Why Intelligence Per Watt Is the Next Phase of AI
As energy becomes the next great constraint, the question shifts from how intelligent AI can become to how efficiently intelligence can be produced.
For most of the AI era, the objective was simple:
make the models smarter.
More parameters.
More data.
More compute.
More GPUs.
More electricity.
For over a decade, progress largely followed this path.
If you wanted more intelligence, you simply spent more resources.
And for a while, it worked.
But every technological era eventually encounters a constraint.
For AI, that constraint increasingly looks like energy.
This is why researchers are beginning to focus on a different metric.
Stanford researchers have helped formalise the concept of Intelligence Per Watt (IPW) — a measure of how effectively an AI system converts energy into useful answers.
In simple terms:
How much useful intelligence can a system generate for every watt of energy it consumes?
At first glance, this sounds like a hardware problem.
A question of chips, cooling systems, power grids and data centres.
But I increasingly suspect it points to something much larger.
Because once energy becomes a limiting factor, the optimisation function of AI changes.
The question is no longer simply:
How intelligent can a system become?
The question becomes:
How efficiently can a system produce intelligence?
And that shift may define the next phase of AI.
The Hidden Cost of Uncertainty
Much of the computational cost inside AI systems comes from uncertainty.
When a system encounters ambiguity it must:
→ evaluate possibilities
→ compare alternatives
→ estimate probabilities
→ reason across competing pathways
All of this consumes compute.
And compute consumes energy.
The relationship is surprisingly straightforward:
More uncertainty
↓
More evaluation
↓
More compute
↓
More energy
Which suggests the opposite is also true:
Less uncertainty
↓
Less evaluation
↓
Less compute
↓
More efficiency
This is where intelligence per watt starts becoming much more than an engineering metric.
It becomes a systems-design principle.
…
Why This Matters
Stanford’s definition focuses on how efficiently AI converts energy into useful answers.
But if useful answers are the goal, then anything that reduces the computational effort required to reach those answers becomes valuable.
That includes:
→ reuse
→ trusted pathways
→ priors
→ coherence
→ uncertainty reduction
Viewed through this lens, intelligence per watt is not only about hardware efficiency.
It is also about cognitive efficiency.
The ability to reach reliable outcomes with less evaluation, less computation and less energy.
The first phase of AI was about increasing intelligence.
The next phase may be about increasing intelligence per watt.
And that could change almost everything about how AI systems evolve.