Journal · Data & Analysis

Moore’s Law held for 55 years. The dividend was not speed.

Transistor counts rose about 146,086,957 times between 1971 and 2026, a doubling every 2.03 years. What that actually bought was not faster software. It was analysis at almost no cost, and most businesses never collected it. This is the long version: the scaling laws, what broke, why AI is on a far steeper curve, and what it all means for running a company on statistics.

Highland Willow · Boone, North Carolina · 18 October 20265,459 words · about 24 minutes

Moore’s Law is the most quoted sentence in computing and one of the least examined. It gets produced to explain why a phone is fast, then put away. That is a waste, because the interesting part is not how much computing power grew. It is what the growth made cheap, what replaced it when it slowed, and why a business sitting on fifty-five years of compounding abundance still cannot answer a question about its own pricing.

This is a long piece. It covers what Moore actually said, whether it held, the family of scaling laws around it, what broke and when, why AI is on a far steeper curve than Moore ever was, and what all of it means for running a business on statistics rather than on instinct.

What is Moore’s Law?

In April 1965, Gordon Moore, then director of research at Fairchild Semiconductor and later a co-founder of Intel, wrote a short article for Electronics magazine titled “Cramming more components onto integrated circuits.”¹ He observed that the number of components on a chip had been doubling at a steady rate, and projected the trend forward. He had roughly a decade of data and very few points to fit.

He put the doubling period at about one year. In 1975 he revised it to about two. The familiar “every eighteen months” appears nowhere in his work; it is an average of the two figures that other people coined later and attributed back to him.

It is worth being precise about what kind of claim this is, because the imprecision is where most commentary goes wrong. Moore’s Law is not a law of physics. It is not even a prediction in the ordinary sense. It is an observation about the rate at which an industry was choosing to improve, which then became a planning target the entire industry organised itself around. Equipment makers, materials suppliers, fabs and design houses all set their roadmaps to it. The prophecy was self-fulfilling by construction, and that is a more interesting fact about coordination than it is a fact about silicon.

Did Moore’s Law actually hold?

From the Intel 4004 in 1971, carrying 2,300 transistors, to NVIDIA’s Rubin in 2026, carrying 336 billion, the count rose about 146 million times. Solve for the implied doubling period and you get 2.03 years.

Moore’s 1975 revision said two years. He was wrong by about one and a half percent across fifty-five years. There is nothing else in business forecasting with that record, and it is worth sitting with the strangeness of it before moving on.

Fifty-five years. A hundred and forty-six million times. And the number of questions a business asks of its data has barely moved.

The family of laws Moore’s belongs to

Moore’s Law never operated alone. It is one member of a family of empirical scaling observations, and most of the confusion about whether it is “dead” comes from collapsing them into one. Some have ended. Some are still running. Two of them work against you.

Read that table and the picture is not “Moore’s Law is dead.” It is that the particular combination that delivered free annual improvement to ordinary businesses has come apart, while other curves, some much steeper, have taken over.

Dennard scaling, and why your laptop stopped getting faster

Robert Dennard’s 1974 observation was the companion that made Moore’s Law valuable. As transistors shrank, power density stayed constant, so you could pack more of them in and run them faster without the chip melting.³ Transistor count and clock speed rose together, and the second one is what a user feels.

Dennard scaling broke down around 2005. Leakage current at small geometries stopped cooperating, and the thermal budget became the binding constraint. Clock frequency stalled in the four to six gigahertz band and has stayed there for two decades. A processor bought today runs at roughly the clock speed of one bought in 2006.

The gains did not stop. They changed shape. Instead of one core running faster, you got more cores, wider vector units, and eventually dedicated accelerators. Every one of those requires software written deliberately to use it. This is the single most important consequence of 2005 for a business: after that date, hardware improvement stopped being automatic and started being conditional on software. Most line-of-business software never met the condition.

Amdahl’s Law, the ceiling on that consolation

There is a limit to the more-cores answer, and it was described in 1967. Amdahl’s Law says the speedup available from parallelising a task is capped by the fraction of the task that must remain sequential. If a tenth of a job cannot be parallelised, then a thousand processors buy you less than a tenfold improvement, and ten thousand buy you barely more than a thousand did.

Month-end close, a credit check, a pricing decision: these are full of sequential steps that no quantity of cores improves. Which is why a business that waited for hardware to fix a slow process after 2005 was waiting for something that was never coming.

Rock’s Law, the one nobody mentions

Moore’s second law, usually credited to Arthur Rock, says the capital cost of a semiconductor fabrication plant doubles roughly every four years.⁴ A fab in the late 1960s cost on the order of four million dollars. A modern leading-edge fab costs ten to twenty billion and upward.

This is the quiet counter-current running underneath fifty-five years of falling prices. The chips got cheaper; the factory got catastrophically more expensive. The industry absorbed it through volume for decades. The reason that stopped working is arithmetic: when a fab costs twenty billion dollars, the volume required to amortise it exceeds what any single market can provide, and the number of companies able to attempt it falls toward one or two.

Koomey’s Law, the one that matters for data centres

Jonathan Koomey observed that computations per joule of energy doubled about every 1.57 years, a trend running back to the vacuum tube era.⁵ For a long time this was more consequential than raw speed, because it is what made laptops and then phones possible: the battery needed for a fixed computing load fell by a factor of a hundred every decade.

After 2000 the doubling slowed to roughly every 2.6 years, for the same reasons Dennard scaling failed. This now matters enormously, because the binding constraint on AI compute is increasingly electricity and cooling rather than silicon.

Wirth’s Law, the compensator

Niklaus Wirth put it as an adage in 1995: software gets slower faster than hardware gets faster.⁶ It is sometimes called the Great Moore’s Law Compensator. Successive generations of software grow in size and complexity until they absorb whatever the hardware delivered.

For fifty years, the two curves cancelled, and the cancellation felt like stability. The quoting spreadsheet got slower, the laptop got faster, and nobody noticed either. That arrangement required one of the curves to keep moving.

Is Moore’s Law dead?

The honest answer is that density is still improving and the economics underneath it have inverted.

Cost per transistor flattened at around the 28nm node.⁷ Buyers kept getting more transistors, and stopped getting cheaper ones. Then at 2nm, cost per transistor rises: wafer pricing runs near $30,000 against roughly $20,000 at 3nm.⁸ For the first time in a major node transition, the next generation costs more per unit of computing than the last.

So the useful formulation is not that Moore’s Law died. It is that Moore’s Law became a technical fact without an economic dividend. Density improves. Your cost per unit of work no longer falls automatically because of it.

What replaced it: the new golden age

In their 2018 Turing Lecture, John Hennessy and David Patterson argued that the end of Dennard scaling and the slowing of Moore’s Law would produce not stagnation but a new golden age for computer architecture.⁹ If you can no longer get performance by waiting, you get it by designing hardware for a specific domain.

That prediction has aged well. The graphics processor, repurposed for linear algebra, is the reason modern AI exists at all. Tensor processing units, neural engines in phones, and specialised inference silicon are all the same move: give up generality, win an order of magnitude on one workload.

The business consequence is subtle and important. Performance improvement is now something you buy deliberately for a specific purpose, rather than something that arrives with next year’s hardware. It has become a procurement and architecture decision instead of a force of nature.

What did all that computing power actually buy?

Follow the price rather than the transistor count, because the price is what reaches a business.

Gigaflop figures from published cost-of-computing series.¹⁰ Token prices vary by provider and by what counts as equivalent performance; see the caveats.

A gigaflop fell from the price of a good house to the price of nothing at all. In terms a business recognises: a statistical model that needed a mainframe and an overnight run in 1984 now finishes before you have finished reading the question. The cost of asking your own data a hard question collapsed by a factor somewhere in the hundreds of millions.

Demand did not follow. Most mid-sized businesses ask roughly the number of questions of their data that they asked in 1995, which is to say a monthly profit and loss statement and an argument about it. The supply curve moved eight orders of magnitude. The demand curve did not move.

AI is not on Moore’s curve. It is far steeper.

This is the part most commentary gets wrong, usually by describing AI progress as “Moore’s Law for intelligence.” The comparison understates it by a factor of about five.

Training compute for frontier models has grown about five times a year since 2020, which is a doubling roughly every 5.2 months.¹¹ Moore’s Law doubled every twenty-four. AI compute scaling has been running at something like four and a half times Moore’s rate, and it has been doing so by spending money rather than by waiting for physics.

That distinction matters. Moore’s Law was supply-side: the same dollar bought more each year. The AI compute curve is substantially demand-side: organisations chose to spend far more, on hardware that was not itself improving at anything like that rate. Curves driven by spending decisions can stop much more abruptly than curves driven by manufacturing improvement.

Scaling laws and the Chinchilla correction

The reason anyone spent that money is that model performance turned out to be predictable from compute, parameters and data. DeepMind’s 2022 Chinchilla work refined the recipe, finding that earlier large models were badly undertrained: for a given compute budget, roughly twenty training tokens per parameter is near optimal.¹² Smaller models trained on far more data beat larger models trained on less.

For a business, the useful translation is that AI capability has behaved like an engineering input with a known response curve rather than like a series of inventions. You can forecast it, within limits, which is a strange and useful property.

The inference collapse

Training is a frontier-lab problem. Inference, the cost of actually running a model, is the one that reaches your invoice, and it has fallen faster than almost any input price in industrial history. Running a GPT-4 class model cost roughly twenty dollars per million tokens in late 2022. By 2026 the equivalent is under a dollar.¹³ Published estimates of the exact multiple range from about fifty times to a thousand times depending entirely on which models are treated as equivalent, but every series points the same direction and all of them are steeper than Moore ever was.

Jevons paradox, or why your AI bill went up anyway

In 1865 the economist William Stanley Jevons observed that improvements in the efficiency of coal use did not reduce coal consumption. They increased it, because efficiency made coal economical for applications that had previously not been worth it.

The same thing is happening now, visibly. Per-token cost fell by roughly an order of magnitude in two years while enterprise consumption rose by more than a hundred times.¹⁴ Falling unit costs did not reduce total AI spend. They restructured it, moving money from labour lines to compute lines, while the rebound effect kept aggregate consumption climbing.

What AI actually changes for business data

Strip away the demonstrations and the durable change is narrow and large: unstructured material became queryable at a cost that does not require anyone’s approval.

Fifty-five months of invoice PDFs. Every support email. Field notes, supplier quotes, inspection reports, warranty claims, the free-text field nobody ever analysed because extracting it would have taken a person a quarter. That material was always the majority of what a business knew and almost none of it reached a spreadsheet. It now can.

What AI does not change is the part everyone hopes it will. It does not supply the structure. A language model reading your invoices still needs to be told what a job is, what a unit of output consumes, and which events cause which. That remains a modelling problem, and modelling problems are solved by thinking.

Statistical management of businesses: the idea is a century old

Everything above is about supply. This section is about the thing being supplied to, and it did not begin with dashboards or data lakes. It began in a factory in 1924.

Shewhart, and the most useful statistical idea a business owner can hold

On 16 May 1924, Walter Shewhart at Bell Telephone Laboratories wrote a one-page memo containing the first control chart.¹⁵ The method was simple: plot a measurement over time, add statistically derived limits, and watch.

The insight underneath it was not simple, and it is still the single most valuable statistical idea for running a business. Shewhart separated variation into two kinds. He called them chance-cause and assignable-cause; they are now usually called common-cause and special-cause.

Common-cause variation is the ordinary noise of a stable process. It has no story. Nothing happened. Special-cause variation is a genuine signal that something changed.

Almost every management meeting in the world confuses the two. Sales were down four percent last week, so somebody is asked what went wrong, and somebody produces an explanation, and an intervention follows. If that four percent sat inside the normal range of a stable process, then the explanation was fiction and the intervention was harmful. Deming called this tampering, and demonstrated that reacting to common-cause variation as though it were a signal reliably makes output worse than leaving the process alone.

Deming, the learning loop, and why sampling existed

Shewhart’s other lasting contribution was the cycle that became Plan, Do, Study, Act, carried forward and popularised by W. Edwards Deming, who took these methods to Japanese manufacturers after the Second World War.¹⁵ The loop is the whole of what people now call a data-driven business: form a hypothesis, make a change, measure what happened, decide, and go round again.

Alongside it, Shewhart worked with Harold Dodge and Harry Romig to put sampling inspection on a rational statistical footing. That work was brilliant, and it existed because measuring everything was impossible. You sampled because inspecting every unit would have cost more than the defects.

Hold that next to the price table above. The entire apparatus of twentieth-century statistical quality control was a set of elegant techniques for economising on measurement. That constraint is gone. For most businesses the population is now cheaper to analyse than the sample was in 1950.

From sample to census, and what that changes

When you can measure everything, some things get easier and one important thing does not.

What gets easier: you no longer need to infer population properties from a sample, because you have the population. Every order, every line item, every timestamp. Confidence intervals about sampling error become irrelevant when there is no sampling error.

What does not get easier: causation. A census tells you exactly what happened. It still cannot tell you what would have happened otherwise. The counterfactual is not in your data at any sample size, which is why experiments, holdouts and causal designs matter more now rather than less. More data makes correlations tighter and more confident. It does not make them causal.

The techniques that are now affordable

None of this is new mathematics. Most of it predates the personal computer. What changed is that running it used to cost more than the answer was worth.

• Statistical process control. Control charts on the metrics that matter, so you know which movements are signals. Nearly free, almost never done outside manufacturing, and arguably the highest-return analysis available to a mid-sized business.
• Price elasticity. What actually happens to volume when you move a price, estimated from your own transactions rather than assumed from a competitor’s behaviour.
• Demand forecasting. What to hold and in what quantity, with a stated interval rather than a feeling about last spring. Seasonality, trend and promotion effects separated rather than averaged together.
• Cohort analysis and lifetime value. Which kinds of customer are worth acquiring, distinguished from which were merely cheap to acquire.
• Survival analysis. Borrowed from medicine. When is an account likely to churn, and what changes the hazard. Most businesses approximate this with a spreadsheet and get it wrong, because churn is a timing question rather than a yes-or-no question.
• Causal inference. Difference-in-differences, synthetic controls, regression discontinuity. These answer “did this intervention work” when a clean experiment is impossible, which in a real business is most of the time.
• Anomaly detection. Finding the account that quietly stopped ordering eleven months ago before it becomes a lost customer rather than after.
• Bayesian methods. Starting from what you already know rather than pretending to ignorance, and producing credible intervals a decision-maker can act on instead of a p-value nobody can interpret.

Marketing analytics, where this is most visible right now

Marketing measurement is the clearest live example of statistical management changing under pressure, and it is worth being accurate about why, because the usual explanation is wrong.

The common story is that third-party cookies went away. They did not. As of 2026 third-party cookies remain live in Chrome by default: Google walked back its deprecation timeline in July 2024 and then retired ten Privacy Sandbox APIs in October 2025.¹⁶ What actually happened is signal loss from a different direction. Apple’s App Tracking Transparency hollowed out user-level tracking, and the promised replacement infrastructure was cancelled rather than delivered. Something like 71% of brands reduced reliance on user-level tracking after iOS 14.5.¹⁶

The response has been a return to methods that were standard before anybody could follow an individual around the internet.

• Marketing mix modelling. Econometric modelling of aggregate spend against aggregate outcome, separating incremental effect from baseline and seasonality. It needs no cookies, no device identifiers and no consent signals, because it models patterns rather than people. Bayesian formulations are preferred now because they produce credible intervals and cope better with the heavy multicollinearity of real media plans.
• Incrementality testing. Deliberate geographic or audience holdouts: withhold spend somewhere and measure the difference. This is the only method on the list that produces causal ground truth rather than an estimate. In a January 2026 survey of 500 senior US decision-makers, independent incrementality testing was the most trusted measurement method, ahead of mix modelling and well ahead of the in-platform reporting that most budgets are still steered by.¹⁶
• Platform attribution. Useful as a tactical signal about what is happening inside one channel. Not useful as a portfolio view, and systematically self-flattering, since every platform is marking its own homework.

The defensible arrangement triangulates all three: mix modelling for the portfolio view, incrementality for causal truth, platform numbers for tactical feedback, with disagreements treated as information rather than as a reconciliation problem to be made to go away.

Notice what this is. It is a return to econometrics, which was the state of the art in 1970, made cheap enough to run continuously rather than annually. The surveillance era was an interlude in which an easier-looking answer was available. The easier answer is leaving, and the older answer is now affordable at a cadence that was never possible before.

The loop, and why cadence beats sophistication

Here is the part that decides outcomes, and it is not the part most businesses invest in.

The advantage does not come from having better models. It comes from the rate at which a business revises its decisions in light of what the data said. A company that re-examines pricing, mix and acquisition cost every month, and acts on what it finds, will pass a company that does it annually, even when the annual company has better analysts and better tools.

This is the same arithmetic that produced a hundred and forty-six million times. Moore’s Law is not impressive because any one generation of chips was a breakthrough. It is impressive because a modest multiple repeated on a short interval for fifty-five years. A business that closes its learning loop twelve times a year rather than once is running the same compounding at a rate its competitor cannot match by being cleverer once.

Which gives a practical hierarchy, roughly in order of return:

1. Shorten the interval before improving the method. A crude analysis monthly beats an excellent analysis annually, because the crude one gets eleven more corrections.
2. Separate signal from noise before explaining anything. Control limits on the handful of metrics you actually steer by. Stop explaining common-cause variation.
3. Make one decision reversible. Pick the decision you revisit most often and make changing it cheap. Expensive-to-reverse decisions get revisited late and defended badly.
4. Only then invest in sophistication. Bayesian mix models and survival curves are worth real money, and they are worth it after the loop exists, not before.

The traps that cheap compute makes worse

Abundance cuts both ways. Every constraint that compute used to impose was also, accidentally, a discipline. Removing it removes the discipline too.

• Multiple comparisons. Test enough hypotheses and some will clear any significance threshold by chance. At a p-value of 0.05, one in twenty pure noise findings looks real. When running a test cost a week of mainframe time, nobody ran a hundred. Now a hundred is a keystroke.
• Overfitting. More model capacity against the same history produces a model that explains the past beautifully and forecasts nothing. Holdout data is the defence and it is the first thing skipped under deadline.
• Simpson’s paradox. A trend that holds in every subgroup can reverse when the groups are pooled, and vice versa. Aggregate dashboards hide this by design.
• Survivorship bias. Analysing the customers you still have tells you about the ones who stayed. The ones who left are the data you actually needed, and they are usually not in the system any more.
• Goodhart’s Law. When a measure becomes a target, it stops being a good measure. Cheap measurement means more metrics, more targets, and more of the organisation optimising numbers rather than outcomes.
• Confident automation of a wrong model. The most expensive failure available now. A badly specified model, run continuously and wired into decisions, produces wrong answers faster than any human could, with a tidy interface in front of it.

The lesson is not caution for its own sake. It is that the discipline which compute scarcity used to enforce now has to be imposed deliberately, by whoever is accountable for the model.

So why do most businesses still not do this?

Not cost, and not tooling. The obstacle is that you cannot run a regression on a business nobody has modelled.

The data exists. It is in the order system, the time clock, the production floor, the purchase ledger, and three spreadsheets on somebody’s desktop. What is missing is the structure that says how those things relate: which event causes which, what a unit of output actually consumes, where a cost belongs, and what the identity of a customer is when three systems spell the name differently.

Without that structure, every question becomes a bespoke archaeology project. The first one costs three weeks. The second costs three weeks. After the third, nobody asks a fourth, and the organisation concludes that analytics does not work here.

Building the structure is unglamorous and it is the whole job. It is also what makes the next question cheap, and the one after that, which is how a business starts to learn at a rate that compounds rather than at a rate that exhausts people.

We ran this loop on one client’s own books. The study covered fifty-five months of transaction data and roughly seventy model specifications. It identified lapsed revenue that had been invisible to the business, and it cut a modelled monthly production requirement by roughly three quarters by showing that the binding constraint sat at a single workstation rather than in capacity. The compute for all of it cost less than lunch. The modelling took weeks, which is the correct ratio and the opposite of what most people expect.

What this argument does not prove

The token price comparison is soft. Published figures for AI inference decline range from roughly fifty times to a thousand times over three to four years, depending entirely on which model pair is treated as equivalent performance. We have given the direction and a conservative range rather than pick the most impressive number, and we would treat any single precise multiple with suspicion.

Transistor counts are not performance. The chart plots density because density is what Moore described. Real-world speed depends on architecture, memory bandwidth, software and workload. Instructions per second improved by roughly four million times over the same period, which is the more honest performance figure and still enormous.

The doubling period is computed from two points. We solved for it using the first and last processors plotted, not by fitting a regression across all of them. A fit would give a slightly different number and would be more defensible; we chose the transparent calculation over the better one and are telling you so.

Cheap analysis is not correct analysis. Falling compute costs make it cheaper to produce a confident wrong answer at scale, as the traps section sets out at length.

And our figures are ours. That was one study of one business. It predicts nothing about yours until somebody measures yours.

Common questions

Sources

1. G. E. Moore, “Cramming more components onto integrated circuits”, Electronics (1965); doubling period revised to roughly two years in 1975. Moore’s law.
2. Published transistor counts for the processors plotted. Transistor count.
3. Dennard scaling and its breakdown from approximately 2005. Dennard scaling.
4. Rock’s Law, or Moore’s second law: fab capital cost doubling roughly every four years. Moore’s second law.
5. Koomey’s law: computations per joule doubling every 1.57 years, slowing to about 2.6 years after 2000. Koomey’s law.
6. Wirth’s law, also called the Great Moore’s Law Compensator. Wirth’s law.
7. Cost per transistor flattening from the 28nm node. McKinsey, Moore’s law: repeal or renewal?.
8. 2nm wafer pricing near $30,000 against roughly $20,000 at 3nm. TSMC’s chip scaling efforts reach crossroads at 2nm.
9. J. Hennessy and D. Patterson, “A New Golden Age for Computer Architecture”, Turing Lecture, ISCA 2018. Communications of the ACM.
10. Cost of one gigaflop over time. Trends in the cost of computing.
11. Frontier training compute growing about 5x a year since 2020, a doubling time near 5.2 months. Epoch AI, compute trends.
12. Chinchilla compute-optimal scaling, approximately 20 training tokens per parameter. Neural scaling law.
13. AI inference price decline, 2022 to 2026. Series vary by provider and by which models count as equivalent. LLM token price index.
14. Jevons paradox in enterprise AI: unit price falling while aggregate consumption rises. The Jevons paradox explained.
15. W. A. Shewhart, first control chart, Bell Telephone Laboratories, 16 May 1924; common-cause and special-cause variation; the cycle that became Plan-Do-Study-Act. Walter A. Shewhart.
16. Marketing mix modelling resurgence, signal loss, Privacy Sandbox retirement, and incrementality testing trust survey. Marketing mix modeling in 2026.

Client figures shown with permission. The doubling period is computed from the first and last points plotted rather than fitted across all of them, as stated in the limits above.