The Compute Paradox
The Assumption
For years, AI followed a simple rule.
More compute produced more intelligence.
Larger clusters.
More GPUs.
Higher costs.
Better models.
The economics seemed obvious.
If demand increased, compute demand would increase with it.
Infrastructure became the bet.
But something interesting is happening.
Models continue getting better.
At the same time...
they keep getting cheaper.
That shouldn't be happening if compute alone explained intelligence.
The Compute Paradox
Across the industry, two conversations are happening simultaneously.
Infrastructure providers are committing hundreds of billions of dollars to future compute.
Meanwhile, model developers keep discovering ways to need less of it.
Better routing.
Better architectures.
Better inference.
Better memory.
Better reasoning.
The destination hasn't changed.
The journey has become dramatically more efficient.
Intelligence Is Becoming More Efficient
Imagine driving from London to Edinburgh.
The first journey follows every side road.
Gets lost.
Repeats sections.
Stops unnecessarily.
The second journey uses a motorway.
Both reach the same destination.
One simply burns far less fuel.
That increasingly feels like what's happening inside modern AI systems.
They're not only becoming more intelligent.
They're becoming better at reaching intelligence.
The Missing Variable
For years we focused on three variables.
Compute.
Parameters.
Data.
Those remain important.
But another variable increasingly deserves attention.
Coherence.
How well information fits together.
How efficiently reasoning progresses.
How rarely the system wastes effort.
Every unnecessary detour costs tokens.
Every repeated evaluation costs inference.
Every redundant calculation costs energy.
Coherence removes those costs.
The Economics Change
This changes more than engineering.
It changes strategy.
If intelligence per unit of compute keeps improving, then today's infrastructure decisions are also bets on tomorrow's efficiency.
The critical question is no longer simply:
"How much AI will people use?"
It's also:
"How much compute will future AI actually require?"
Those are fundamentally different questions.
Demand may continue rising rapidly.
But the compute required to satisfy that demand may not rise in the same proportion.
Beyond AI
The same principle appears inside organisations.
When leadership aligns...
when knowledge connects...
when workflows become coherent...
less effort is required to produce useful outcomes.
Humans waste less energy.
AI wastes less inference.
The pattern is remarkably similar.
Coherence reduces effort.
Reduced effort compounds.
A Different Way to Measure Progress
Perhaps we've been measuring the wrong thing.
The last decade measured AI by the amount of compute it consumed.
The next decade may increasingly be measured by the amount of intelligence produced from each unit of compute.
Not because compute stops mattering.
But because coherence makes every unit of compute more valuable.
The biggest breakthroughs may not come from building larger systems.
They may come from building systems that think with fewer unnecessary steps.
Because in the long run, intelligence isn't defined by how much energy it consumes.
It's defined by how little energy it wastes.