The AI Paradox
Jevon’s Paradox
There’s a 160-something year old idea from economics that might be the most important thing nobody in the AI industry is talking about. It’s called Jevons Paradox, and once you understand it, you start seeing it everywhere.
Back in 1865
A man named William Stanley Jevons was a British economist who noticed something strange happening during the Industrial Revolution. Steam engines were getting better and better at converting coal into useful work. You’d think that means Britain would need less coal, right? Wrong. Coal consumption went through the roof.
His idea was simple but kind of mind bending: when you make something more efficient, you make it cheaper to use. And when something is cheaper to use, people use a lot more of it. The total consumption goes up, not down. Efficiency doesn’t save resources at the system level. It unleashes demand.
That’s Jevons Paradox. And it’s back, in a big way, in the age of AI.
Okay, how does this relate to AI?
AI is developing at a mind blowing pace, it’s genuinely crazy. But aside from the speed it’s progressing, it also gets faster and more power efficient. Every year, AI models get better at doing more with less compute. OpenAI, Google, Anthropic and others are in a constant race to make their models cheaper to run. On the surface this sounds like great news for the environment and for resource consumption generally. I’m sure you’ve heard about how much water AI data centers use.
But zoom out and look at the actual numbers. Global data center energy consumption is projected to double or more within just a few years. The companies building the most efficient AI chips are also announcing the most aggressive plans to build more data centers. There’s some Canadian billionaire named Kevin O’ Leary who wants to build a 40,000 acre data center in Utah, yes FORTY THOUSAND acres. Efficiency gains are being swallowed whole by exploding demand.
This is Jevons Paradox playing out in real time.
The Cost Per Thought Problem
Here’s one way to think about it. When GPT-3 launched, running a query cost a noticeable amount. Not huge, but enough that companies thought carefully about when and how to use it. Then GPT-4 came along and was dramatically more capable but also, over time, dramatically cheaper per token. Then the models after that were cheaper still.
Each time the cost drops, the math changes for a whole new category of use cases. Things that weren’t worth automating before suddenly are. New products get built. New industries form around AI capabilities. The total number of queries being run explodes, thus increasing the cost amount, and worsening the surrounding environment.
Cheaper intelligence doesn’t mean we use less intelligence, actually the complete opposite. It means we find a thousand new things to do with it.
The Developer Productivity Loop
Now this, this is why I’m writing this article to begin with.
The same pattern we just talked about shows up in AI coding tools. Antigravity, Claude Code, Codex, and whatever else is out there make individual developers more productive. A reasonable first guess would be that this reduces the total amount of programming work that needs to happen. Fewer developer hours needed means smaller teams, less software being built, etc.
But that’s not what happens. When software becomes cheaper and faster to build, more software gets built. Startups that couldn’t afford to build a product before can now build one. Companies that wanted to automate something but couldn’t justify the cost now can. The bar for what’s worth building drops, so more things get built, and the total demand for software development keeps growing.
The idea that AI will replace programmers has been flawed from the very beginning. This isn’t to say programmers would stop using AI for their work, it’s to say it’s in the best interest of companies to avoid giving their employees free credits and plans to AI coding tools, because in the long run it costs them more. Sure, your developer puts out more code, but you also spent a substantial amount on AI for that code to even be written.
More work being done = more money being spent.
This debunks what people say about hiring AI agents to develop software being cheaper than hiring a human developer. In fact, take a look at this with me.
Microsoft canceled its internal Claude Code licenses this week after token-based billing made the cost untenable, even for a company with effectively infinite cloud resources. Uber’s CTO sent an internal memo warning the company burned through its entire 2026 AI budget in just four months. American AI software prices have jumped 20% to 37%, and GitHub (owned by Microsoft) is dropping flat-rate plans for usage-based billing across its products.
It costs more to lessen the work load!
Efficiency lowers the threshold. Lower thresholds mean more activity. More activity means more total resource use.
The AI industry has largely bet on efficiency as its answer to sustainability concerns. The argument goes something like: yes, current AI uses a lot of energy, but we’re working on making it more efficient, and eventually the energy cost per query will be so low it won’t matter. Massive data centers are saying we’ll find a way to use less water to cool our data centers, but in the long run that increases water usage, and so and so forth.
Jevons Paradox says if you make AI cheaper and more efficient, you don’t get the same amount of AI at lower cost. You get vastly more AI at roughly the same or higher total cost. The efficiency gains get reinvested into more usage, not banked as savings.
This doesn’t mean efficiency research is pointless. More efficient models are still better than less efficient ones all else being equal. But efficiency alone is not a path to lower total resource consumption. It might actually be the opposite.
The Rebound Effect
Economists talk about something called the rebound effect, which is basically Jevons Paradox applied to specific sectors. Direct rebound is when you personally use more of something because it got cheaper for you. Indirect rebound is when the money you saved gets spent elsewhere, driving consumption in other areas. And economy wide rebound is when efficiency gains across the whole sector just translate into expanded activity.
AI is experiencing all three simultaneously. Users run more queries. Companies spend their AI savings on more infrastructure. And the sector as a whole grows to fill every niche that cheaper intelligence opens up.
So What Do We Do With This?
Jevons Paradox doesn’t mean we’re doomed or that efficiency is bad. What it does mean is that efficiency gains alone cannot be expected to solve resource and energy problems at the system level. If that’s a goal you actually care about, you need something beyond just making the technology better.
Jevons didn’t have an answer to his own paradox. We don’t have a clean one either. But at least we can start by being honest about what efficiency does and doesn’t solve.
In the long run, AI is worse for companies than they might think. In 5 years, maybe you’ll be replaced by a Claude Code agent, but in 10 years when companies realize how much they’re losing, you’ll be hired.
Let’s go 🚀
Yeah you even get this renewing effect where if you adapt quickly to the new situation you can actually pull a competitive advantage out of it. Because there’s this transition phase where a lot of skeptics are still pushing back against the new development. The longer they resist the later they’ll be in the race and end up falling behind long term. This applies to individuals as well as to nations.
Write a comment