Why AI Is Getting Cheaper and Your AI Bill Keeps Climbing

The Jevons Paradox, and the one thing about AI that makes it different from every piece of software that came before it.

Somewhere in your finance team’s inbox this quarter is a chart that shouldn’t make sense. The per-token cost of running a frontier AI model has collapsed — by some estimates, roughly 280-fold between 2023 and 2025, with base inference fees still dropping 60–70% a year. By any normal logic, that should mean your AI bill is shrinking too.

It isn’t. Enterprise AI spending is on track to hit somewhere in the range of $2.5 trillion globally in 2026, and it’s still climbing. Something is producing this contradiction, and understanding it is now a genuine strategic question, not a finance department curiosity.

The something has a name: the Jevons Paradox. And once you see the actual mechanism driving it in AI’s case, it points to something that’s easy to miss if you’re still thinking about AI the way you thought about software.

A 19th-century economist explains your 2026 bill

In 1865, the English economist William Stanley Jevons noticed something odd about coal. As steam engines got more fuel-efficient, he expected coal consumption to drop. Instead, it rose — because cheaper, more efficient engines made coal-powered machinery viable for far more uses than before, and total demand outpaced the efficiency gains.

The same pattern is now playing out in computation. Model providers have driven the cost of an individual token down through faster hardware, smaller distilled models, and brutal price competition. But that price drop didn’t just make existing AI use cheaper — it made whole categories of AI use viable that weren’t before. And that’s where the second half of the paradox comes in: it’s not just that more people are using AI. It’s that the nature of what “using AI” even means has changed underneath the pricing model.

The SaaS assumption that quietly broke

To see why, it helps to go back to what enterprise software economics used to look like — and Jevons Paradox alone doesn’t fully explain what’s happening. There’s a structural shift underneath it that’s easy to miss if you’re only looking at the price curve.

Traditional software — the code running your CRM, your accounting platform, your internal tools — does the same thing every time. The function is fixed. The output is fixed, or at least fixed in kind: the same query against the same database returns a comparably-sized answer, computed the same way, at a marginal cost so close to zero that software companies could sell it on a flat monthly seat license and still book gross margins north of 75–85%. Whether a customer used the product once a day or fifty times, the underlying compute barely moved the needle. That’s the entire economic foundation the SaaS industry was built on for two decades: predictable inputs, predictable outputs, near-zero marginal cost.

AI breaks that assumption at the root, not at the edges. A prompt isn’t a fixed function call — it’s an open-ended instruction whose computational shape depends entirely on what’s being asked. A one-line factual question might cost a few hundred tokens. A multi-step reasoning task, an autonomous agent debugging a codebase, or a system checking its own work before answering can burn through 5 to 30 times more tokens than a simple completion — and in the most complex coding or quantitative workflows, the token depth can expand by a factor of a thousand. None of that is inefficiency. It’s the system doing what it was actually asked to do: explore, backtrack, check itself, and try again, because the problem itself didn’t come with a fixed-size answer attached.

This is the real distinction worth sitting with. Old software didn’t need to experiment to produce an output — the logic was already written, and running it again produced the same shape of answer. AI, especially for anything beyond a single-turn lookup, often has to search a space of possible reasoning paths before landing on something usable. Multi-step, multi-agent problems in particular don’t just add cost linearly — token volume in collaborative multi-agent setups scales closer to the square of the number of agents involved, because each one is checking, reasoning about, and responding to the others. A monolithic, fixed-output software product was never built to price that. It’s not a software product in the old sense at all — it’s a metered, variable-cost production process wearing a software interface.

Why the falling price doesn’t save you

Once you see AI this way, the aggregate-spending contradiction stops being a contradiction. The industry didn’t respond to cheaper tokens by doing the same amount of work for less money. It responded by asking AI to attempt harder, longer, more exploratory problems that would have been prohibitively expensive to even try two years ago — reasoning chains, autonomous agents, self-correcting workflows, systems that run hundreds of internal verification passes before producing a final answer. The falling unit price didn’t shrink the bill. It expanded what was worth asking the system to do, and the new scope of what’s being asked more than absorbed the savings.

That’s also why the old SaaS pricing model is quietly cracking under real products right now. A flat $30-per-seat subscription made sense when the marginal cost of a user’s twentieth query was effectively zero. It stops making sense the moment that twentieth query might trigger an agentic workflow burning through hundreds of dollars of underlying compute — and highly active users can now do exactly that. Enterprise software is visibly shifting in response, moving from pure per-seat licensing toward hybrid models with token quotas layered on top of a base subscription, and increasingly toward pure outcome-based billing — pay per resolved ticket, per reviewed contract, per completed pull request — where the vendor absorbs the token variability directly rather than pretending it isn’t there.

The part that doesn’t get cheaper

There’s a second twist worth knowing about, because it explains where the money is actually going even as software gets more automated. While the digital side of this economy is deflationary — the marginal cost of generating text, code, or analysis is falling toward the cost of the compute itself — the physical side is doing the opposite. Data center buildout is projected to reach into the trillions of dollars by the end of the decade, and the bottleneck has quietly shifted from clever algorithms to genuinely physical constraints: electrical grid capacity, high-bandwidth memory, cooling systems, and the semiconductor supply chain. Global AI-related data center electricity consumption reportedly grew by roughly 50% in a single year recently, against a low-single-digit growth rate for electrical grids generally. You can optimize software all day; you can’t route around a substation that’s already at capacity.

That divergence — cheaper digital output, more expensive physical input — is arguably the more durable trend here. It’s why the current wave of engineering effort (model routing to smaller models for easy queries, caching repeated context instead of reprocessing it, generating executable code instead of verbose tool-calling schemas) exists at all: not because anyone ran out of clever tricks, but because the floor under all of it — power, silicon, physical infrastructure — doesn’t get more efficient at the same pace software does.

What this actually means for anyone running a business

The practical upshot isn’t “AI costs are out of control” or “AI costs are collapsing” — both are true simultaneously, and mistaking one for the whole picture leads to bad decisions in opposite directions. The unit is genuinely getting cheaper. The scope of what people are asking it to do is expanding faster than that. Treating a token like a fixed, predictable line item — the way IT budgets used to treat a software license — misses the fact that its total cost depends on something a license fee never had to account for: how much exploration a given problem actually requires, and that’s now a design choice, not a fixed constant.

That has a direct implication for how leaders should be thinking about AI cost, not just AI capability. The organizations gaining real advantage aren’t the ones chasing the cheapest per-token rate. They’re the ones that have figured out which problems in their business genuinely need that expensive, multi-step exploratory reasoning — and which ones are being run through it out of habit, when a cheaper, more constrained tool would do the job just as well. The token isn’t a utility bill to be minimized. It’s closer to a raw material whose right quantity depends entirely on the shape of the problem you’re actually solving — and getting that match right is quickly becoming the difference between AI spend that compounds into advantage and AI spend that just compounds.

Will the Paradox Keep Existing?

A caution is worth adding, though: this divergence isn’t guaranteed to run forever, because it’s really being driven by two separate forces stacked on top of each other, not one. The first is capability growth — models become better able to do genuinely new kinds of work. The second is the price-elasticity effect proper — the same falling per-token cost that classic Jevons Paradox describes. Both need to slow down before the bill stops climbing, and they don’t necessarily move together; token prices could keep grinding down 60–70% a year for a while after capability gains level off, or the reverse. Even then, “enterprises have fully integrated AI into their workflows” is a receding target, not a fixed one — the electrification of the last century shows the same pattern, where cheaper, more capable power didn’t just serve existing uses, it kept creating new appliance categories that hadn’t existed before, and consumption climbed for decades before it leveled off. Something similar could easily happen here: new products and job categories built because AI exists, arriving faster than existing use cases get saturated. If and when both forces do finally slow — capability gains flattening, and the space of things worth automating actually getting exhausted rather than continually reinvented — the paradox stops looking like a paradox at all. It just becomes what falling prices are supposed to do: the total bill starts shrinking along with the unit cost, instead of climbing despite it.

Is this happening in your enterprise right now? How is your enterprise coping with the increase in AI costs? Will be keen to hear what happens! Share them with me on LinkedIn!

Are you a researcher or developer working on Low-Resource Languages? Can I interest you in Project Atlas, global data catalog for such by AI Singapore. :)