What is Moore's law?

Moore's law is the observation, stated by Gordon Moore in 1965 and revised by him in 1975, that the number of components worth integrating on a single circuit doubles at regular intervals, at the minimum cost per component. It is not a law of physics: it is an economic forecast about the pace of an industry, one that held for half a century and made computing steadily denser and steadily cheaper, generation after generation, without any customer having to ask for it. The side effect matters more than the statement itself. For decades, anyone designing software could assume that next year's hardware would absorb this year's inefficiencies, and could therefore treat a performance problem as a problem that waiting would solve. That rent has thinned out, and with AI inference workloads the difference shows up on the bill well before it shows up in the benchmarks.

What it actually says, and what it never said

In the article "Cramming more components onto integrated circuits", published in Electronics on April 19, 1965, Moore projected a roughly annual doubling, and from there estimated 65,000 components per chip by 1975. In 1975 itself, with more data in hand, he corrected the forecast to a doubling every two years. The "eighteen months" almost everyone quotes appears in neither statement: it is attributed to David House, then an Intel executive, who combined the doubling of transistor count with the fact that transistors were also getting faster, and got a doubling in performance. One caveat on modern figures is worth adding: manufacturers' process node names are commercial labels and no longer correspond to any measurable physical dimension, so comparing generations by counting the nanometers in the name is a meaningless exercise.

An enterprise example

An insurer puts into production an assistant that reads incoming claims and drafts the reply. It works, adoption grows, and the volume of model calls triples in two quarters. The historical answer would have been to wait for the next hardware cycle, or buy bigger machines, and let the unit cost fall on its own. That plan no longer exists: the price paid is per token consumed, and no infrastructure refresh lowers it unless somebody designs the reduction. What actually brought the bill down was deliberate work: shorter prompts, caching for recurring answers, a small model for classification and the large one only for the final draft. None of those gains arrive from a chip generation. They are design decisions, taken because somebody put them in a budget.

Why it matters for decision makers

Whether the law is over is a matter of dated positions, not settled fact: in September 2022 Nvidia's Jensen Huang declared it finished at GTC, while days later Pat Gelsinger, then Intel's chief executive, called it alive and well. For a decision maker the useful question is a different one: how much of your cost strategy rests on a gain that arrives by itself. The economic engine has moved from generic scaling to specialization, dedicated accelerators, parallelism, chiplets, and every one of those gains has to be chosen to be had. The practical consequence is that optimizing inference stops being a technical topic to postpone and becomes a budget line, with the same visibility and attribution discipline as FinOps and the metrics of AI tokenomics. A parallel observation about the energy efficiency of computing is Koomey's law.