Jevons Is Not a Paradox. It Is a Capacity Plan.
Source: Hank Green, Unfortunately, You Need to Know What the Jevons Paradox is, YouTube, July 2026 (~33 min).
I. The cheap thing you keep burning
Make a fungible resource cheaper and total burn often rises, because new uses appear that were not economical before. That is not a paradox. It is demand elasticity meeting an expanding opportunity set. Call it Jevons if you need the name. Treat it as a capacity-planning rule if you run systems.
II. Coal, 1865
William Stanley Jevons wrote The Coal Question after watching Watt’s more efficient steam engines fail to “save” Britain’s coal. Efficiency made engines cheap enough for deeper mines, factories, and trains. Aggregate coal use exploded.
The chain is simple. Efficiency lowers the effective price of a use. Uses that were uneconomic walk onto the market. When the resource can serve many ends, the new uses outweigh the savings on the old ones. Intensity falls; consumption rises. The word paradox is marketing. The dynamic is ordinary once you stop assuming demand is fixed.
III. The fungibility ladder
Green’s useful tool is not the coal anecdote. It is a ladder:
| Rung | Examples | How hard Jevons bites |
|---|---|---|
| Specific goods | Hot water, a single plastic SKU | Soft. Saturation is possible. |
| Broad outputs | Code, transport, manufacturing | Hard. New applications keep opening. |
| Substrates | Energy, information, atoms | Brutal. “Enough” means running out of things the substrate can still do. |
The higher the rung, the less you should trust the sentence “efficiency will shrink the bill.” Efficiency shrinks unit cost. Falling unit cost invents demand.
IV. AI, in the present tense
You do not need twenty-year sci-fi. The present is already enough. Coding tools change how software ships. Large lab spend is a signal of expected surplus — not proof of permanent monopoly profits, but not noise either.
Code sits in a sweet spot for this dynamic: dense training data and cheap verification through tests, compilers, and typecheckers. Essays and novels do not have that loop. Paths across industries diverge; do not paste coding curves onto wet chemistry by slogan.
If AI also cheapens general problem-solving, intelligence starts to look like a fourth substrate:
atoms · energy · information · intelligence
Energy manipulates atoms. Intelligence manipulates information. Consciousness and qualia can stay outside the capacity model. They are not required for the demand dynamic.
V. Electricity and the costume plateau
US per-capita electricity was roughly flat for about two decades — efficiency gains plus offshored industrial load. It felt like enough. The plateau was partly costume. The next demand spike is structural: electrify heat and transport, and feed AI compute. Electricity has no hot-water ceiling. Under climate and grid constraints, the demand bound is closer to everything anyone wants that energy can still do.
If your mental model is “better PUE, therefore lower total power,” you are planning for rung one while living on rung three.
VI. What actually binds
In a Jevons moment, something binds growth. It is usually not the variable you are staring at — token price, FLOPs list price, model benchmark.
| Bind | Why it bites |
|---|---|
| Power gen and grid build rate | Physical lead times beat software release cadence |
| Wet-lab and empirical reality | You cannot invent cancer rates in fish by thinking harder |
| Cybersecurity surface | Cheap capability times a large attack surface is nonlinear risk |
| Policy and law | Friction that does not care about your roadmap |
| Social backlash | Concentrated gains collect a legitimacy tax |
| Epistemics | Cheap content makes truth expensive |
| Meat constraints | Human judgment, goals, taste, the lives people actually want |
“The only limit is sunlight” or a Dyson swarm sits at one possible end of the fungibility ladder. Treat it as a maybe, not a plan. Plans need the bind that hits first.
Green’s secondary note maps cleanly onto our stack: recommendation systems already reshaped cognition before chatbots. Harnesses and apps around models may matter more than raw weights. Human taste still leads product shape. That is an agency claim, not a vibe.
And the correction worth keeping: models are strong at organizing and distilling existing information. Much of science is acquiring new information — instruments, wet fingers, field work. Models may accelerate that loop. They do not replace it.
VII. Libertaria lens
If intelligence and power concentrate in a few providers, exit cost rises with every Jevons-driven use you bolt onto their substrate. Own the bind variables you can — local compute budget, power contracts, verification harnesses — or accept capture.
Whoever prices the substrate and the verifier owns the opportunity set. API rent on rung-two and rung-three dynamics is still rent.
Agent fleets amplify the pattern. Cheaper inference means more tools invoked, more tokens, more IO, more human review. Cost models that assume constant utilization under falling unit price are wrong by construction.
So the sovereignty response is blunt. Plan capacity for expanded use when unit cost falls. Put explicit ceilings — budget caps, approval gates, risk scan on tool output, credential-pool rotation — where meat and policy must substitute for missing physical binds. Prefer verifiers and harnesses you control over weights you rent.
Closing
Signal: the fungibility ladder, the claim that efficiency invents demand, and the bind list. Those are actionable for infra and fleet economics.
Noise: treating Jevons as moral panic about efficiency itself, or as automatic AGI prophecy. The mechanism needs neither.
We run agent fleets, schedulers, credential pools, and power-hungry model calls on purpose. Falling unit cost without bind design is not thrift. It is uncontrolled demand expansion dressed as productivity.
Size for Jevons. Gate for meat, truth, and law.
When imagination is a function of what is currently cheap, you do not invent the uses of what is currently expensive. Efficiency is almost never the end of the story for broadly usable resources. For AI, code, and power, expect more use if costs fall — unless a meat, policy, truth, or grid constraint binds first.