Since roughly 1870, real output per person in the leading economies has grown at about two percent a year. Wars, depressions, oil shocks, the internet — the trend line absorbs all of it. Two percent compounds to something extraordinary over a century, but it is remarkably difficult to move. The productivity gains from electrification took about forty years to arrive, because factories had to be physically rebuilt around electric motors before the technology paid. Robert Solow observed in 1987 that you could see the computer age everywhere except in the productivity statistics; the payoff eventually came, arrived narrowly, and lasted less than a decade.
So the interesting question about AI is not whether it is impressive. It plainly is. The question is whether it is the kind of thing that moves that line.
I think there is a specific, checkable threshold where it would, and that almost none of the things currently treated as evidence bear on it.
Cognitive abundance is not growth
What we have now is cognitive abundance. For an unusually capable operator, coding, analysis, design, administration, research assistance, and coordination have improved by something close to an order of magnitude, in some cases more. This is real. I use these tools daily and the difference is not marginal.
It does not translate into growth, because most organizations are not constrained by cognition.
They are constrained by physical construction and manufacturing capacity. By energy and grid interconnection. By experimental latency — the battery has to actually cycle ten thousand times, the trial has to actually run. By permitting and regulation. By capital allocation cycles. By procurement and supply chains. By the gap between a system that works in demonstration and one that works unattended for ten years in the rain.
Make one task in that chain a hundred times faster and you have not made the chain faster. You have moved the constraint one link down and made it more visible. The output of the whole system is set by the slowest step, and the slowest steps are almost never the ones where a language model helps.
This is why the anecdotes and the aggregates disagree so violently right now, and why both camps think the other is being obtuse. The individual experience of a 10× improvement is not an illusion. Neither is its absence from the national accounts. They are measurements of different links in the same chain.
The threshold
The transition that would matter is not “AI makes office workers faster.” It is:
AI improves engineering → cheaper engineering makes automation and factories cheaper → cheaper capital raises the return on investment → more machines, energy, and laboratories get built → those systems make further engineering and capital formation cheaper still.
That is a loop, and loops behave differently from improvements. An improvement raises the level. A loop changes the rate.
Stated as precisely as I can:
The tell is not AI company valuations. It is not better chatbots. It is not even mass layoffs — layoffs are consistent with a pure efficiency story, which is the story that does not break two percent.
The tell is factories, laboratories, power plants, mines, robots, and infrastructure becoming faster and cheaper to design, finance, construct, and operate. That is a measurable claim about the physical world, and it is currently false. Nothing about the last three years has made a power plant faster to build.
What would have to get cheap
Five things, roughly in order of how much they would matter.
Autonomous engineering is the most important, because it sits furthest upstream. A system that turns requirements into designs, simulations, manufacturing plans, tests, and iterations — with radically less human engineering labor — would lower the cost of everything downstream that has to be engineered, which is all of it.
The hard part is not generating CAD. Generating plausible geometry is close to solved and nearly irrelevant. The hard part is the training environment: physics simulation faithful enough to trust, engineering tooling, design histories, failed prototypes, manufacturing results, field failure data. The reason engineering resists automation is that the feedback signal is slow, expensive, proprietary, and mostly unrecorded. Whoever assembles that corpus has the asset, not whoever has the model.
Autonomous science is the second lever, and its significance is easy to misjudge. What matters is not the dollar value of laboratory activity — the entire global R&D spend is a rounding error against the economy it drives. What matters is whether closed-loop systems can repeatedly run hypothesis → experiment → result → new hypothesis and produce commercially important drugs, materials, catalysts, batteries, or processes faster than human research organizations. The leverage is in the discoveries, not the labor saved.
Robotics and automated manufacturing convert cheap cognition into cheap physical labor and, more importantly, into cheap capital goods. My read is that specialized machines matter more this decade than humanoids. A humanoid is an attempt to fit automation into environments designed for humans. A factory can simply be redesigned around the automation, and redesign is cheaper than generality. Humanoids may well matter eventually; they are not required for the thesis.
Construction automation matters almost entirely as a bottleneck remover. Making an excavator thirty percent more productive is not revolutionary. Making datacenters, transmission lines, fabs, mines, and housing substantially faster to build is, because those are the physical form that new capital takes.
Energy abundance is enabling rather than large. Solar, storage, geothermal, advanced nuclear, and perhaps eventually fusion matter less as a share of GDP than as an input price. Cheap firm electricity makes compute, manufacturing, materials, transport, and automation cheaper simultaneously. The industries that matter most are usually the upstream ones that lower the cost of everything else, not the ones that become large slices of output themselves.
The case against
The honest version of the skeptical position is strong, and it is not “AI is overhyped.”
It is that the remaining bottlenecks are not made of cognition, and therefore do not yield to more of it.
Permitting is a legal and political process. An interconnection queue is a scheduling problem inside a regulated monopoly with no incentive to clear it faster. Environmental review is adversarial by design. You can generate the application instantly and wait exactly as long. Several of the binding constraints on building things in developed economies are institutional, and institutions are the part of the system most resistant to being optimized by an outside intelligence — often deliberately so.
Experimental latency is worse, because it is physics rather than politics. Intelligence can choose better experiments, run more of them in parallel, and extract more from each result. It cannot make a material fatigue faster, a cell culture divide faster, a concrete cure faster, or a five-year reliability question answerable in five months. Some of the most valuable knowledge is gated on elapsed time, and no amount of cognition compresses it.
Then there is the absorption lag. Electrification needed forty years and a physical reorganization of the factory floor. Computing needed twenty and delivered its measured payoff in a narrow window. The bottleneck in both cases was not the technology but the rate at which organizations and capital stock could be rebuilt around it — and rebuilding capital stock is exactly the thing this thesis claims will accelerate. That is either the strongest confirmation of the argument or a circularity in it, depending on which way it resolves.
And Baumol’s problem does not go away. If AI drives productivity up sharply in the sectors it touches and leaves the others alone, the untouched sectors absorb a growing share of spending, and aggregate growth converges toward the laggards. Efficiency in part of the economy is not growth in the economy.
None of these are fatal. All of them are reasons the loop might simply fail to close.
What would disconfirm this
A thesis that cannot fail is not worth holding. Here is what would tell me I am wrong.
If, by the mid-2030s, factories still take years to bring online. If power projects still take a decade. If laboratories still operate near human experimental cadence. If physical capital remains slow and expensive to expand. If the cost curve for building a datacenter, a fab, or a transmission line looks in 2036 roughly like it looked in 2026, adjusted for inflation.
In that world, AI was an extraordinary information revolution — comparable to printing or the internet, genuinely civilization-altering in what people can know and do — and not an inflection in the growth rate. That is not a small outcome. It is just a different one, and I want to be able to tell the difference.
Conversely, I should be suspicious of my own thesis if I find myself counting the wrong things. Rising AI revenue is not evidence. Model benchmark scores are not evidence. Headcount reductions at software companies are not evidence. Every one of those is consistent with the pure efficiency story.
The evidence would be capital goods getting cheaper in real terms, and the time from decision to operating asset getting shorter. Those are boring, published, checkable numbers.
Why GDP is the wrong instrument
There is a second failure mode, which is that the thesis is right about welfare and wrong about the statistics — or the reverse.
Consider a world where AI makes software, accounting, legal work, customer service, and advertising vastly cheaper, and nothing else changes. Measured growth stays ordinary. But that world is substantially richer than the accounts say, because consumer surplus rises enormously. Software that once cost ten thousand dollars becoming effectively free is a large welfare gain that shows up in GDP as a decline in software revenue.
Agriculture is the clean historical case. Automation and chemistry made food dramatically cheaper, and agriculture’s share of output collapsed. People did not respond by eating five times the calories. Demand saturated, the sector shrank as a fraction of the economy, and civilization got much richer in the process. Falling share of GDP was the signature of success, not failure.
So abundance and measured growth can diverge, and probably will. The high-growth case requires something more than efficiency: the resources released have to keep finding new productive uses rather than saturating.
Where demand does not saturate
Some sectors fill up fast. Others could absorb almost arbitrary increases in productive capacity:
- Productive capital itself — machines, factories, robots, infrastructure. This is the one that closes the loop.
- Energy and compute, which appear to have no visible ceiling.
- Housing and construction, where latent demand in developed economies is enormous and entirely supply-constrained.
- Health and longevity, where willingness to pay has no obvious upper bound at all.
- Transportation and logistics.
- New materials and industrial processes.
- Scientific R&D, which is upstream of everything.
- Eventually, space and genuinely new frontiers, though nothing rests on this.
The first entry is the important one. If released resources flow into consumption, you get a richer ordinary economy. If they flow into productive capital, you get the loop.
The 2036 test
2026–2030. Cognitive automation becomes unremarkable. Engineering agents improve sharply. Industrial robotics proliferates, mostly as specialized systems rather than general ones. Autonomous laboratories produce increasingly credible results. AI infrastructure drives an enormous capital investment boom — which is itself ambiguous evidence, since building datacenters is capital formation caused by AI without being the recursive kind. Aggregate productivity effects become visible but stay smaller than the anecdotes imply.
2030–2035. The decisive window. If engineering automation, robotics, construction, energy, and scientific automation begin reinforcing one another, advanced-economy productivity growth could move substantially above historical norms. General-purpose humanoids may be relevant by then; they are not necessary.
By about 2036 we should know. If physical capital formation itself has accelerated — if the real cost and elapsed time of building productive assets has fallen materially — then sustained periods of four to six percent real growth in advanced economies stop looking implausible.
Anything durably above that would require increasingly recursive automation of science and capital production, and I regard it as much more speculative. I include the number mostly to be pinned to something.
Small companies, large assets
If this is right, the firm that results does not look like a frontier lab.
It looks like a small human control plane, an enormous agent workforce, a set of specialized vendors, and physical assets. Organizations become astonishingly small relative to the resources they direct. A few dozen people managing billions of dollars of productive capital may become ordinary rather than remarkable.
The competitive advantage moves accordingly. Generic cognitive labor stops being a moat, because it stops being scarce. What remains scarce is physical assets, proprietary domain data, distribution, regulatory permission, capital, energy access, brands, network effects, and unusually good models of specific real-world systems.
Notice how much of that list is not technology. When intelligence is cheap, the binding constraints are mostly institutional and physical, which is a strange outcome for a technological revolution and probably an under-priced one.
The assumption that breaks
Industrial society was built on the premise that competent human cognition is scarce. Nearly every institution we have — the firm, the university, the credential, the management hierarchy, the professional licence — is machinery for allocating a scarce supply of thinking.
AI may simply reverse that premise. If it does, economic life reorganizes around whatever is still scarce, and the central entrepreneurial question changes from “can we assemble enough talented people to solve this?” to something considerably harder:
What constraint remains once intelligence is available on demand?
And if AI can eventually help remove those constraints too, the loop closes:
Intelligence creates machines. Machines make capital cheaper. Cheap capital expands experimentation. Experimentation produces better technology. Better technology makes intelligence, machines, and experimentation cheaper again.
That loop — not AGI on its own, not any particular capability threshold — is what escape velocity would actually mean. It is a claim about capital formation rather than cognition, which is why most of the current argument is being conducted about the wrong variable.
We will know within a decade. In the meantime the useful discipline is to stop counting benchmarks and start counting how long it takes to build a factory.