AK or just okay? AI and economic growth

After a typically dark and damp winter in Oxford focused on every sane economist’s least favorite of topics — business cycles — the last two months have treated my vampiric flesh with feeling the sun and feeling the AGI. In concreto, I attended the Stripe “Economics of AI fellowship” conference in SF and then helped teach a two-week “Economics of Transformative AI” summer course in Palo Alto.

While I’m still feeling the afterglow of the ultraviolet therapy I figured I’d write up some brief thoughts on how I see the current state of affairs: can AI continue scaling at such a rapid pace, and if it does, what sort of growth impacts from AI should the world expect? I don’t have a blistering take to light the discourse with — instead, a few simple points and a beyond the doctor’s daily recommended dose of uncertainty which I hope will equally aggravate the AGI-pilled and the AGI-skeptics amongst us.

One prompt for writing up these thoughts is a remark I heard Leopold Aschenbrenner make: what’s so great about the current debate raging around the macroeconomic impact of AI is that — unlike in nearly every other macroeconomic battleground — we’re going to figure out who was right and wrong. And we’re going to figure it out soon. Come 2040, I doubt we’ll have resolved whether post-2020 inflation was supply or demand driven, but whether or not AGI manages to automate vast swathes of labor, hyper-charge R&D, and induce a growth explosion? The impacts (or lack thereof) of “transformative" AI will be everywhere including the productivity statistics.

So, in the spirit of de-pseudoing my science, let’s try and make some falsifiable claims.

There are four items of business on today’s agenda:

1. Conditional on “true” AGI capabilities, explosive growth will occur. Your pet favorite bottleneck is no match for the AK model.

2. There is no decisive reason to rule out the possibility of true AGI capabilities in the near(ish) future. If you’re so sure that AGI (and explosive growth) is impossible by 2040, I would happily bet you at (say) 20:1 odds.

3. Simply “extrapolating current trends” to the best of our current abilities should lead us to have 2035-40 AGI timelines, but there are good arguments to be made in favor of either faster or slower development.

4. If you’re going to be a skeptic, the two strongest arguments are (1) the financing won’t be there for scaling to continue (2) the data wall.

Conditional on “true” AGI capabilities, explosive growth will occur. Your pet favorite bottleneck is no match for the AK model

One of my biggest pet-peeves with economists’ expressions of skepticism about the possibility of “transformative” growth (>10%) from AI is the conflation of capabilities skepticism with growth conditional on capabilities skepticism. Any belief you have about the importance of some bottleneck in constraining growth is a joint-hypothesis over how that bottleneck operates and how a given level of capabilities will run up against that bottleneck. And if you’ve ever spent time wondering how efficient the market really is, you should know that we still haven’t developed rapid-test tech for joint-hypotheses.

Consider this thread from Luis Garicano, or this post from Tyler Cowen, both of which I think are representative of the transformative growth doubters. Garicano’s thread is premised on the fact that AI will automate some but not all tasks. That’s a fine premise to have, but it’s first and foremost a belief about future AI capabilities. If you tell me there is some bound on tasks that AI will be able to automate, of course I agree the growth impacts will be mitigated (though I still think Garicano understates the case for how much growth would occur from say ~20% automation).

Cowen’s post makes a number of points, but, once again, all of the explicit bottlenecks are bottlenecks that come from AI being able to do some but not all tasks. That’s a completely reasonable belief to have! But he ends the post by writing “None of these views are based on pessimism about the capabilities of AI models” — I don’t see how that’s consistent with the assumption that only partial automation is possible.

I imagine Cowen would respond to the above by saying that some of the bottlenecks he envisions are not about “whether or not AI can do a task” but “whether or not we will let AI do the task”. For example, he writes: “Let’s say AI increases the rate of good pharma ideas by 10x. Well, until the FDA gets its act together, the relevant constraint is the rate of drug approval, not the rate of drug discovery.”

It’s important to recognize that the constraint that is “the rate of drug approval” is itself a joint hypothesis about bottlenecks and capabilities. Suppose that in addition to increasing the rate of good pharma ideas by 10x, AI’s come up with ways of simulating the experiments we perform on humans, allowing for (at least) a 10x speed-up of the procedures required to approve a drug. Sure, maybe the FDA won’t adapt these new mechanisms for a few years, but the point stands that the degree to which the bottleneck constrains growth is itself dependent on capabilities progress.

The hill I will happily die on here — with a thank you to Parker Whitfill for putting this forcefully to me — is that if we end up living in an AK world, we’re going to get AK growth: double-exponential, where growth rates themselves grow exponentially — going from (say) 2% to 4% to 8% and so on. For those who managed to avoid undergrad macro or forget what they were taught in those first few hazy weeks of defining GDP and reading Bob Solow quips, what I mean by an AK world is the following: typically, aggregate output (GDP) is modeled as coming from a combination of labor inputs, capital inputs, and the level of technology. For our purposes, the key feature of such a world is that there are diminishing returns to capital accumulation. If we double the amount of capital in the economy (you get two houses! you get two factories! etc..), we don’t double the amount of output in the economy. Factories would sit idle, houses vacant, Waymo’s both driverless and passenger-less. The reason being, we don’t have the workers to utilize all the new capital.

By contrast, in an AK world, where we have just technology (A) and capital (K) — no more pesky labor — if you get to pull the “double capital” lever, whiz-bang, next thing you know, double output. The logic for “why would we be in an AK world” is that the AI capital becomes a perfect substitute for labor: GPT-9o mini-high-thinking-mode is going to be able to do your desk job, and the robot that Claude Sonnet 7.413 makes is going to be able to do your manual job.

Why does this world lead to double-exponential growth? The nice thing about capital is that it doesn’t need to be told to go forth and multiply. If you save and invest into new capital at a faster rate than capital depreciates, capital accumulates. Therefore, once you’re in an AK world, you get exponential growth in output, even without technological progress. The growth rate of the economy is given by the growth rate of capital, which grows at the gap between the savings and depreciation rates. If technological progress is also growing exponentially (as we tend to believe it has for the last ~200 years), there is more and more output to be saved for a given amount of capital, and growth becomes double-exponential.

And all of that is before factoring in that AGI would almost certainly speed up the rate of technological progress, which, if true, could lead to even more extreme growth scenarios — i.e. hyperbolic growth.

The point of this section is not to argue that we will get true AGI that can substitute for all human labor in the economy. The point is that conditional on doing so, we are going to get explosive growth. So if you don’t think we are going to get explosive growth any time in the next 20-30 years, you are making a prediction about future AI capabilities, not just about economic bottlenecks.¹

There is no decisive reason to rule out the possibility of true AGI capabilities in the near(ish) future. If you’re so sure that AGI (and explosive growth) is impossible by 2040, I would happily bet you at (say) 20:1 odds

Let me level with my reader: at my core, I am deeply skeptical that AGI in the near future is possible. Due to some mixture of inherent “nothing ever happens”-itis and a naive desire to cling to the belief that there is something about the human mind that makes it special (the unity of our apperception!), I do not feel the AGI in my bones.

And yet.

Inexorably, metronomically, beating on against the current, the models just keep getting better. Task-length horizons widen apace; superforecasters wind up with egg on their face; and the labs continue to race.

Furthermore, the logic for how the models could continue to improve is eminently reasonable. Pre-training scaling laws have continued to hold over many OOMs. RL environments offer a plausible path to superhuman capabilities in any given domain — especially software engineering — and if they crack software, unless you’ve been granted a vision of the true value of lambda/beta, I don’t see how you can rule out the possibility of recursive self-improvement. Again — I’m not saying AGI will happen in the next 15 years — but being entirely dismissive of the prospect seems based on faith, not reason.

You might protest that extrapolating straight lines on a graph is not a Nobel-worthy methodology, but then I would ask you — on what basis are you expecting 2% growth to continue? We’re in a game of chicken between two straight lines, and fine, feel free to favor the line that’s stayed straight for longer, but I don’t see how you can be sure 2% growth won’t be the one to swerve.

Simply “extrapolating current trends” to the best of our current abilities should lead us to have 2035-40 AGI timelines, but there are good arguments to be made in favor of either faster or slower development

Right now, the METR task-time horizon framework is the democracy of AI benchmarks: the worst form of benchmark except for all the others. It’s true that the tasks are not representative of many “real-world” tasks; it’s true that beyond a 4-hour or so time-horizon there aren’t enough tasks to provide much signal; and it’s true that we have no idea what the “task-time” is of most economically relevant tasks.

Nonetheless, the fact that models are already so good at “test” questions but not able to automate many (if any) jobs in the economy seems to me to boil down to a lack of sufficient task-time horizon (and reliability). Most things people do in (cognitive) jobs either rely on context from previous work in the job (i.e. emails), or require a few hours of dedicated time on one problem.² The reason AI can’t do the job (yet) is that it doesn’t have sufficient coherence across a longer time-horizon. Ergo, monitoring task-time’s is, for now, the best way of monitoring AI progress.

So, off to the task-time monitoring station we go. If we cross our fingers and pray that straight lines continue, we find that a one-week time-horizon with 80% completion-reliability will be reached by 2030-2031, while a one-year horizon will be reached in 2034.

How does this lead me to have 2035-40 AGI timelines? My approach here is admittedly speculative. Most tasks done in the economy take (far) less than a week to accomplish, but to be a productive worker, you typically need multiple months (if not years) of context, and there are some tasks that do seem to require something like a year-long horizon. Starting a business could be broken down into a bunch of smaller horizon tasks, but there is some relevant sense in which you have to plan and execute coherently over (something like) a year.

Adding in that the METR measure is focused on unrepresentative tasks — that are more well defined than many actual tasks — leads me to shade back a few years. Same for the fact that some tasks require greater than 80% reliability. Finally, going from cognitive tasks to manual tasks will require building lots of robots — pushing in the same direction. But I don’t think these considerations are enough to push the trend-extrapolation estimate beyond 2035-40. If the robots step seems implausible to you, I urge you to take seriously what a world where the AI’s can do all cognitive tasks looks like: think about how much cognitive power would be in the machines’ heads already, how many resources could be devoted to figuring out the robots, and maybe also go read this transcript and this transcript.

Putting all this together, I claim that, if we remain on trend, we will get AGI in the 2035-40 neighborhood. The definition of AGI here is a strong one: the AI’s can do any job equally as well if not better (and cheaper) than humans.³

It should go without saying that there are myriad considerations which could push one to expect progress to not remain on trend. I’ll discuss my pet favorite “stagnation” theories in greater detail in the next section, but the straightforward argument in relation to what’s come thus far is that I am underestimating how severe the “unrepresentativeness” issue is, underestimating how important reliability is, or that I am being too credulous about the robots.

On the other side of the ledger, there are two primary arguments which push in favor of faster timelines. 1) Labs are only just starting to dedicate lots of resources to “RL environments”, and we know from AlphaZero that (the right) dedicated RL environments can rapidly achieve superhuman performance. 2) Even if the RL environments only initially work on well-defined tasks in software-engineering/math, getting those abilities to superhuman may be sufficient to kick-off recursive self-improvement and drag the rest of the capabilities along for the ride.

For similar reasons as Ryan Greenblatt lays out here I am not persuaded by those faster timelines arguments, but in a recurring theme, I don’t think they can be entirely dismissed.

If you’re going to be a skeptic, the two strongest arguments are (1) the data wall (2) the financing won’t be there for scaling to continue

If my signposting has failed to keep you on the straight-and-narrow, let me reiterate: the “2035-40” number is conditional on current trends continuing. I do think there are two quite strong arguments for why current trends may not continue.

Argument (1) is more well-trodden ground. Barring dramatic sample efficiency improvements, for the models to keep getting better — through either pre-training or RL — they are going to continually need new and relevant data. Whether sufficiently good synthetic data (for pre-training) or RL environments can be unearthed seems, at this stage of things, a mostly open question. The people closest to the drilling operations in the labs seem optimistic, but the outputs of the labs thus far give little reason to trust that optimism.

I continue to think there are boring worlds where AI progress slows dramatically due to this constraint. I’m not going to pretend I can put a probability on it, but it remains to me the simplest and strongest reason for progress skepticism. However, as Tyler Cowen likes to say — never underrate the elasticity of supply!

Argument (2) is premised on the fact that if we expect progress to continue “on-trend”, that should rely on investment into AI continuing on trend. And if investment is to continue on trend, the pre-AGI products of the AI labs — what they produce 2026-2030 — is going to need to have substantial economic value.

Vladimir Nesov has an excellent post on just how much cash we are talking. It is going to cost ~$770 billion for one company to build out the data centers necessary to train a model in 2030 that is 3 OOMs larger than the current largest models (thereby remaining on pre-training scale-up trends). By that point in time, we should expect post-training to represent about 25-50% of development costs — bringing us to $1-1.5 trillion in overall spending. That will be about 3-4% of US GDP — for one model.

Another great post by Jeffrey Heninger points out that if all of the major labs continue scaling up their equity raises at current rates, they will demand roughly the entire universe of US venture capital funding as soon as 2027.

Open AI’s own current projection is that they will have $145 billion of revenue in 2029 and $200 billion in 2030. Total annual software developer wages in the US are ~$200-300 billion. Google’s total revenue in 2024 was $350 billion. Therefore, at a minimum, if any lab wants to spend $1-1.5T in 2030, I expect them to need to raise $500 billion in debt and equity financing. Are investors going to be willing to do that? One could go through a much more detailed scenario analysis here, but I think the point is clear: it’s far from a given. Even if the labs are able to fully automate all software engineering jobs by then, it still would require immense faith on the part of investors that further scaling is going to pay off.

Who knows what the animal spirits will be like at the time — I’m not saying it’s impossible. I’m sure some will quibble with my exact numbers here. But in my modal scenario, the pace of investment growth does slow, and likewise the rate of progress.

I’d be happy to be corrected if I’m wrong about this — I’m far from an expert — but I see the “energy-use” constraint as a sub-issue of the financing issue. Epoch estimates that a single frontier training run in 2030 will require approximately as much power as the entire power capacity of US data centers circa 2023. Factoring in the compute needed for experiments and inference could as much as triple that demand. Since I’m not in the business of underrating the elasticity of supply, I believe such demand could be met if investors are willing to finance the build-out. It is the investor side I am more worried about.

Conclusion

For my next trick, after initially trying to convince the skeptical amongst us that they should take the possibility of AGI in the next fifteen years seriously, I will admit that my modal scenario is relatively bearish.

It goes something like this: dedicated RL environments and the next round or two of pre-training scale-up do cause the software engineering domino to fall. Other than some high-level jobs which require long-horizon planning, 90%+ of software engineer jobs are fully automatable by 2028-30. However, AI R&D is not solved by 2030. It’s much harder to come up with RL environments for innovative research than for completing a task with a fixed end-goal, meaning we do not enter the recursive self-improvement world.

Beyond software engineering and bit-players like call centers’ reps, most jobs in the economy will not be fully automatable by 2030. But the LLMs will be a significant productivity-amplifier in the majority of cognitive jobs. How big will that impact be? TFP growth has been running 0.5-0.75pp above its 2010-20 trend the past two years, and BOTEC calculations from current usage and estimated productivity impacts imply something like a 1pp boost to TFP already. So it’s straightforward to imagine AI boosting TFP growth by ~0.5pp per-year over the next five years, and I don’t think it’s outlandish to put the number at 1-2pp.

The investment numbers are even more dramatic. AI investment was already responsible for 20-43% of Q2 2025 GDP growth. Heninger’s numbers imply that AI labs (collectively) would be investing $720 billion to $1.2 trillion by 2027 if they remain on trend — that investment alone would generate 2-4% nominal GDP growth. I think it’s unlikely investors will pony up that much capital unless the models surprise significantly to the upside in the next year or two, but even still, 1-2% nominal and 0.5-1% real GDP growth coming from just AI investment in 2026-27 seems entirely plausible.

By 2027, even with the labs well on their way to fully automating software engineering, I expect the rate of investment growth —and, correspondingly, AI progress — to slow substantially. If the productivity impacts are as large as I’m speculating, there won’t be a true “bust”, but the scale-up will slow down with something between a bang and a whimper.

I’m rate-capped on speculation by Substack so I won’t forecast much beyond the point of slowdown. Come 2027, we will be living in a crazy world: the combined productivity and investment effects will be boosting real GDP by growth by at least 1pp. But progress will slow, keeping the effects in that ballpark for the next few years, and, unless an unforeseen breakthrough occurs, AGI will not be on track to occur by 2040.

Is the scenario I’m outlining consistent with current market pricing? Long-term real interest rates are about 1-1.5pp higher than they were 2010-2020, and while some of that is likely due to other factors (fiscal expansion, risk-premia), part of the reason is surely that higher productivity, in part from AI, has allowed the economy to “weather” higher-rates. That being said, the best direct evidence we have is that rates have declined in response to model releases. Moreover, existing “mainstream” forecasts of the next few years of AI-capex are on the low side of what I’m expecting, leading me to believe that the market is not fully pricing in what is to come. Implications of that are left for the reader.

I look forward to discovering all the different ways I was wrong.

1. Two caveats worth discussing: 1) I don’t think this argument applies equally to developing countries. Clearly, there are abundant frictions that prevent such countries from magically teleporting to the technological frontier. Explosive growth could kick-in in developed economies and only diffuse to developing countries with a lag. But I still think it would diffuse. Thanks to Arjun Ramani and Joseph Levine for discussion on this point.
2) If regulation actively blocks AI from doing certain jobs, that will of course slow growth. But I expect such a friction to be temporary, especially given the incentives of different countries to compete with each other. See, for example, Albania’s AI finance minister.

2. Source: I made it up.

3. If you want to make very limited exceptions for performers — athletes/musicians/actors — and care-oriented service jobs (nurses/masseuses), that’s fine. I simply do not think the Baumol effects would be strong enough to block explosive growth.