How Fast Could an AI Economy Actually Grow? A New Model Says the Ceiling Is Physical, Not Cognitive

Strip away wages and the scarcity of workers, and ask a narrower question: how fast can a pile of factories, mines, power plants and machine tools build more of itself? That is the question at the center of "The AI Industrial Explosion — Part 1: Maximum growth rates with current production methods," published in May 2026 by the pseudonymous author djbinder and now circulating well beyond the forums where it first appeared. Its answer is a doubling time measured in months to a couple of years — and a claim that the binding constraint on an automated economy is not intelligence but concrete, copper and kilowatt-hours.

"Once human labor is automated, the constraint on growth becomes the speed at which the economy's physical capital can reproduce itself," the author writes. "Using US input-output data, I find this economy could double in about a year, in line with other estimates that assume full automation."

What the model actually does

The method is deliberately old-fashioned. The author takes the 2017 Bureau of Economic Analysis Supply-Use tables at detail level — 398 commodity codes, extended to 400 with sectors for robots and compute — and treats them as a fixed Leontief technology. Two new "workers" are inserted: a humanoid robot body at a central assumption of $30,000, and an eight-GPU AI server at roughly $425,000 of all-in data center capital. Labor is then set free, all output reinvested, and the model solves for the maximum self-sustaining expansion rate — the classic Von Neumann growth rate.

The outputs depend heavily on how hard you assume the machines run. At today's utilization the model gives 0.61 per year, a 1.1-year doubling; at round-the-clock, 168-hours-a-week operation, 1.38 per year, a six-month doubling. Layering in robot and compute costs, doubled mining and fossil fuel costs, and construction lags of one month for equipment and six months for structures, the author lands on "the Von Neumann growth rate is 0.59 yr⁻¹ at emergency utilization and 0.95 yr⁻¹ at full 168 hr/wk utilization, corresponding to doubling times of 1.2 years and 9 months respectively." Assume households consume some of the output — a Ramsey savings rule — and rates fall to 0.32 and 0.57 per year: doublings of roughly 2.2 years and 1.2 years.

Sectoral composition is where the "physical, not cognitive" framing bites. In the model's fastest-growing economy, construction absorbs 19.5% of output against 4.7% today; machinery jumps from 1.1% to 17.1%, a 16-fold shift; primary metals rise 11-fold. Robot manufacturing takes 1.4%, data centers 0.5%. As the author puts it: "The present-day US economy is far from the fastest-growing composition, with too many services and consumer goods relative to mining and heavy manufacturing."

Energy and materials bite as friction, not a wall. Swapping the fossil grid for solar plus four hours of battery storage at 2030 costs drops the emergency-utilization rate from 0.79 to 0.45 per year; nuclear gives 0.50. Depletion is harsher for hydrocarbons than metals: "Oil cannot supply the energy needed to build more than about 2x the current capital stock" — a claim the author later softened in comments, noting substitutes exist but cost more today.

Why It Matters

The case for taking this seriously is that it is framed as a lower bound. It assumes no technological improvement beyond whatever automates labor, uses recipes designed around human workers at human wages, and is cross-checked against OECD data for 18 countries, giving a mean of 0.54 per year. If the numbers are even directionally right, mainstream forecasts are not merely conservative: "Most economists predict that AI will add only a few percentage points to growth. If AI can fully automate the economy, these predictions are not just wrong but off by an order of magnitude." Epoch AI's Andrei Potlogea and Anson Ho reached a compatible conclusion from a different model, writing that "we are increasingly puzzled by the views of highly confident AI skeptics, currently dominant in the economic profession."

The case against is that the assumptions do enormous work. Fixed input-output coefficients mean no substitution, no price adjustment, no demand constraint — the author concedes his sector shares "should not be taken too literally: relative prices would shift substantially as the economy rebalanced." Nobody is asked whether anyone can buy the output. On LessWrong, the commenter p.b. pressed exactly that: "as long as humans retain control, the growth rate of the economy should be limited by how fast humans want more stuff... But I can't help but wonder, who is buying a car every other day?" AI researcher Ryan Greenblatt pushed on the supply side, arguing the model waves past semiconductors: "maybe it's the bottleneck after the 4th doubling rather than as of the first, but it seems super plausible it's the bottleneck to me."

Economists have made the more fundamental objection for years. AEI senior fellow Will Rinehart argues that "AI can revolutionize our world without causing explosive growth," and that such results depend on frictionless adoption — instant permitting, instant retooling, instant consumer adaptation. Stanford's Charles Jones has shown that replacing all labor with capital in standard task models produces "a singularity in which knowledge and incomes go to infinity in finite time" — less a prediction than a sign the models have been pushed past where they mean anything.

What to Watch

The testable pieces here are the input prices, not the growth rates. Watch whether humanoid robot bills of materials converge toward the $30,000 assumption or stall in the low hundreds of thousands; whether leading-edge fab capacity scales faster than its long lead times allow; and whether US data center power procurement keeps clearing at multi-gigawatt scale or hits interconnection queues that look a lot like the author's construction-lag term made real. Watch, too, for engagement from credentialed economists — Epoch has publicly invited it, and the debate is so far conducted largely among people who already agree full automation is the relevant scenario. Part 5 landed July 24; the argument is still being written.