Takeoff Tracker

We track progress towards explosive economic growth driven by AI. In this scenario, production shifts primarily to capital, which includes artificial intelligence. Human labor comprises an increasingly smaller share of the input to production. This presents a unique challenge to the existing mechanism of distributing resources through the labor market, stressing political and economic systems.

Our indicators of takeoff are all consistent with takeoff from AI-driven explosive growth, but they are not necessarily proof of impending explosive economic growth.

First, we track a collection of indicators to avoid excess sensitivity to a single indicator.

Second, we define a set of rules to assess whether each indicator shows evidence of takeoff, and how strong that evidence is. We tabulate these assignments over time, allowing for comparison against perceived impact from AI and contextualization against longer-running economic trends.

Nordhaus (2021) proposed six indicators of whether we are approaching an “economic singularity”: a regime of rapid, automation-driven growth.

1. Gross substitutability of capital for labor
2. A rising rate of productivity growth
3. A rising capital share
4. A rising, and even accelerating, capital-to-output ratio
5. A rising information capital share
6. A rising amount of productivity growth not captured in standard economic accounts

We include three of these indicators:

• A rising rate of productivity growth
• A rising capital share
• A rising information capital share

We exclude the remaining three indicators:

• Gross substitutability of capital for labor: Timely and consistent estimates of the substitutability of capital and labor are difficult to maintain and interpret. (Measurement challenge)
• A rising (and even accelerating) capital-to-output ratio: We may see certain forms of capital proliferate before automation grows extensive enough to greatly increase output. Ultimately, however, each unit of capital would likely be able to produce ever more output under a regime in which growth is driven by capital accumulation. This expansion in output per unit of capital would lead to a falling capital-to-output ratio. (Ambiguity about the direction of the effect)
• A rising amount of productivity growth not captured in standard economic accounts: Given that these gains are definitionally not included in standard accounts, it is difficult to source regularly maintained estimates. However, we plan to use the Lab’s GDP-B project to add this indicator to future versions of this project. (Measurement challenge)

We supplement the original indicators proposed by Nordhaus with additional series. More detailed explainers for each metric accompany the charts below, but we provide some context here.

• Output growth: We aim to track whether we are approaching a world of rapid, automation-driven growth.
• Labor productivity growth: With output growing and labor inputs constant or falling, labor productivity rises mechanically.
• Real risk-free interest rates: In anticipation of rising incomes from explosive economic growth, individuals are disinclined to save unless they receive outsize interest. Chow et al. (2025) formalize this argument. Additionally, a rising marginal product of capital should raise interest rates mechanically.
• Network-adjusted private capital shares (NAPCS): A good’s network-adjusted private capital share is the answer to the question: of a given dollar spent on that good, what share of it is ultimately paid out in exchange for value added by capital (as opposed to labor), all the way down the supply chain? We look at the NAPCS in specific industries to assess progress toward fully self-replicating production loops.
• Energy: Rapid automation-driven growth, and the concomitant expansion of the stock of capital (especially information capital), are likely to be energy intensive, even if the energy efficiency of AI systems and other information technologies improve rapidly.

We assign each indicator to “contradictory evidence”, “neutral evidence”, “mild evidence”, or “strong evidence” in stages:

• We first transform each indicator such that increases in the indicator point toward takeoff. We take the logits of the “share” indicators, so that, like the “growth rate” indicators, their ranges are unbounded above and below. Each indicator begins in the neutral evidence category; we then assign it to contradictory, mild, or strong evidence according to the following conditions.
• We fit a linear time trend with AR(p) residuals to each indicator on data from the beginning of the sample period (typically 1997) to 2019. We tune the AR lag p on this sample.
• We next bootstrap a distribution for each indicator. In each bootstrap draw, we first simulate a synthetic history from the fitted trend with AR(p) residual after burn-in. We then refit the same model (linear trend with AR(p) residuals) to this synthetic history. This stage allows us to account for estimation uncertainty in fitting the linear trend with AR(p) residuals.
• Conditional on this refitted model, we simulate post-2019 predictive values, centered at each draw’s refitted 2010 value. In other words, we hold the trend held flat after a 2010 kink date. Accordingly, this post-2019 predictive distribution is centered around each draw’s refitted 2010 level, according to the linear trend, rather than around the trend extrapolated forward to the target date.
• We use the 2010 fitted value to center these draws, rather than extrapolating forward a trend, (a) to require indicators with a downward [upward] trend through 2019 to revert to a higher [lower] level before being assigned to the mild or strong [contradictory] evidence categories and (b) to consider the persistence of trends—or the maintenance of level gains [losses] from 2010 through 2019—as evidence. 
• We compare a one-year moving average of the realizations of the data to a one-year moving average of the distribution of simulated draws from the AR process.
• We assign realizations at or below the 20th percentile of this distribution to the contradictory evidence category. Realizations at or above the 80th percentile but below the 95th percentile are assigned to the mild evidence category, while any realization at the 95th percentile or higher is classified as strong evidence.