In February 2018, I sat in a Cape Town co-working space reviewing a token whitepaper for the fortieth time. The project promised "democratized access to computing power." Its pricing model was elegant on the slide and predatory in the footnotes: the token was the only way to pay, the supply was controlled by insiders, and the "democratization" was a subsidy that would vanish the moment the treasury emptied. I wrote a column warning that access without durable economics is not democratization — it is a promotional rate. The backlash was immediate and personal, and it cost me a few friendships I did not mourn.
Eight years later, the same story is being retold in a different language. Frontier AI laboratories are rethinking their pricing models as budgets tighten. The headlines frame this as a strategic pivot. I read it as something older: a maturity event, and a warning. Hype burns out; robustness remains in the ledger.
Let me lay out what is actually happening, stripped of the press-release gloss. For three years, the dominant pricing paradigm for large language models has been metered inference: you pay per token, in and out, at a rate set by the provider. This model worked while model capability was scarce and inference was expensive. It no longer does.
Two forces have converged. The first is cost. Engineering optimizations — mixture-of-experts architectures that activate only a fraction of parameters per token, quantization down to FP8 and INT4, speculative decoding, prefix caching, and continuous batching — have driven the unit cost of inference down by roughly an order of magnitude per year. None of these are exotic. They are the unglamorous work of systems engineering, and they compound quietly while everyone watches the model releases.
The second is commoditization. Open-weight models now approach closed frontier performance on most practical tasks, and they cost almost nothing to serve at the margin. When a capable model is a download rather than a subscription, the scarcity that justified premium pricing evaporates.
When a scarce good becomes abundant, the price you can charge for the good collapses. The laboratories are not rethinking pricing out of philosophy. They are rethinking it because the meter no longer measures scarcity.
There is a third force that the coverage names but does not examine: budgets. Enterprise AI spending has entered a return-on-investment reckoning. The era of experimental procurement — buy seats, see what happens — is ending. Buyers now ask what a deployment returns, not what it promises. And capital markets, which funded enormous infrastructure buildouts on the assumption of explosive revenue, are growing less patient with the gap between spending and income.
I watched this exact sequence play out in a different market. In 2017, capital flooded into token projects on the promise of utility, with no mechanism to verify that the utility would ever arrive. The correction was brutal, and it was clarifying. The AI market is now approaching its own version of that correction, except the underlying technology is real and the correction will be about margins rather than existence.
I spent 200 hours in 2020 auditing a governance mechanism for a decentralized finance protocol. The lesson that stayed with me was not about voting weights. It was about where value accrues when coordination becomes cheap. That lesson applies directly here.
The pricing shift is not one shift. It is three, and conflating them is the central error in the current discourse.
First, the mid-tier layer is collapsing in price. Where a general-purpose model once commanded a premium, near-equivalent open-weight alternatives are available at a fraction of the cost. This is genuine deflation, and it is healthy — it is the mechanism by which access widens. A startup in Lagos or Jakarta can now build on capability that cost six figures to access two years ago.
Second, the frontier reasoning layer is becoming more expensive, not less. Models that spend large amounts of test-time compute to reason through hard problems consume far more inference per query. The compute is the product. These models will not follow the deflationary curve, because their cost structure is fundamentally different. A reasoning model that deliberates for thirty seconds is not a cheaper product than a model that answers instantly; it is a more expensive one, and it should be priced accordingly.
Third, the enterprise layer is moving away from token metering entirely, toward outcome-based and seat-based pricing. This is where the real restructuring is happening, and it is the least discussed. When a buyer pays per resolved ticket rather than per token, the vendor absorbs the variance in cost and the buyer absorbs only the outcome. That is a fundamentally different risk allocation, and it is a sign that the market is maturing past the metered-utility model.
The mistake is to speak of "AI pricing" as a single thing. There is no single AI price. There is a stratified market, and the strata are moving in opposite directions.
This is precisely the pattern I documented during the 2017 token boom. Projects claimed a single, unifying "token" would price all activity — compute, storage, governance, attention. In practice, the moment a market matures, it stratifies. A single price for everything is a symptom of immaturity, not sophistication. The AI industry is now discovering what the crypto industry learned at a cost: pricing power is not a function of the narrative you attach to your product; it is a function of the cost structure underneath it.
Consider the arithmetic. If the unit cost of inference falls by a factor of ten but usage grows by a factor of fifteen, total revenue from the same customer base rises while price per unit collapses. This is the Jevons paradox, and it is the quiet engine beneath the current repricing. Lower prices do not shrink the market; they expand it, shifting value from the meter to the volume. It is why the collapse in inference cost has not collapsed demand for compute — it has inflated it, because every application that was previously uneconomical becomes viable at a tenth of the cost.
Consider what this means for the infrastructure layer. The same deflation that pressures model pricing expands the demand for inference itself. Every price cut converts a marginal use case into a viable one, and every viable use case consumes compute. The chipmakers and cloud providers are not the losers in this repricing; they are the beneficiaries of the volume it unlocks. The losers are the intermediaries who assumed the meter was the moat.
But there is a harder arithmetic that the optimistic reading skips. If usage grows by a factor of five instead of fifteen, total revenue falls even as your infrastructure costs rise. The repricing is not automatically healthy. It is healthy only if the elasticity of demand exceeds the rate of deflation. No headline captures this conditional. No press release will.
I learned to distrust conditional-free claims during the ICO review season, when I read over forty whitepapers and found predatory tokenomics in roughly a third of them. The tell was always the same: a projection of explosive adoption with no elasticity analysis behind it. The AI repricing narrative has the same tell. It assumes that cheaper access produces proportionally more usage, without demonstrating it. The assumption may be correct. But an assumption stated as a certainty is a warning sign in any market, and I have learned to read it as one.
Here is the insight I would offer to anyone trying to read this moment clearly. The question is not whether AI gets cheaper. It is who captures the surplus when it does. In a stratified market, value migrates to the two ends. Downstream, application builders and their customers capture the benefit of cheaper capability. Upstream, compute suppliers capture the benefit of higher volume. The middle — the pure model layer with no differentiation — is where the surplus is squeezed.
That is why the language of "frontier labs" as a collective is analytically useless. The category conceals a brutal internal divergence. A laboratory with proprietary silicon, exceptional inference efficiency, or a genuine capability gap can sustain premium pricing. A laboratory selling undifferentiated API access to a commoditized model is selling a product whose price is being set by its cheapest competitor. These are not the same business, and they will not survive the repricing the same way. The cost leaders will set the price. Everyone else will accept it.
The deeper anxiety, the one the headlines avoid, is that the repricing tests the investment thesis of the entire sector. Enormous capital has been committed to infrastructure on the assumption that revenue would scale to meet it. If prices fall faster than usage rises, that assumption breaks, and the correction is not a pricing adjustment — it is a valuation one. I have watched this movie before, in a market that had far less substance underneath it, and the ending was not gentle.
This is the same dynamic that separated the decentralized finance protocols that survived 2020 from those that did not. The survivors had a structural reason to exist — a cost advantage, a network effect, a genuine lock-in — and the others had a narrative. Narrative does not survive a bear market. Neither does a pricing page built on one.
The consensus reading of this moment is that AI is becoming democratized, and that this is unambiguously good. I want to test that claim against the pragmatism of the ledger, because I have seen the word "democratization" used to sell things that were neither democratic nor durable.
Democratization is not a pricing decision. It is a structural condition. Access becomes genuinely democratic when no single party can revoke it, when the cost of participation is governed by open standards rather than the goodwill of a vendor, and when the underlying capability cannot be withdrawn by a change in corporate strategy. A promotional rate is not democratization. A temporary price cut financed by venture capital is not democratization. They are marketing, and marketing burns out.
Open source is a covenant, not just a license. The open-weight model ecosystem is the closest thing to genuine democratization in this market, precisely because the weights cannot be recalled. When you can run a capable model on your own hardware, the vendor's pricing page stops being a source of authority. That is structural. Everything else is a discount with an expiry date. Faith in people is costly; faith in math is free.
The second contrarian point is subtler. Lower prices lower the barrier to use, but they also lower the barrier to misuse. In my work on the Verifiable Human Standard — a framework I helped draft to preserve provenance and authenticity as synthetic media proliferated — the recurring tension was never about capability. It was about accountability. When access is nearly free, the cost of abuse approaches zero, and the mechanisms that distinguish a legitimate participant from a malicious one become the scarce resource. The pricing debate is silent on this, and it should not be. A market that celebrates cheaper access without pricing the externalities of that access is not mature. It is merely cheap.
I do not think the repricing is a crisis. I think it is a reckoning, and reckoning is how markets mature. The laboratories that survive will not be the ones with the loudest narrative about access. They will be the ones whose cost structures can withstand the deflation they themselves triggered.
We audit the logic, for humans will always err. The meter will keep running. The question that matters is not what a token of intelligence costs today, but whether the ledger underneath it is honest enough to still be standing when the hype has gone.