The Carnival Before the Extraction - Part 2
Five months later, I was mostly right and a little bit wrong.
Last January, in The Carnival Before The Extraction (Part 1), I said this: the cheap, abundant, almost magical AI we’re all enjoying right now is a temporary carnival, not a stable state. Consumers are the beneficiary of a brief “capitalist subsidy” before the extraction phase comes due.
For those of us old enough to remember, I drew the parallel with the early Web of the late nineties. Open, exploratory, full of possibility. An era that started optimistically with “Information wants to be free” and ended with the current modus of the “Attention Economy” where “We are the Product”, where nothing is owned, and everything is milked “as a service.”
I argued in that Substack piece that AI is on the same trajectory. Only faster. I warned about “The Intelligence Divide” (one of many) that is coming. Between:
the AI Sovereign - entities with real models, real compute, real productivity, and
the AI Serf - the rest of us, with weaker AI versions, and monetized for attention and data.
That was mid January. It is now late May. It’s time for an update.
The 20$ per month was always a fiction
Around $20 (or euros) per month or so - the monthly price of Netflix or Spotify - it’s the psychological number that today’s consumers hold as “things I can pay monthly without thinking too much about it.”
When OpenAI launched ChatGPT Plus, their finance team did not run the unit economics. They just borrowed a number that consumers had already been trained to accept, a price completely disconnected from the actual cost. And there lies the problem.
To be clear, there has been a significant improvement in the inference cost of AI. Two years ago, top models cost $30 per million tokens. Now, they cost $2.50 to $5. So yes, AI is getting cheaper.
But the math still doesn’t square (and cannot square).
The mathematic reality behind the bubbly exuberance
In 2026 alone, the industry will spend close to $700 billion on capital expenditure. Data centers. Chips. Power. Talent. OpenAI itself has pledged to invest approximately $600 billion into AI and computing infrastructure by 2030.
But, of the alleged 900 million weekly users of ChatGPT, only 5.5% pay anything. Which means that every paying user of ChatGPT is subsidizing roughly 18 free riders.
In February, OpenAI started injecting ads into the free tier of ChatGPT. The same Sam Altman who once called the idea of mixing ads with AI “uniquely unsettling” reversed course.
They tried to launch an age-verified “Adult Mode” to find new revenue streams, but had to quietly shelve it after a huge pushback.
They introduced a Pro tier at $200 a month because $20 wasn’t pulling its weight.
They added a hidden surcharge on long-context queries because power users were burning their compute budget.
All this because OpenAI is on track to lose somewhere between $14 billion and $36 billion this year. In the first quarter alone, their operating margin was negative 122%. For every dollar of revenue, they lost an additional $1.22.
The simple math is that you can’t amortize a $600 billion build-out with $20 consumer subscriptions, no matter how cheap inference becomes.
So they had to invent the divide
Basic business 101 gives us two choices here: raise the price, or change the customer (i.e. find a new market). OpenAI (and others, not just them!) chose to do both. And it’s clear which one is the real strategy.
The new and real market is the enterprise customer. The bank, the law firm, the developer team, the Fortune 500 IT department. These are not people paying $20 a month. These are workloads. Autonomous agents running 24 hours a day, generating tens of thousands of dollars in monthly billing.
Anthropic just turned an operating profit on roughly this customer alone. Their revenue went from 1 billion-dollar run rate at the start of last year to something approaching 45 billion now. Not because consumers fell in love with Claude. Because enterprises figured out that one autonomous agent costs less than one engineer.
Microsoft - a 3 trillion-dollar company - recently forced its own engineers off Claude Code because the token burn was blowing up internal budgets. And that tells us everything about the coming divide.
If even Microsoft can’t comfortably afford frontier autonomous AI for its own staff, the rest of the market is not going to get the real thing at consumer prices.
The AI Sovereign and the AI Serfs
The rest of us, the AI serfs, the free tier. The 10-dollar tier. The 20-dollar tier, all of us will quietly get capped, throttled, surcharged, and eventually advertised at.
We will get a version of AI. It will be useful enough to be sticky. It will be cheap enough to be popular. It will help us write our emails, summarize our reading, draft our messages, fool around a bit. It will be enough to feel like we have AI. It will NOT be enough to compete with anyone who has the real thing.
That version of AI - the one that deeply reasons, the one that runs autonomously, the one that compounds - will be afforded only by the AI Sovereign. It’s the math of AI economics today, the business model of the industry.
The question is: is that the business model for our society?
What I underestimated
In January, I described the Intelligence Divide as something that was coming. A risk we still had time to prevent. Open-source. Data dignity. Public AI infrastructure. Cognitive independence. All still worth fighting for.
But I should have been sharper about the timeline. The divide is not coming. The divide is already the only path by which the industry can survive its own capex.
When the unit economics cannot support the consumer at the price the consumer expects, the consumer is no longer the customer. The consumer is the product (once again!). And alas, we have seen this movie before.
The AI carnival is still loud. The lights are still on. There is still real wonder in what these AI tools can do.
But the gates have already been built.
We are just still inside them, paying our 20 dollars, or enjoying our free tier, not yet noticing that the door behind us is closed.
This is part 2 of an essay first published on my Substack in January 2026. You can read the original - The Carnival Before The Extraction - at ericraza.substack.com.