Efficiency Is Not Enough

Compute keeps getting more efficient and using more energy anyway; honest impact accounting starts by admitting that.

The rebound we keep rediscovering

Every generation of computing hardware does more work per unit of energy, and each generation of AI models has found ways to compress capability into fewer operations. Yet total energy used by computing keeps climbing. When a resource gets cheaper to use, we tend to use far more of it — a rebound effect observed across technologies for over a century.

This matters for how we talk about fusion and AI. It would be comfortable to argue that efficiency will tame compute demand and fusion will mop up the rest. Neither half is safe to assume. Efficiency has not flattened demand, and fusion is unbuilt.

What follows honestly

• Efficiency is necessary and should be pursued aggressively
• Efficiency alone has not, historically, reduced total demand
• New clean firm capacity is likely needed regardless of efficiency gains
• Any such capacity, including fusion, must clear its own gates first

There is a discipline here that mirrors our approach to physics: do not credit a favorable trend with more than it has delivered. Just as we refuse to claim net gain before FOAK, we refuse to claim that efficiency will absorb compute growth. Both are hopeful stories that the evidence does not yet support.

The constructive version

The useful framing is a portfolio: aggressive efficiency, smarter scheduling of flexible compute, and new firm clean generation, with fusion as one unproven candidate in the generation tier. That is less satisfying than a single silver bullet and considerably more likely to be true.

See the compute-energy problem and powering the AI era responsibly.