AI could restart exponential growth by making intelligence accumulable like capital. But it might also narrow the frontier of discovery.

Here’s the central economic puzzle of our time. Population growth stalled in the 1950s. Idea generation — which depends on people — slowed with it. Economic growth has been coasting on the accumulated momentum of the post-war baby boom and the computing revolution. The engine is decelerating. And then, quite suddenly, AI arrives with a proposition that changes the equation entirely: what if labour could be accumulated?

Throughout human history, labour has been the one input to production that couldn’t be stockpiled. You can accumulate capital — build factories, hoard gold, invest in bonds. You can accumulate resources — fill warehouses, secure supply chains, build reserves. But you can’t stockpile human thinking. People go home. They retire. They forget. The production function of civilisation has always had this fundamental asymmetry.

AI breaks it. Intelligence — through compute, through GPUs, through data centres — is becoming accumulable like capital. You can invest in it, scale it, deploy it around the clock, and compound it. The United States is currently investing over $1 trillion annually in data centres and software — 3.5% of GDP, a figure that would have been inconceivable five years ago. That investment isn’t consumption. It’s the accumulation of a new factor of production.

Tamay Besiroglu and the team at Epoch AI have mapped the historical correlation between population growth and idea generation with remarkable precision. Their work suggests that semi-endogenous growth theory — the idea that sustained economic expansion requires a growing population of idea-generators — may be the best model we have. If that’s right, then AI’s ability to augment and eventually substitute for human idea generation could restart the feedback loop that has been decelerating for seventy years.

Even on conservative assumptions, the numbers are compelling. Research suggests that existing AI tools — not future breakthroughs, but tools available today — could lift US labour productivity by approximately 1.8% annually for the next decade. That compounds. Over ten years, it’s a 20% improvement in economic output without adding a single worker.

But I’m not an uncritical optimist, and the data demands honesty about a troubling counter-signal.

A study published in Nature tracked what happens when AI enters scientific research. AI-augmented researchers produced three times more papers and five times more citations. Impressive. But topic diversity contracted by 4.6%, and collaboration between research groups declined. AI made science faster — and narrower. It favoured exploitation of known paths over exploration of new ones.

This is the exploration deficit, and it may be the most important risk in the AI economy. If AI systems are trained on existing knowledge and optimise for measurable outcomes, they will systematically favour the familiar over the novel. They will strip-mine known veins of insight with extraordinary efficiency while neglecting the speculative, poorly defined, high-risk inquiries that produce genuine breakthroughs.

A Harvard political theorist recently wrote a publishable academic paper in two hours using AI. He described feeling “elated and depressed” — elated at the capability, depressed at what it implied about the value of human scholarship. A Lancet study found that doctors using AI-assisted colonoscopy detection saw their independent detection rates fall from 28.4% to 22.4% when the AI was removed. The humans didn’t just delegate. They atrophied.

The question, then, isn’t whether AI creates growth. It almost certainly does, and probably at a scale that justifies the current investment boom. The question is whether it creates the right kind of growth — the kind that opens new frontiers rather than strip-mining existing ones. The kind that expands the space of possible ideas rather than concentrating effort on the most measurable subset.

This is not a technical problem. It’s a design problem — and ultimately a civilisational choice. The institutions, incentive structures, and research funding models we build around AI will determine whether the next century is characterised by explosive discovery or efficient stagnation.

The difference between exponential growth and truly infinite growth lies in what we choose to optimise for.

Azeem explores this topic with leadership teams. Begin a conversation.