Every defence of AI's energy use rests on the same promise. Efficiency will fix it.
Models get cheaper to run. Chips get faster. The energy cost per token keeps falling, and that fall gets treated as the argument settled.
It isn't. It might be the opposite.
An old observation about coal
In 1865, the economist William Stanley Jevons noticed something odd. James Watt's steam engine used coal far more efficiently than earlier designs.
Efficiency was supposed to mean less coal burned. Instead, total coal consumption rose.
Cheaper power opened up uses that hadn't been worth the cost before. Volume grew faster than efficiency shrank it.
That's the Jevons paradox. Make something more efficient, and you don't necessarily use less of it. You often use more, because the lower cost removes the reason not to.
It's a nineteenth-century observation about coal. It maps uncomfortably well onto a 2026 GPU cluster.
The same pattern, inside a company
Uber gives you the corporate-scale version.
The company encouraged staff to use AI coding tools "as much as possible" and ranked usage on internal leaderboards. By this year's second quarter, 95% of its engineers were using tools like Claude Code every month.
Uber burned through its entire annual AI budget in around four months.
The response wasn't to ask why usage had grown so fast. It was to cap spend at $1,500 per employee per tool per month, and build a dashboard to track it.
Nobody set out to be reckless. The tools got cheaper and faster. Usage expanded to fill the space that created.
The bill arrived before anyone had modelled what "as much as possible" would actually cost.
What cheap removes
Cost used to do a job nobody assigned to it. If something was expensive, you thought about whether it was worth doing before you did it.
Expense was a crude proxy for judgement. But it worked as one.
Take the cost away and the proxy goes with it. Nothing forces the question anymore. You generate first and ask whether it was worth it later, if you ask at all.
That's not a flaw specific to AI. It's what happens whenever the cost of an action falls towards zero. The volume of the action rises to meet it.
Efficiency was never going to be enough on its own
None of this is an argument against efficient models. Efficiency is good.
It's just not sufficient. Treating it as sufficient is the mistake.
The organisations managing this well aren't waiting for a more efficient model to arrive. They're putting judgement back in deliberately, the same judgement that expense used to provide by accident.
Limits, not because AI is bad. Unlimited access to something cheap and fast always finds more to do. Not all of it is worth doing.
Efficiency was supposed to be the answer to AI's cost and climate questions. Measured against Jevons, and against Uber's ledger, it looks more like the accelerant.
The fix isn't a more efficient model. It's the discipline that was needed before efficiency arrived, the discipline efficiency made easier to skip.