problem. If the easiest path in these products is the one that thinks for you, it is not because somebody wants students to stop thinking, it is because a product built for usage has no reason to hand the work back to you. An assistant that does not need you to stay can afford to ask you to do part of the work yourself.

## You can put tap water in a vial

Here is an objection I made to myself while thinking about AI as a public utility: public systems fail, and sometimes they fail badly. Flint in Michigan changed its water source in 2014 and put lead in the taps of a hundred thousand people, and the officials repeated for more than a year that the water was fine. Thames Water has been a permanent scandal for years, with sewage in the rivers. And if the public utility is my model, then its failures are part of my model too.

But look at how Flint was exposed. Residents who did not believe the reassurances collected samples, Marc Edwards and his team at Virginia Tech tested them, the paediatrician Mona Hanna-Attisha looked at the blood records of the children, and after that the numbers were hard to argue with. This was possible because tap water has a property we never think about, which is that anybody can put it in a vial and send it to a laboratory that does not work for the city. The system failed, but it could be caught, and for me these are two different things.

A closed model cannot be put in a vial. It changes without telling you, the version of yesterday is not available anymore, and an external lab cannot really test it, also because the terms of service usually forbid it. So when a closed system fails somebody, the only evidence is in the hands of the company that runs it.

This is why open source matters in this argument, and not as a detail. Open weights, open evaluation, published benchmarks, data flows that can be inspected, this is the vial. I am not saying that the public version will never fail. I am saying that when it fails, it can be shown, and for me this is worth more than any promise about good intentions.

## Water was not always public

Water used to be sold by the bucket, by private carriers, at prices the poor could not pay, and the cities took it over in the nineteenth century. They did not do it because the municipal water tasted better, they did it because of cholera and because of fire. The case for the utility was never the premium experience, it was that everybody needed a baseline, that the baseline was cheaper to provide together than separately, and that everything else got built on top of it.

The Romans had a term for this, res publica, the public thing, which is where the word republic comes from. The aqueduct is a good example, a thing built once, for everybody, that no single household could have built for itself.

Two things they had in the nineteenth century that we do not have: cholera, which forced the decision, and a legal monopoly with tax money once the decision was made. Our version is voluntary all the way down, so the product has to be actually good, and not only safer than the alternative. And this history is not finished, and it is not everybody's history. [One person in four in the world](https://www.who.int/news/item/26-08-2025-1-in-4-people-globally-still-lack-access-to-safe-drinking-water---who--unicef?ref=julie.beliao.fr) still has no safely managed drinking water at home, and where the pipe never arrived people buy water by the sachet and by the jerrycan, at a price per litre far above what a connected household pays. The tap is not the natural state of things. It was decided, city by city, and in most of the world it has not been decided yet.

## At the beginning nobody was in charge

The part of this history I find the most useful is that nobody really planned it from the top.

In Britain, the Health of Towns Association was formed in 1844, a voluntary group of doctors, engineers, statisticians and clergymen, with branches in the industrial cities. They produced evidence, they argued in public, and they dissolved around 1849 once the Public Health Act existed. In the United States, the American Water Works Association was founded in 1881 by the water people themselves, and it wrote the technical standards that the utilities, and later the regulators, adopted. The federal law on drinking water arrived only in 1974, so the body that defined what good water means held the field for ninety years before the law caught up. The municipalisation itself spread city by city, by imitation, because Birmingham in the 1870s worked and the numbers were public, so the other cities copied it. In France, many cities run the water themselves, in what we call a régie (the city operates the service directly), and the others delegate it to companies, and the two companies founded then, Compagnie Generale des Eaux in 1853 and Lyonnaise des Eaux in 1880, are today Veolia and Suez.

So there was an evidence body, a standards body, a demonstration city and private operators, and none of them was in charge, and all of them were necessary. It also shows what lasts, because each of these utilities was eventually absorbed or replaced, and the association that wrote the standards is still there. Standards for machine intelligence are being written today, at ISO, at NIST, in the European bodies working on the AI Act, and since 2025 at the United Nations, which now has an independent scientific panel on AI and a [global dialogue](https://www.un.org/global-dialogue-ai-governance/en?ref=julie.beliao.fr) that met in Geneva in July. What does not exist anywhere is the other half of what the water people did, which is independent measurement of the systems people actually use, published, and solid enough for a regulator to adopt later. That seat is empty, and filling it is arguably worth more than another model.

## Drinking and washing

Now, what is this utility for? Thinking about how a household actually uses water, there is the water you wash with, the showers, the laundry, the dishes and the floors, which is almost all of the volume, and where nobody would ever pay for premium. There is the water you drink, a few litres per day, where quality and trust matter, and the tap serves this too, which is exactly why tap water is held to a drinkable standard everywhere. And then there is the bottle, the same drink at several hundred times the price, that some people prefer for the taste or the brand. The same three levels exist, in my own reading, in machine intelligence, the bulk everyday work, the careful uses where quality really matters and where a public utility has to be good enough as well, and the premium branded product, which is allowed to exist but as a preference and not as the only safe option.

The categories below come from the[ Harvard Business Review work of Marc Zao-Sanders,](https://hbr.org/2025/04/how-people-are-really-using-gen-ai-in-2025?ref=julie.beliao.fr) which runs every year since 2024 and is built from what people say in public about what they actually do, and the shares are from the 2025 edition, because the 2026 one does not publish them in the same form. What the list does not say is which of these uses belong to the tap and which belong to the bottle.

| Category                                 | Share      | Tap or bottle                                       |
| ---------------------------------------- | ---------- | --------------------------------------------------- |
| Personal and professional support        | 30 percent | Tap, washing and drinking both                      |
| Content creation and editing             | 18 percent | Tap, mostly washing                                 |
| Learning and education                   | 16 percent | Tap, and it has to be drinkable here                |
| Technical assistance and troubleshooting | 16 percent | Tap for the everyday half, bottle for frontier code |
| Creativity and recreation                | 11 percent | Bottle                                              |
| Research, analysis and decision making   | 9 percent  | Tap, drinkable quality when the decision matters    |

The person who buys Evian for the table still fills the dishwasher from the tap, and would find it silly to do otherwise.

So my position is that we should go after the tap, all of it, the washing and the drinking, and be greedy about it instead of modest. It is the very big majority of what this technology is actually used for, and it is the part where the cost per answer decides everything and where the last five percent of quality decides almost nothing. Today it is served by products that are much too powerful for the job and priced accordingly, by companies whose interest is that you keep the tap running. Libraries, clinics, city offices and community organisations do this work every day, with budgets that do not survive a price decided somewhere else.

If somebody wants to read this as an admission that we will not win against the bottle, that is fine, because I am not trying to sell bottles. The bottle survives where the tap is not trusted, and my point is that the tap has to be drinkable, so that choosing the bottle stays a preference instead of being the only safe thing to do.

## Four places where my analogy breaks

I prefer to break my own analogy myself, before somebody else does it for me. There are four places where it does not hold, and for each one I turned the problem into something to ask for.

Water has almost no marginal cost once the pipes are laid, and here every answer costs money, forever. This is true and it is the strongest thing you can say against me. It is also why I do not think the meter itself is the problem. What I have a problem with in Altman's sentence is who holds the meter and what happens to the people who cannot pay it. So the demand is a published price per unit of work, low, and the same for everybody.

Water is undifferentiated and machine intelligence is not. Tap water that is twenty percent worse than bottled is invisible, and a model that is twenty percent worse is obvious in the first reply. So the demand is a quality bar, in the segment we chose, verified by somebody who is not us, which is simply the drinkability standard applied to us.

Water is local and here the competitor is one browser tab away. There is no natural monopoly protecting anybody, and honestly I find this healthy, because it means the argument has to be won on merit. The people who really cannot substitute are the ones blocked by cost, language, sovereignty or capacity, and they are the audience anyway.

Water quality is regulated and measured by law, and model quality is neither. So whoever builds a credible public measurement of models will decide what good means, and nobody has built it yet.

## Who pays

This is where public interest projects are usually the least honest. Every answer costs money, forever, and a service that is free to the person is never free to run. The first question any organisation asks is what it costs to serve its own people, and anybody who wants to build this has to be able to answer it before asking anybody to adopt it.

Which gives three ways of paying, and they should not be confused. The build is paid once. The consumption is paid continuously, by the organisation that deploys it, at a published price. And the person at the end pays the way people pay for water, through a bill, through taxes, or through a subsidy when they cannot pay the bill. Nothing here is free, the question is only who pays and how.

## The test

So when somebody uses the word utility when speaking about AI, including me, I propose four questions:

- Can you say where the data sits and who can read it.
- Can you say what a request costs, to you and to the planet, and do you show it to the person who makes the request.
- Can somebody who does not work for you put your system in a vial, test it and publish the result.
- And what happens to the person who cannot pay.

For the sentence that started this piece, the answer to the last question is nothing. The people who cannot pay are simply not in the future it describes, and for me this is the whole difference.