The correction windowThe cost of delay nobody budgets for

Fifty-three years later, that window is still open. And the tool meant to close it may be widening it.

One model, one mechanism, one variable

The Limits to Growth, published in 1972 by the Club of Rome, rests on a simulation model called World3. Five global variables (population, industrial output, food, resources, pollution) interact over two centuries. The report became one of the most cited and most misunderstood texts of the twentieth century, because most people take away a conclusion it never reached: “the world is going to collapse.”

What the model shows is more precise and more useful. It measures what happens when a signal arrives and the correction comes late. The central scenario does not describe a shortage of resources. It describes a delay in decision-making. The cascade works like this: less industrial output means less investment, which means less food per person, which means declining well-being. Each variable drags the next one down.

The variable that separates the scenarios is not the quantity of resources. It is the speed of the response.

Fifty years of data, and the dynamic holds

In 2021, Gaya Herrington, then an analyst at KPMG, published an empirical update of the model in the Journal of Industrial Ecology. She laid the real data collected since 1972 over the twelve original scenarios. The data track two trajectories most closely: BAU2 (resources more plentiful than expected, slow response) and CT, Comprehensive Technology (technology widely deployed, decline avoided, growth capped).

In 2024, Nebel and co-authors fully recalibrated the World3 model with real-world parameters. The main effect is to raise the peaks of most variables and push them a few years into the future. The model’s dynamic holds. The dates move. The mechanism does not.

For a CEO, this is operational information: the trajectory is still open, and the control variable is the delay between the signal and the decision.

AI and the unkept promise

The CT scenario assumes technological progress of 4% a year. The pace actually observed is closer to 1.5%. The gap between those two numbers is the correction window.

The most attractive hypothesis is simple: if artificial intelligence speeds up technological progress, it could bring the real pace closer to the threshold the CT scenario requires. In 2023, McKinsey estimated that AI could add up to 3.4 points of productivity a year by 2040. Goldman Sachs projected an increase of 1.5 points a year over a decade.

The actual measurements tell a different story. In March 2026, Goldman Sachs published an analysis that finds no significant correlation between AI adoption and productivity at the macroeconomic level, while recording gains of 30% on specific tasks. In 2024, Daron Acemoglu of MIT estimated that AI-driven gains in total factor productivity will not exceed 0.53 to 0.66% in total over ten years.

The promise is 3.4 points a year. The measurement is 0.66% in total over ten years. The gap between the two is the correction window itself.

The gap does not rule out AI’s potential. But it forbids taking that potential for granted.

The energy paradox

Data centers dedicated to artificial intelligence will more than triple their electricity consumption between 2025 and 2030, according to the International Energy Agency. Total data center consumption doubles over the same period, from 485 to 950 terawatt-hours. These centers connect to the same power grids as factories.

According to Lawrence Berkeley National Laboratory, the median time between a grid connection request and commercial operation has gone from less than two years to nearly five years in fifteen years. Regional markets are absorbing data center demand faster than they are building the matching capacity.

Two feedback loops face each other. The first is virtuous: AI speeds up the detection of signals, compresses decision times and pushes technological progress toward the required threshold. The second is destructive: AI consumes a growing amount of electricity, competes with industry for grid access, and contributes to the slowdown in industrial output that triggers the Meadows cascade.

The 2025-2035 window is the decade in which one of the two loops will win out.

The signal nobody read for ten years

The dynamic Meadows described is not limited to planetary models. Every organization has its own curve.

La Redoute was France’s leading mail-order retailer. The e-commerce signal was readable from the early 2000s. The correction came late. In June 2014, Kering sold the company for a symbolic one euro, after committing about €315 million to a transformation plan and €180 million to employee measures, along with a plan to cut 1,178 jobs out of 3,437 employees. The signal had been public for ten years. What closed was the correction window.

The question for a CEO is not whether Meadows was right. It is how long a signal takes to become a decision in your own organization.

What this changes in practice

I run an organization augmented by artificial intelligence myself. Analyses that used to take days come out in a few minutes. Data flows are processed continuously without tying up a permanent team. The compression of lead times is real, measurable and daily.

And that is where the paradox becomes invisible. The only resource I manage consciously is informational: tokens, context memory, speed of output. The electricity that powers those tokens sits completely outside my scope of decision. The individual experience is one of liberation. But when every user feels only the acceleration, without feeling the energy weight of their own requests, nobody feels the signal the model measures. The disconnect between individual experience and systemic consequence is exactly the kind of delay Meadows described.

The variable has not changed

It is not the resource that runs out first. It is not the technology that fails. It is the time between the moment you know and the moment you act.

Signals that arrive and that nobody reads. Decisions that drag on because the cost of delay is not on this month’s invoice. Processes that pile up because removing them takes a courage the quarterly calendar does not reward.

This reading can be proven wrong: if measured productivity gains reach the required pace before the power constraint bites, the correction window closes on its own and this analysis will have been wrong. It is measurable. We will know within ten years.

The curve does not punish growth. It punishes delay.

The correction window is a concept proposed by Peter Guillon to name the gap between the signal received and the decision made. It applies to the trajectories of organizations as much as to macroeconomic trajectories. See the definition in the glossary.