AI can get cheaper like electricity even if scaling laws slow down

Nathan Lambert argues that the real Moore’s law comparison for AI is not scaling laws, but year over year gains in intelligence efficiency. If training and inference efficiency improve 2x per year at a fixed capability level, that capability becomes 32x cheaper after 5 years and roughly 1000x cheaper after 10 years.

Lambert moves the debate from model size to price per capability

The tweet is not a product announcement, but it is a useful frame. Much of the public debate asks whether bigger models will keep producing more intelligence. Lambert shifts attention to what happens when the same capability becomes dramatically cheaper.

That is the more practical market question. A company does not buy abstract intelligence. It buys cost per task, latency, reliability and the ability to put a model into a product without inference costs eating the margin.

Products care about cheaper existing capability, not only higher scores

If a model can do the same work for half the cost next year, the set of viable use cases changes. Tasks that were too cheap for human labor and too expensive for AI can suddenly make sense.

For product and engineering teams, the signal is to watch cost curves, not only leaderboards. Cheaper inference can unlock personalization, long contexts, continuous quality checks and agentic workflows that currently hit the price of every step.

Doubling efficiency each year is a scenario, not a law of nature

The 32x and 1000x numbers come from simple compounding at 2x improvement per year. That is a useful mental calculator, not physics. Hardware, energy prices, data availability, regulation or stubborn task complexity can slow it down.

Cheaper intelligence also does not automatically become a commodity like electricity. Electricity is standardized. Models differ in behavior, legal risk, data footprint and whether you trust a vendor with your product runtime.

The invoice for real tasks will decide, not the commodity slogan

The next signal will not be in a tweet or a manifesto. It will be in invoices: the cost of resolving a support ticket, writing a pull request, reviewing 1000 documents or serving a million personalized recommendations.

If those unit costs fall by orders of magnitude, AI really does start looking like infrastructure. If they stay uneven and vendor locked, it remains an expensive service with a nice curve in a slide deck.

Lilith's verdict

A commodity is not born when someone draws a beautiful chart. It is born when the accountant stops raising an eyebrow at every AI step.