The Collapsing Cost of Intelligence

Until recently, the usage rule in artificial intelligence was simple. The best models cost the most, and the most expensive were the scarcest. If you wanted to look cutting-edge, you paid up and left the cost-benefit question for later. 

Silicon Valley coined a word for it: “tokenmaxxing,” and it applied at the individual and the enterprise level alike. It meant burning tokens, the billing units of AI, with little regard for whether the spend was justified. Pushing semi-blindly ahead was the smart strategy, and the market punished anyone who would not play. 

This phase is ending. The cost of intelligence is falling fast, both from competitive pressure and because the models themselves keep improving. The speed and scale of the change are enough to reorder who wins and who loses. 

It is worth working through what that means — who it helps, and who it threatens. 

How Far, How Fast 

The price of capable intelligence has fallen more than ten times over, and the clearest proof came from outside the American labs. 

As seen in The Decoder’s article by Matthias Bastian, in May 2026, DeepSeek, a Chinese developer, cut the price of its flagship model and made the cut permanent¹ . That put its top model at a fraction of what the leading American models cost, while it performed nearly as well on standard tests of coding and math. 

What matters is not the price but the reason for it. DeepSeek did not run a promotion; it made the model cheaper to run, on a fraction of the compute its predecessor needed, and passed the saving along. A discount can be waited out. A lower cost is a new floor. 

DeepSeek was not alone. The major providers cut prices tier after tier through the same stretch, and the direction of the cost of intelligence has been steeply, repeatedly down. 

The Pressure Reaches the Top 

The pressure has now reached the premium sellers themselves. 

As reported by The Wall Street Journal’s authors Keach Hagey and Berber Jin in June, OpenAI was weighing significant cuts to what it charges for tokens. It expected its closest rival to cut too. The talks were unsettled, and no decision had been made — but the direction is the point. Its own chief executive has called AI costs a huge issue for customers, and some have burned through a year’s budget in a few months.²

This is what the end of scarcity looks like from the inside. The most prominent seller of premium intelligence is being pulled toward the prices of the challengers beneath it. A premium depends on staying rare, and cheap, nearly-as-good alternatives erode it. 

The chip industry already taught us the lesson. The companies that lasted were not always the ones with the best raw performance; they were the ones who built defensible moats around their lead. A model that is merely the smartest this quarter has no such moat — and it is exposed the moment something close enough ships for a fraction of the price. 

Part of how the lead leaks downhill is a technique called distillation: training a cheaper model on the outputs of a more capable one. Done to a lab’s own models it is routine — the whole industry uses it to make smaller, cheaper versions. Done to a competitor’s model at scale, it is contested. 

As stated by CNBC’s Matthew Chin, through 2026 the leading American labs have publicly accused Chinese developers, including DeepSeek, of extracting capability this way, and the accusations have widened and sharpened over the year. The legal ground is genuinely untested, but likely to be challenged.³  These components were also mentioned in my blog dating back to February 24, 2025 called US Exceptionalism? Time to Ditch the Training Wheels.⁴

The point for the cost of intelligence is simpler. Whether by original engineering or by learning from the leaders, capability is reaching the cheap tier faster than the frontier can hold it, and the buyers doing the routing rarely care. 

Who Gains and Who Loses 

Cheaper intelligence is a gift to the companies that buy it and a threat to the ones that sell it at a premium. 

Businesses that consume intelligence benefit: software companies, cloud platforms, banks, anyone who buys AI as an ingredient and sells something built on top of it. When the ingredient gets cheaper, their costs fall and their products improve at once. And paradoxically, while cost per unit may decline, total spending on AI may hold or even rise. 

On the other side are the labs whose whole business is selling the most advanced model at a premium. As capability becomes abundant and its price falls, the advantage shifts from the sellers of scarce intelligence to its users. Whether usage will accelerate fast enough to offset that shift is a real debate. 

There is discipline in being a good user. Not every job needs the most powerful model, any more than every computing task needs the top-end chip. The skill is routing each task to the right tool, saving the expensive model for where it earns its price. 

Reaching for the priciest option by reflex was the condition of the maxxing phase; now it is just profligacy. 

The Paradox Underneath 

Falling costs do more than shift advantage. It pulls adoption forward. 

When intelligence was scarce, only the obvious jobs cleared the hurdle. As the cost of each unit collapses, tasks that did not pencil out a year ago suddenly do, and adoption broadens across the economy sooner than it otherwise would. 

This has a name, and I have invoked it before: the Jevons paradox, which I also described in my blog US Exceptionalism? Time to Ditch the Training Wheels.4 Make a resource cheaper to use and total consumption tends to rise, not fall, because the lower price expands demand faster than efficiency cuts it. 

It reconciles the two facts in this note that seem to fight. Cost is collapsing, yet AI bills are exploding, because intelligence got cheap enough that the world began using far more of it. Falling unit cost is the cause; rising total spend is the effect. 

That growing pool of spend is the prize, and it accrues unevenly. The companies quickest to put cheap intelligence to work pull ahead, the disciplined ones fastest, while the rest arrive late and some arrive badly. 

It also softens the threat to the premium sellers. If total demand is expanding, volume can offset some of the pressure on price, so what they face is a compression of margin and pricing power rather than the simple loss of a market. The chip industry shows the shape of it. The dominant designer still leads the field even as more buyers make their own silicon for the jobs that do not need the best — losing select battles without losing the war. 

The premium model tier is no different. It does not vanish, and the hardest work will still demand the top model and pay for it. But that band is smaller and more contested than the spending phase assumed, when the reflex was simply to buy the best. 

For most of this cycle, the market cheered the sellers of scarce intelligence and treated cheapness as a threat. The market had it backwards. The only question left is how long the sellers can resist gravity. 

Footnotes 

¹ Matthias Bastian, The Decoder, “DeepSeek Makes Its 75% Discount Permanent,” May 23, 2026.
² Keach Hagey and Berber Jin, The Wall Street Journal, “OpenAI Considers Drastic Price Cuts, Anticipating War for Users With Anthropic”,” June 10, 2026.
³ Matthew Chin, CNBC, “Anthropic joined OpenAI in flagging ‘industrial-scale’ distillation campaigns by Chinese AI firms,” February 24, 2026.
⁴ Mark Scalzo, “US Exceptionalism? Time to Ditch the Training Wheels“, validex.co, February 24, 2025.