The value of AI - what do we see?
“Us self-claiming some AGI milestone, that's just nonsensical benchmark hacking to me. The real benchmark: is the world growing at 10%” Satya Nadella, February 2025
The diffusion of AI verses other general purpose technologies - how fast?
AI has raised productivity for knowledge workers in domains including coding and law; and increased the pace of scientific discovery in domains including bioscience.
If Nobel Laureate Robert Solow were writing today he might say “You can see the AI age everywhere but in the productivity statistics.” Solow was referring to computers in 1987, and it wasn’t long before you did see computers in the productivity statistics. How long will it be before we see AI in the economy wide productivity statistics?
It took decades for general purpose technologies such as steam, electricity and computers to move from initial adoption to material impact (see chart).
However, there are grounds for thinking we should see a material impact from AI soon. AI such as image recognition has been applied for around two decades, and whilst generative AI is comparatively new, it has seen rapid adoption.
We also already have the building blocks for a deeper transformation: cloud computing, software as a service, networks and devices that can support rapid adoption of AI. This time we don’t have to wait to develop the internet and build networks, as we did before computers would have a material impact.
Even were AI having a material impact on productivity growth it would take time to see an unambiguous signal in economic data since availability of sectoral productivity datasets comes with a significant lag.
We are also yet to see new AI-first business models and institutions, but significant gains are likely from an initial phase where AI is infused into existing business models and institutions. And this is happening, often initially via individual workers adopting AI, rather than top-down.
The value of AI - what to expect?
How much might AI add to productivity and GDP growth?
One benchmark is previous general purpose technologies which added about 0.5% per annum in the case of steam (and rail) over many decades, and around 1-2% per annum in the case of computers, but over a much shorter period, before the contribution fell away. This time it may of course be different, but what is striking is the divergence of views.
Dario Amodei in his October 2024 essay Machines of Loving Grace had this to say:
“Overall, a dream scenario—perhaps a goal to aim for—would be 20% annual GDP growth rate in the developing world....”
Contrast that with Daren Acemoglu in his April 2024 paper The Simple Macroeconomics of AI:
“Using existing estimates on exposure to AI and productivity improvements at the task level, these macroeconomic effects appear nontrivial but modest—no more than a 0.71% increase in total factor productivity over 10 years [0.07% per year].”
Compared to Dario’s dream scenario, Daren’s estimate is over two orders of magnitude lower.
There are reasons for being sceptical about these outliers, and Tyler Cowen has critiqued both.
Tyler Cowen argued that Acemoglu failed to consider a deepening of automation and disagreed with his dismissal of near-term impacts on science and the pace of innovation.
Tyler Cowen has also argued that even if AI is revolutionary, the impact may be slow to materialise. Constraints include the tendency for those things that are hard to automate to grow as a share of the economy (Baumol’s cost disease) and the fact that a revolutionary impact of AI, say drug discovery, may be bottlenecked by humans, the need for clinical trails.
An intermediate estimate between the <0.1% and >10% growth contributions is offered by Philippe Aghion and Daren Bunel who consider both the experience of previous general purpose technologies and the approach adopted by Acemoglu, and conclude with an estimated growth contribution of around 1% per annum:
“Based on the first approach, we estimate that the AI revolution should increase aggregate productivity growth by between 0.8 and 1.3pp per year over the next decade. Using the second approach but with our own reading of the recent empirical literature on the various components of the task-based formula, we obtain a median estimate of 0.68pp additional annual total factor productivity (TFP) growth.”
This is before considering the impact of AI on the production of ideas, which appears likely to be material, and ideas are the only sustainable source of growth:
“Our estimates do not take into account the fact that AI automates tasks not only in the production of goods and services, our focus in this note, but also in the production of ideas.”
AI is the intention of a method of invention, as the following illustrate.
The ability of AI to predict protein structures won Demis Hassabis and John Jumper of Google DeepMind a Nobel prize in Chemistry in 2024 for ‘Alpha-Fold’. AlphaFold is now in use by over 2 million scientists and has hugely compressed the time it takes to establish the structure of a protein. Anticipate advances in medicine, biology and materials to flow from this.
Another example is the AI tool "co-scientist" by Google. Professor José R Penadés and his team at Imperial College London had spent years working out and proving why some superbugs are immune to antibiotics. He gave "co-scientist" a prompt asking it about the problem he had been investigating and it reached the same conclusion in 48 hours.
These examples illustrate that AI will contribute to health, as well as wealth, and possibly more quickly.
Is there an AI bubble, or do current market valuations reasonably reflect the potential of AI to spur growth?
“That a correction will arrive at some point seems a racing certainty…” The Economist, June 2024
Technology bubbles, which may leave behind lasting benefits, are hardly unusual. The dot-com bubble at the start of the 21st Century and the Railway Mania during the 19th Century are examples. How do these episodes compare with AI?
A way of comparing different technology driven growth episodes, and associated booms and busts in market value, is to focus on the present value of growth contributions and the share of value captured by the relevant ‘tech’ companies.
To illustrate, the present value of a productivity uplift of 0.5% per annum over base growth of 1.5% per annum for the global $100 trillion economy would be over $667 trillion i.e. many times initial GDP (calculated in perpetuity with a discount rate of 4.5%).
By comparison, the increase in value of the top ten ‘tech’ companies in the era of generative AI, starting with the release of Chat GPT n November 2022, is around $7 trillion, or around 1% of the present value of a 0.5% growth uplift in perpetuity.
That may seem small, but enterprise typically capture a small share of the value of innovation with consumers the main beneficiaries. But how does the AI-related boom compare with the railway mania and the dot-com bubble?
Railway mania
Campbell (2013) estimated that railway stocks reached a peak of £250 million, around 45 per cent of estimated UK GDP in the 1840s ). From their peak, railway stocks fell around 75 per cent.
Crafts (2002) estimated that rail contributed around 0.2 per cent per annum to productivity growth from 1830-1860.
Since the present value of 0.2% per annum growth is 3-times initial GDP, and railway stock reached 45% of GDP, the peak value captured was 15%, falling by around two-thirds to 5% of the present value post-mania.
Dot-com bubble
Tech stocks reached a peak of $6.7T during the dot-com bubble before falling 78% in value to $1.2T. Relative to US GDP of $10T at the time the peak was 67%, falling to 11%.
I assume tech, pre-generative AI, contributed 0.4% per annum to growth, a judgement based on the growth literature including Ho, Nomura and Samuels, June 2024.
The present value of 0.4% per annum growth is 5.5-times initial GDP, and tech stock reached 67% of GDP, so the peak value captured was 12%, falling by 78% to 1.5% of the present value post the dot-com bubble.
Comparing AI, rail and the internet
The following chart shows value uplift curves of $10, 20 and 30 trillion plotted against the GDP growth uplift (over an assumed underlying growth rate of 1.5% per annum and discount rate of 4.5%1) and share of value captured.
The Railway Mania (1840s) and dot-com bubble (2000) arrows show value pre- and post-crash values positioned above their estimated per annum growth contributions.
The AI related value uplift of tech stocks of around $7 trillion is around 1% of the present value of a 0.5% growth uplift. The valuation is low - even in comparison to the post-crash shares for rail and dot-com stocks of around 5% and 1.5% respectively.
If Satya Nadella is right and AI could boost global growth by 10%, AI related tech is either significantly undervalued, or is set to capture a tiny fraction of the value created.
Alternative market signals
An alternative signal that AI is expected to contribute high economic growth could be elevated real interest rates. Anticipating future wealth might be expected to consume more now versus in the future, driving up borrowing and interest rates. Alternatively strategic capital accumulation in order to capture a greater share of future income from AI could have the same effect. We do not observe a significant rise in real interest rates.
Unlocking AI
AI will require computing infrastructure and electricity, and both should be a focus in terms of reducing possible constraints on AI.
AI also requires permission-less innovation for contested and widely distributed progress. As with steam, electricity and computing permission-less innovation does not imply that applications will necessarily be unregulated, rather it implies that people are free to innovate in relation to the underlying technology.
In relation to the use of AI, Baumol’s cost disease and human bottlenecks will constraint the growth contribution. However, we can to an extent, anticipate and address such constraints.
Any job or task that requires a human credential (including professional licensing) will be off limits to autonomous AI, unless we change the rules. For example, an AI could not prescribe medicine, unless we permit it.
We can also anticipate an acceleration in new drug discovery, but unless we increase the resources available for approving new drugs, and perhaps expand mutual recognition, these discoveries will be bottlenecked by a human approval process scaled for a lower rate of discovery.
Present value of incremental growth is given by the following, where r=2.5+(g+∆g) i.e. 4.5% for base growth (g) of 1.5% and a growth increment (∆g) of 0.5%.