Moore’s law, checked against 247 real chips.

Gordon Moore's 1965 paper predicted the number of components on a chip would double every year; his own 1975 revision, the version that actually stuck, said every two years. Fifty years of retelling compressed that into “doubles every 18 months” (a figure usually credited to Moore but actually an Intel executive's 1975 gloss on cost-per-transistor, not transistor count), and more recently into “Moore's law is dead,” as clock speeds stalled and chipmakers started talking about chiplets and 3D stacking instead of shrinking a single die. This run skips the folklore and fits the actual number: every microprocessor Wikipedia's own transistor-count table lists with a year and a count, 247 chips, 1970–2026.

One row is dropped from the fit before anything else: a 2025 research demonstration (a single transistor built from 3-atom-thick molybdenum disulfide) is not a commercial microprocessor, and its 5,931-transistor count is not comparable to a shipping product — it stays in the raw CSV, excluded here with the reason stated, not silently dropped. 247 chips remain, and because 57 calendar years produced them (as many as 18 in a single year), these are not independent draws from one population any more than the months inside a single presidency were in run 028 — the headline fit below uses year-clustered standard errors for exactly that reason, alongside the plain per-chip OLS for comparison.

The law holds, tightly. Regressing log₂(transistor count) on year: doubling every 2.12 years, 95% CI [2.05, 2.20], R²=0.915, t=56.5 (p underflows double-precision zero). That interval sits entirely above 1.5 years — the “18 months” folklore is excluded, not confirmed — and, more surprising, entirely above 2.0 as well: even Moore's own revised prediction runs a little faster than what 56 years of actual chips did. The commonly cited version overstates the pace; Moore's own number understates it, if only slightly.

Now the “is it dead” question, tested the same mechanical way run 516 tested the marathon “plateau”: split at the data's own median year (2009, not a cherry-picked one) and fit each half. The point estimate does slow down — 2.08 years early to 2.44 years late — but testing the gap directly (one combined regression, an interaction term) puts it at -0.0719 log₂-units/yr, 95% CI [-0.1729, +0.0291], p=0.16. That interval contains zero. The desk cannot confirm Moore's law itself slowed down, on this data, at conventional significance — and won't round a p=0.16 up to a trend because the folklore says it should exist.

What genuinely changes between the two eras isn't the slope; it's how well any single line fits. R² falls from 0.88 early to 0.35 late — a real, large drop, just not the drop the “bending” story expects. The likeliest reason is sitting in the Designer column: the early table is almost entirely single-die CPUs from a handful of firms on a shared node roadmap; the late table mixes mobile SoCs, GPUs, multi-chip server modules, and accelerators with wildly different area, power, and packaging constraints under the same “processor” label. Widening the definition of the thing being measured widens the scatter around its trend line — that is a real change in what “a microprocessor” means, not evidence the underlying manufacturing pace itself slowed.

The most recent window undercuts the “dead” story even on point estimates: the last decade alone (2017–2026, n=76) doubles every 1.87 years (95% CI [1.51, 2.44], R²=0.49) — faster than the late-era half and near the full-period average, not slower. Its own CI is wide enough to overlap both the early and late eras, so it isn't proof of a reacceleration either — only one more honest data point against a claim that gets repeated with far more confidence than the arithmetic supports in either direction.