Why the Firms Waiting for This to Settle Will Not Survive It
cr4fts.com/n/8mrwn
The most expensive belief in business today is that this is a cycle. It is expressed in board papers as prudence and in strategy offsites as discipline. Wait for the hype to settle. Let the early movers make the expensive mistakes. Move when the technology matures and the use cases are proven. This posture has been rewarded repeatedly over the past three decades, and it is about to be punished.
The distinction between a cycle and a shift is not academic. A cycle is a deviation from a stable state, and the correct response is patience, because the conditions that made you successful will reassert themselves. A shift is a permanent change in those conditions, and patience is the one response that guarantees failure. The firms that die in a shift do not die from recklessness. They die from doing sensible things, competently, for a decade, against an environment that stopped rewarding them in year two.
Almost every institution currently deciding how to respond to machine learning is running the cycle playbook. They are wrong, and the historical record is unambiguous about what happens next.
Edison patented a commercially viable incandescent lamp in 1880. Two decades later, roughly three percent of American residences had electric lighting and electric motors accounted for under five percent of mechanical drive in factories. Both figures crossed fifty percent only in the 1920s. The productivity statistics moved later still.
Paul David documented this lag in his 1990 paper for the American Economic Association, written in answer to Robert Solow's observation three years earlier that you could see the computer age everywhere except in the productivity statistics. David's finding was that forty years separated technical availability from measurable economic effect, and that the delay was not a deficiency in the technology but the time required to rebuild everything around it.
The mechanism is the part that matters, and it is the part almost every retelling omits. Steam factories were organised around a single central engine, with power distributed through overhead line shafts and leather belting. Machines had to sit near the shaft. Plants were therefore built tall and narrow, and laid out according to the mechanics of power transmission rather than the sequence of the work. When electricity arrived, the first generation of factory owners did the obvious, prudent thing. They replaced the steam engine with one large electric motor turning the same shafts.
It worked perfectly. It produced almost no gain. The engine had never been the constraint.
The gain arrived only when a later generation stopped retrofitting and asked what a factory would look like if power were available at any point on the floor at negligible cost. The answer was the single-storey plant, laid out along the flow of work, with a small motor on each machine. That is not an electrical innovation. It is an organisational one, and it took managers who had never worked with line shafts to see it, because the people who had spent careers optimising around the shaft could not stop seeing the shaft.
The firms that waited did not get a cheaper entry. They got a shorter runway and a workforce trained on the wrong assumptions.
Scepticism is warranted, because the technology industry announces a revolution roughly every eighteen months and most of them are product cycles wearing better clothes. The useful thing is that economics supplies a test rather than a vibe.
Timothy Bresnahan and Manuel Trajtenberg established the criteria in their 1995 Journal of Econometrics paper on general purpose technologies. Three conditions must hold. The technology must be pervasive, meaning it is used as an input by many downstream sectors with nothing in common. It must have inherent potential for improvement, meaning it keeps getting better and cheaper for decades rather than plateauing. And it must exhibit innovational complementarity, meaning progress in the core technology raises the productivity of research in the downstream sectors, so gains compound across an economy rather than accruing to one industry.
Applied honestly, this test is brutal. Distributed ledgers changed the cost structure of a narrow band of processes and never came close to pervasiveness, which is why a decade of investment produced so little outside its own ecosystem. Virtual reality has failed the same criterion for thirty years. Mobile telephony passed all three. The shipping container passed all three and reorganised world trade as a consequence.
The question about machine learning is therefore not whether the demonstrations impress anyone. It is whether an input cost has collapsed in a way that touches unrelated industries and compounds. On that question the evidence is not close.
Stanford's Institute for Human-Centered AI tracks inference cost at a fixed capability level, which is the only honest way to measure it. Its 2025 AI Index found that querying a model performing at GPT-3.5 level on the MMLU benchmark cost twenty dollars per million tokens in November 2022 and seven cents per million tokens by October 2024. That is a decline of more than 280 times in twenty-three months.
Underneath that headline sit two independent curves. Hardware costs are falling around thirty percent annually. Energy efficiency is improving around forty percent annually. And open-weight models narrowed the performance gap to closed models from eight percent to 1.7 percent on some benchmarks within a single year, which means the capability is not controlled by a small number of firms in the way it was eighteen months ago.
Two honest caveats belong here. Holding capability constant while measuring price is flattering as well as correct: frontier capability has not become cheap, it has stayed roughly constant in price while what you get for that price has risen. And model pricing is partly a competitive weapon rather than a pure cost readout, which means some portion of the decline is reversible.
Neither caveat rescues the cycle thesis. A general economic input has fallen in real cost by a factor in the hundreds inside three years, with independently measured hardware and efficiency curves beneath it, available to anyone with a payment card. That is not a product cycle. Product cycles do not do that, and nothing that has ever done that turned out to be a cycle.
Here is where the analogy earns its keep and where most current commentary is misreading the evidence in front of it.
Nearly every large organisation has now attached a model to an existing workflow. Summarise the ticket. Draft the email. Answer the question from the knowledge base. The results have been consistently underwhelming relative to the promises, and a growing body of commentary treats this as evidence that the technology was oversold.
It is evidence of nothing of the sort. It is the line shaft, precisely. These organisations replaced the engine and left the layout alone. They automated a step in a process whose entire shape was determined by the historical cost of thinking, and then expressed surprise that changing one step did not change the economics. A summarised ticket is worth very little when the ticket exists because the process was designed around a queue of humans reading tickets.
The redesign phase, in which someone asks what a process would look like if judgement were available at any point at negligible marginal cost, has barely started. The firms currently concluding that the technology disappoints are, without exception, firms that have not attempted it. Their evidence is drawn entirely from the phase of the transition that historically produces nothing.
This is why the loud period and the productive period are different periods, and why the gap between them is where the money is made. Attention peaks when a technology becomes legible. Productivity rises much later, once enough surrounding arrangements have been rebuilt that the technology becomes the assumption a process is designed around rather than a feature bolted onto it. Between those points lies a long stretch in which many well-funded projects fail and the failure is misread as a verdict on the technology.
Almost every deployment currently running inside a large organisation falls into one of three patterns, and all three are line shafts. Recognising which one you are running is the fastest available diagnostic.
The first is assistance. A model is placed beside a worker to speed up a task the worker was already doing. Drafting, summarising, searching. This produces a measurable individual gain and almost no organisational one, because the process still contains every step it contained before, including the ones that existed only to compensate for the task being slow. Making a bottleneck faster does not help when the bottleneck was not the constraint.
The second is substitution. A model replaces a human at one step. This produces a real cost saving and a permanent ceiling, because the step was designed for a human and encodes every assumption that came with one. Firms in this pattern typically report savings in the first year and stagnation thereafter, then conclude the technology plateaued when what plateaued was their imagination about it.
The third is the demonstration, built by an innovation function, admired at a board meeting, and never integrated with anything that carries revenue. This produces nothing at all and is frequently the most expensive of the three.
The redesign pattern looks different from all of them and is rare enough that most executives have never seen one. It starts by identifying the steps that exist only because judgement was expensive, deleting them, and rebuilding the process around what remains. It usually reduces headcount in one function and increases it in another. It almost always requires changing who reports to whom, which is why it is nearly impossible to do from an innovation budget and nearly inevitable for a company that has no existing process to protect.
The most costly error is not scepticism. It is the belief that being early is the same as being right.
The firms that captured the value of electrification were overwhelmingly not the electrical equipment manufacturers. They were manufacturers who redesigned their plants, and they moved decades after the technology first worked. The firms that captured the value of cheap bandwidth were not the ones building on that assumption in 1999, most of which are gone. Timing a shift means arriving when the complements exist, not when the core first functions.
The historical durations are worth holding. Steam ran from Watt's separate condenser in 1769 to a measurable productivity effect in British manufacturing around the middle of the nineteenth century. Electricity took roughly forty years. Information technology compressed this to perhaps twenty-five, from the microprocessor in 1971 to the American productivity acceleration of the late 1990s, with the final decade doing most of the work.
The gap has been narrowing across successive general purpose technologies, which argues for a shorter wait this time. It has never once been short. Any plan modelling this transition in quarters is working from a case that does not appear in the record.
If the core input is collapsing in price, then the core input cannot be the advantage. This follows directly and it disqualifies most of what is currently being funded.
A company whose position rests on access to a model has no position, because access is a commodity purchasable by anyone with a card and getting cheaper by the month. A company whose position rests on being first to apply a technique has a position measured in months, because techniques diffuse through published papers and job changes faster than they can be defended.
What does not commoditise is specific, unglamorous knowledge of how a particular process actually runs, including the undocumented parts that live in the heads of people who have done the job for a decade. That knowledge is what permits redesign rather than retrofit, and it accrues at the speed of time spent inside a problem regardless of what compute costs. It is the only input in this transition whose price has not fallen and will not fall.
This is why the winners of the last three shifts were domain operators rather than technology vendors, and why we expect the same here. The advantage was never the dynamo. It was knowing what a factory floor should look like.
The cycle believers are not stupid and their playbook is not irrational. It has been correct more often than not, because most announced revolutions are not revolutions and patience has been the higher-expected-value strategy for a generation of executives. That is precisely why it will be so widely applied here, and why it will be so expensive.
An input that every sector uses has fallen in cost by more than two orders of magnitude in under three years. That has happened perhaps half a dozen times since 1750. On every previous occasion the incumbents retrofitted, saw disappointing results, concluded the technology was overstated, and were displaced a decade later by firms that had redesigned around it. There is no recorded instance of the patient strategy working.
The window is long enough to build in and short enough to matter. Four decades of electrification did not mean four decades of opportunity; the defaults were set in a much narrower period and then held for most of a century. The same will be true here, and the people setting them will not be the ones waiting for proof.
This is not a cycle. The conditions are not coming back. The only question that matters is whether you are rebuilding the factory or replacing the engine.