If you use AI every day, sell an AI product, or invest in AI companies — one fact unfolding right now will send the three of you toward very different fates: the price of intelligence is collapsing faster than almost anything in business history.

Start with the number. Per Stanford’s AI Index: intelligence at GPT-3.5 level cost $20.00 per million tokens in November 2022. By October 2024, the same level cost 7 cents. Eighteen months, a 280-fold drop. And it hasn’t stopped: depending on the task, LLM inference prices are falling 9x to 900x per year.

This is “intelligence deflation” — not a metaphor, but a price curve steeper than Moore’s Law. Most people haven’t thought through what it means. Three analogies will do it.

Analogy One: Long-Distance Calls — Deflation Isn’t Decline, It’s Ubiquity

In the nineties, an international call cost a small fortune per minute; people drafted notes before dialing. Today you video-chat with the other side of the planet for free, for hours.

“Calling” didn’t disappear — it became ubiquitous. But nobody makes money selling calls anymore.

Intelligence is walking the same road. So the first meaning of intelligence deflation is good news: intelligence is going from luxury good to utility — everyone can afford it. The real question is the second meaning: when something becomes affordable to everyone, what happens to the businesses that make money selling it?

Analogy Two: Recipes — Why This Price Collapse Can’t Be Stopped

Why has fine whisky stayed expensive for decades? Provenance, aging, terroir — it can’t be copied.

An AI model is more like a new dish: you invent it and dazzle the town; three months later the restaurant across the street (open-source models) serves something 80% as good for a fraction of the price. Models aren’t protected by patents, methods diffuse within months, and strong models can “teach” cheap small ones. Capability is a reproducible engineering artifact, not a non-renewable resource. You can’t charge an exclusivity premium, so you cut prices with everyone else.

This is the deepest difference from oil, land, and license businesses: those get scarcer as you extract them; intelligence gets cheaper as you make it.

Analogy Three: Electricity — Who Cries, Who Laughs

When electricity prices collapsed, ask who actually made money:

• People using electricity won big — that’s you and me and every business using AI; the harsher the deflation, the bigger the windfall;
• Appliance makers got rich — cheaper power meant people bought more appliances — that’s the companies embedding AI into products you can’t leave (workflows, entry points, subscriptions);
• Grids and fuel suppliers got paid first — usage exploded — that’s chips, data centers, and power, the upstream;
• Power plants had the worst seat: prices kept falling, so they had to generate ever more just to earn last year’s income — running up a descending escalator.

In the AI industry, the “power plant” is the model company.

Now the reversal that confuses everyone: if intelligence is deflating, why is AI company revenue exploding? Per multiple reports, the leading lab’s annualized revenue quintupled in half a year, doubling every six weeks. How do deflation and explosion coexist?

Because after a 280x price cut, usage grew far more than 280x — the oldest script in economics (Jevons paradox): the cheaper it gets, the more wildly it’s used. But “exploding revenue” and “making money” are two different things: every unit of intelligence sells for less and less, while the dominant cost — compute — gets paid upstream. Per Epoch AI’s breakdown, roughly seventy cents of every dollar a leading model company spends goes to compute. The revenue is theirs; the profit is the landlord’s.

So Where Does the Money Settle?

Fold the three analogies into one sentence:

Intelligence is deflating, but both ends of the pipe it flows through are getting more expensive.
Upstream (chips, power) earns the scarcity premium — the harsher the deflation, the bigger the usage, the scarcer the power and compute;
Downstream (products, entry points, workflows) earns the lock-in premium — however cheap intelligence gets, you can’t leave the product it’s wrapped in;
Only the middle — selling bare intelligence — is doomed to sell cheaper and cheaper.

When electricity got cheap, the money went to grids and appliances, not power plants. The revenue map and the profit map have never been the same map. Knowing whose hands cheap intelligence flows through is worth more than predicting intelligence itself.

What do you think — is the AI you use, the product you build, the company you back a “grid,” an “appliance,” or a “power plant”? Comments open.

Data sources and evidence levels (June 2026): GPT-3.5-level intelligence $20.00 → $0.07 per million tokens, ~280x over 18 months; inference prices falling 9–900x/year by task (Stanford AI Index 2025); leading lab revenue doubling every six weeks (multiple media reports); ~70% of spend going to compute (Epoch AI).

— A companion piece to a book in progress, working title The Deflation Sandwich · the prologue is in the previous post

#AI #IntelligenceDeflation #DeflationSandwich