superexponentialgrowth.fun Read the paper

Super-exponential growth in the agentic era

The AI economy is small at the moment, and world GDP is dominated by human production. But the AI economy is growing larger every day. Once agents become autonomous and spawn other agents, and those agents create other agents, and machines build machines, there is no limit to the agentic economy. It is going to explode.

Here is that idea as one formula. Keep scrolling to put the numbers in.

output, at
today’s prices
produced
by people
tmachines
people build
tmachines that machines build,
τ years after ordering them
 all machines built so far, by anyone
produced by all those machines
  • productivity: output per machine
  • what extra machines add
  • what people build this year
  • how many times more they build each year
  • share of machine output reinvested
  • how much cheaper machines get with scale
  • years to build what machines order
  • year machines start building

α + λ = … > 1

 

Y is what the economy can produce, counted at 2026 prices. It is not GDP, which counts goods for less as they get cheap. The paper works out the gap.

The loop

The clocks

Everyone is describing one letter. Five voices on the AI build-out and the oldest objection to it, read as terms of the formula.

Sam AltmanThe Gentle SingularityThree Observations

Robots that build robots and datacenters that build datacenters are not far off. The cost of a given level of AI falls about 10× a year.

Jensen HuangDwarkesh Podcast, Apr 2026

No bottleneck lasts more than two or three years. Once you can build one, you can build ten, then a million.

Dario AmodeiDwarkesh Podcast, Feb 2026

Growth of 10 to 20% a year, not 300%. Diffusion is fast, but not instant.

+ < 1
Dylan PatelDwarkesh Podcast, Aug 2026

30 GW of AI capacity added this year, 90 to 100 a year by 2029. Capex just over $1 trillion now, over $2 trillion in 2028.

Theodore WrightWright’s law, 1936

Each doubling of the number of aircraft built cut the cost of the next one by a fixed share.

The fixed-factor objectionRicardo, 1817Aghion, Jones and Jones, 2019

Something essential is always in fixed supply: land, metal, energy. As machines multiply it gets dearer, and each extra machine is worth less.

+ < 1

Working paper · November 2026 · Josu San Martin

Want to know where the formula comes from? Read the paper.

$125TYH · what people produce in 2026
$0.18TAKα · what machines produce in 2026
$1.5TK · chips, fabs, robots and AI power in 2026

A model, not a forecast. The page runs the formula as written, less 10% a year of wear on K, with A, n and s set so that it reproduces the baseline of the paper’s full model. The readings of what each person said are ours.