There is a metaphor going around in various forms in posts about artificial intelligence: AI will become like electricity. We will pay for it by consumption. It will be a utility, one of those infrastructures under which daily life flows and of which we know essentially nothing, except for checking the bill once a month.
The metaphor is largely apt and captures the macro trajectory honestly. The problem is that it breaks down at two critical points, and this entirely shifts the question we should be asking now.
What the standard narrative has gotten right
Before the critique, a fair acknowledgement. The metaphor holds at the macro level on three points:
- First: Artificial intelligence is becoming an infrastructural layer, beneath next-generation applications. Editors, browsers, chats, business software, devices. Almost everything. Like electricity a century ago, AI is being layered under the trades, and within a few years it will no longer be a separate product. It will be a cognitive substrate that powers every object we use.
- Second: We will pay by consumption, largely so. There will be flat tariffs for certain categories, exactly as with electricity. But only for those who will have earned that position of standing, enough to negotiate as peers with the major provider. Underneath the flat there will always be real, metered consumption. Tokens per second, context windows, model calls, reasoning levels enabled. The provider knows it, and bills it, even when the user does not notice.
- Third: The margin of average operators will tend to drop over time. Competition between providers, the arrival of open source alternatives, regulatory pressure, the consolidation of architectures all push the price of the base unit toward a stable level over the medium term. Wealth shifts toward those who control complementary assets (data, patents, brands, distribution, specialised hardware).
So far the parallel holds almost perfectly. It is a useful map for describing the general trajectory, and it is the map that public debate has given itself in order to talk about AI without complete hysteria. Fine.
I swear, last preamble, then we get concrete.
In the previous essay I argued that the divide of the next decade would not be limited to the economic middle class and to the resulting irrelevance about to overtake us. There is a level beneath, more subtle, that manifests itself in parallel, on a plane different from the economic one: the existential plane.
There will be no economic middle class anymore, and that was already visible in the Nineties. Above all, however, there will be no existential middle class either. The split that runs along the axis of intelligence-on-tap is no longer played out on incomes. It is played on the median position between those who articulate their own direction and those who are articulated by others. And on that axis, a median position no longer effectively exists.
Where the metaphor shatters
Electricity has a feature that the standard narrative keeps overlooking: you can self-host it. Solar panels on the roof, batteries in the basement, a small wind turbine in the garden, and in many contexts, especially single houses, you have a way out.
For many, the self-host route costs more than the grid, yet it remains possible. It exists as a mental fallback option, even when not used.
With frontier artificial intelligence models, this way out is not there. Let me be precise, because it confuses easily. I am not talking about lightweight models you can run on a MacBook or on a small home server for the price of a moped. Those exist, will do useful things, and for many use cases will be sufficient to plug holes. I am talking about frontier models. The ones that today carry the best quality available on the market. The ones that, if you have to write a structured essay, do a serious analysis, generate complex code that works without spending days hand-polishing it, you will be forced to use. Those models, you do not build artisanally at home.
The reason is that frontier models require, to be trained, infrastructures costing billions of dollars, months of machine time on tens of thousands of specialised GPUs, enormous datasets, dedicated research teams. And they require, to be run in inference at full quality, dedicated clusters of specialised hardware and an infrastructure-management chain that does not replicate on a single desktop machine.
You can try to compress and distil them into smaller versions. People are already doing it. But a distillation always loses something, and that something is often the thing you need: the depth of reasoning, the useful length of context, the robustness in edge cases.
The practical result, on the strategic axis, is this. With electricity, if tomorrow the provider raises prices, you have mitigation options ranging from contract renegotiation to switching to another operator to self-hosting. With frontier models, if tomorrow the provider raises prices or changes policies, the options are tragically narrower. You can switch provider, but the alternative provider has different characteristics and the integration is certainly not free. You can fall back to an open source model, but you take a quality hit. Self-hosting state-of-the-art? Out of the question for us mere mortals.
Brutally translated: it is not a simple commodity. It is a damn dependency.
There is also a more radical consequence, which deserves a slightly old-fashioned formulation. For those who have read a bit of last-century philosophy, the discourse will not sound new. Those who own the means of production also own the freedom to do the things those means make possible. Move the formula from the industry of a century ago to today's craft. Those who own the means of production of intelligence, also own the freedom of thought, action, and reasoning at the levels that matter.
The analogy has very concrete consequences and does not stay on the abstract plane. If frontier models do not self-host, and if the quality of thought they generate is significantly above the alternatives, then the freedom to think at the best available level passes through whoever controls access to the models. Hierarchy does not disappear when intelligence is downgraded to a commodity. It reorganises, and it plays out on an entirely new axis.
The window that is closing
Let us now add a second, temporal level, that the standard narrative around electricity does not even reach.
Right now, in this historical moment, the prices of access to frontier models for power users are enormously below their real cost and value. This is not, of course, an act of generosity from the provider. We are benefiting from a strategic market calculation. Providers calibrated prices on the average customer: a person who uses their AI model to write emails, do searches, summarise documents, generate the occasional image. That average customer, even paying a monthly subscription, consumes a modest fraction of the capacity guaranteed on paper. The provider therefore budgets that average and uses it as the base for the calculation.
What the provider has underestimated is the lateral category of power users. People, and in some cases small organisations, that use the models at intensities twenty, fifty, sometimes a hundred times the average. Personally, if I do not manage to exhaust my weekly quota on a Max account I feel a sharp pain in my chest, as if I were giving the dish from a starred restaurant to pigeons. So for those in this category, the current plans represent a structural discount with respect to the value generated and the actual cost of that consumption for the provider.
This window cannot stay open indefinitely, of course. Compute is not free and the provider has to cover its own cost. The investors, then, who are currently the ones we should thank more than anyone else for this welcome gift, will sooner or later ask for returns.
When tariff policies are updated, the price for the power user will rise, probably by a lot. The window will close in six, twelve, twenty-four months depending on the segment. Without claiming precision, this is an orienting estimate of mine that exists as an order of magnitude in the books of those who run these markets. And it serves to put a bit of pepper on your couch to make you move your backside, in case you still think there is time to waste before the train closes its doors and goodbye to those who were wasting time at the station bar or humouring the chatter of mainstream pundits.
Lateral pressures
Below this mechanism two lateral forces are also moving, worth naming because they shorten the window further.
The first is the open source pressure from China. DeepSeek, January 2025, opened high-quality models at marginal costs compared to U.S. providers. "Oh how generous these Easterners are, we should all take notice." Sure, and I'm Santa Claus.
The move is to be read as government strategy: democratising AI globally means undermining the American advantage, positioning Chinese providers as accessible alternatives, replicating at global scale a scheme China has practiced for decades in local markets with destructive pricing.
The second is Meta's move. Yann LeCun, head of AI research at Meta, has codified a different strategy. When you realise the race for the best model is one you are not winning, you give the model away and turn the model itself into a commodity. You win with what you have more of than anyone else: data. Meta has the most extensive consumer data in the world. Free model plus Meta data, over the long run, is a permanent competitive advantage on the real asset.
Add to this the bootstrap of Mistral, the pre-emptive validations from U.S. companies with their positioning papers, and the picture composes itself. Under the same label coexist at least four different strategies, all converging on a single point: compressing the price of frontier models. And therefore shortening the pricing-error window for whoever pays.
The point of the essay is not these strategies in themselves. The point is their effect. The window exists, is utterly transitory, and inside it everyone is choosing what to do with it, even when they do not realise it. Especially when they pretend not to realise it by burying their heads in the sand. And guess what, do you know what stays nicely exposed to whoever passes when you take the ostrich position?
The Chinese copy at one per cent
Now let us deal with this other pebble in the shoe, because it is the same story every time. "It's the deal of the decade," they say. Chinese AI models at a fraction of the price of the U.S. frontier. Literally on the order of one per cent. "Let's dive in head first without even looking!" Geniuses.
The answer here lies not in ideology but in practical operativity, and it is worth saying in full as it should be said, because outside of AI you also pull this same nonsense, I know you all.
For those who operate where errors cost dearly, the copy is not fungible. The difference between a model at ninety-nine point five per cent reliability and one at ninety-eight per cent looks negligible on paper. On work where errors cost significantly (production code, financial analyses, deliverables that go out to thousands of people, decisions that influence several others downstream), an error every fifty responses instead of every two hundred is an abyssal difference. It is the distance between a system that turns your day around and one that costs you nights of cursing. The proverbial axe that punishes shortcut-takers, scaled to billions of parameters.
For that category of work, the actual cost of the Chinese model goes absolutely beyond the one per cent on the label. On top of the bare price you stack: the cost of the time to spot the errors (if you are not careless), the cost of the risk of the errors you do not catch (if you do not want to spend your life babysitting your machine), and, not least, the cognitive cost of the loss of trust in the system you use. There you have the recipe for the cursing nights mentioned above.
Not everything is to be thrown out, of course. For work where errors do not cost dearly, the situation reverses. For personal entertainment, casual exploration, experiments without consequences, the one-per-cent copy can be acceptable. But at that point, you've gone halfway, you might as well cross the line.
The split between who pays and who does not, then, does not distribute by company size or by wealth. It distributes by intensity of use and by cost of error. A single professional working at a serious level pays the full tariff. A large company that uses AI for secondary things can afford the copy. The hierarchy that the market hid under flat prices, under the pressure of the window, manifests itself here in its full, dramatic splendour. Whoever stays standing is whoever has understood which side of the axis they operate on.
The bubble that does not stand on the paying user base
We come now to talk about money, which I know is what everyone is waiting for.
Over the last two years, capital on the order of trillions of dollars has moved on AI. Investments in datacenters, in chip supply, in new companies, in acquisitions, in network infrastructure, in energy sources to power the datacenters. Numbers that look gigantic outside that market, and that are treated like peanuts at the circus inside it.
The question to ask, with no diplomatic varnish, is then: is this mountain of capital being repaid by those who pay for AI services today?
The honest answer: no. Not even remotely, and certainly not at the required pace. Most end users pay modest tariffs, a small portion pays nothing, and a marginal slice of power users pays subscriptions which, taken individually, do not cover their actual consumption. The numbers cannot lie.
It is the most classic of emerging markets financed by venture capital. The financing covers the gap between cost of service and revenue, while waiting for three things. That the price of chips drops structurally, which is happening. That users grow by an order of magnitude, which is happening, but more slowly. That some operators exit the market, consolidating revenue on the survivors, which is still in a sort of Mexican standoff.
While the rebalancing happens, there are two simultaneous risks. The first: that capital's patience runs out before consolidation, and we enter a phase of budget compression. The second: that the rebalancing happens but leads to such a sharp change in prices and policies that whoever has not built structural infrastructure during the window remains shut out of the AI-enabled labour market.
Calling this phase a bubble may be premature. Calling it a financial window of limited duration, no. That is exactly what it is.
What follows from this
Let us pick up the thread of the previous essay. I had described the divide of the next decade as the distinction between those who know how to align with the machine and give it direction, and those who do not and will have to trust the direction other humans have already given it.
Now let us add the operational level. The machine you align yourself with has the nature of a dependency, and that changes everything. The window in which this dependency costs little to power users is temporary. The difference between those who use this window to accumulate and those who use it to consume will define the position of the next ten years more than the study of AI itself will define it.
And beneath this difference, there is an axis that public discourse still seems afraid to say out loud.
The complete AI infrastructure does not exhaust itself in hardware plus software. It includes processes. It includes human capabilities. It includes above all the quality of the inner direction with which we use the models.
This is the piece that does not appear on company decks, and it makes a bigger difference than all the others. Without articulated inner direction, even with the best complementary assets, you move inside the frame of choices made by whoever sells you the model. With articulated inner direction, even with modest assets, you build your own orbit around those very same models.
A note I leave here because it will serve those who read this text six months from now.
Three years ago, I had set my mind on building a certain system, and it had been quoted at a cost of approximately one hundred and fifty thousand euros. That quote was not absurd. It was correct for the moment.
A year ago, with quite a few of the proverbial cursing nights, I put together by hand a more complex pipeline than that system, in three months of intense work.
Today (May '26), in ten days, I built a system that subsumes that pipeline and adds nearly quadruple functionalities.
Three figures, in thirty-six months. 2023: one hundred and fifty thousand euros quoted (design excluded). 2025: three months of work. 2026: ten days of work with four times the functionalities.
I report it as an operational example of what the window does to those in a position to accumulate. Twelve months from now even this text you are reading will be a piece of archaeology. I declare this now, under timestamp, wayback-machine-proof, because the meaning of what I am writing measures itself only inside the window in which I write it. This is why the window is a window, and not a plain.
Whoever has not gone through the past twenty-four months accumulating the conditions to produce a trajectory like this, twelve months from now will not be able to replicate it. Not because access will be denied to them. Because the price of access, and the learning curve to use it, will have changed. The window does not close in the same way for everyone. It closes earlier for whoever arrived without having built the conditions in the right months.
Three practical things
Three concrete pieces of advice, for whoever is reading, that I have codified personally and that work for others too.
One. Do not buy hardware in-house for the frontier. It is not worth it. The frontier travels on capital scales the individual cannot chase. Buy access, at the best quality you can afford, while the window is open.
Two. Build assets that remain yours regardless of the provider. A personal knowledge base structured, readable by the models, but yours. A brand under which you publish. A network of people who recognise you. Anything that retains residual value even if tomorrow the price of the model doubles.
Three. Identify two or three things you can do with these systems above the average, and specialise on those. No need to be good at every nuance of the universe. You need to be good enough at something to make the current window, for you, a window of accumulation rather than consumption.
It is not catchy advice. It is not viral. It is, however, decidedly more useful, on a horizon of the next two years, than most of what you will read out there on these topics from those who talk a lot and build little.
Who counts the electric sheep?
And now the title.
Philip K. Dick, in 1968, framed it as a question about androids. Do Androids Dream of Electric Sheep? It was a book about what it means to be alive when appearance betrays reality, and about the difficulty of telling the human apart from its perfect imitation.
Almost sixty years later, this text's question is less romantic. It does not concern androids. It concerns you, reading this piece (written entirely by hand, the old-fashioned way, as you may have noticed while reading).
The electric sheep of the title are the units of intelligence we consume every day. The ones you reason with, the ones you write with, the ones you listen to, the ones you watch on devices, or the ones you watch as interfaces deciding what to show you next. Someone is counting them. Someone has decided to make you pay for them. And they are charging not in proportion to the value those units generate for you, but in proportion to a commercial equilibrium that is rebalancing itself in real time, far from your window of attention.
The question becomes: who is counting your electric sheep?
If the answer is you, in a way articulated enough to accumulate something of yours in the window that is still open, you will be in the category that stays standing even when tariffs double or go ten-X.
If the answer is someone else, the sheep are being counted for you by someone else. When the price changes, you will realise it when it is already late, not because anyone will have warned you. You will realise it one morning, opening a bill, and finding a different number from before. Or a popup will ask you to buy new credits, because the usual ones do not cover it anymore. It will not be a disaster.
It will only be a life slightly tighter, slightly more decided from outside, by others, slightly less yours.
"Rented thought, metered intelligence. On the closing window, and on the existential middle class that won't be there." That was what it meant.
The debate is concentrating on how much we will pay for the cognitive kilowatt-hour in 2030. It is a legitimate, important question, but a secondary one. The real question is what we will have built on top of the access we are paying today far below its value. Whoever has built something, after the window, will stay in the game. Whoever has only consumed, will exit with the nostalgia of a season that has closed, and with the same AI on the table, but at a price that will force them to change their habits.
It is not a forecast, nor an assumption. It is geometry. It is already in plain sight for whoever looks at the slope of costs on one side and that of revenues on the other. Only the closing is missing, and its timing.
And in the meantime, someone is counting.
V.