AI and stock returns: lessons from the railways
- TEBI
- 18 minutes ago
- 10 min read
Forecasters expect artificial intelligence to make America substantially richer by 2050. That is not the same thing as making American shareholders richer, and a century of market history explains why.
The assumption is so widely shared that it rarely gets stated out loud. Artificial intelligence will make workers and companies far more productive. A more productive economy is a bigger economy. And a bigger economy, surely, carries the whole stock market up with it — fatter corporate profits, higher share prices and a more comfortable retirement for anyone whose pension quietly tracks the index.
The first two links in that chain are the subject of serious forecasting effort. The last one is not. A team led by Ezra Karger, an economist at the Federal Reserve Bank of Chicago, asked five groups — academic economists, staff at frontier AI companies, AI policy researchers, superforecasters with verified track records and a sample of the general public — how AI will reshape the American economy by 2050. They forecast growth, productivity, employment and the distribution of wealth. Not one question concerned share prices.
This is not a question about whether Nvidia is overpriced, or whether enthusiasm for a handful of chipmakers and cloud providers has run ahead of itself. Those are stock-picking questions, and they will be settled long before 2050. The question here is broader, and for most investors more consequential: what does a genuine, economy-wide productivity boom do to the returns of the market as a whole? A century of data suggests the honest answer is that nobody should assume it does much. The link between how fast an economy grows and how well its shareholders do turns out to be far weaker than that chain of reasoning requires.
What the forecasters actually expect
The survey ran from the middle of October 2025 to the end of February 2026 and covered 69 academic economists, 52 people working at frontier AI companies or on AI policy, 38 superforecasters and 401 members of the general public. The figures below are the economists'. Each respondent forecast the economy unconditionally and then under three defined scenarios for AI progress by 2030, giving their own probability for each. Every forecast concerns the United States.
They put annual US GDP growth at 2.5 per cent in both 2030 and 2050, against a baseline of 2.39 per cent for 2021 to 2025. That is a higher figure than it looks: it sits above eight of the nine institutional forecasts the authors set beside it.
The productivity forecasts matter more, because productivity is where a technology's effect on output shows up. Total factor productivity — the part of growth not explained by more workers or more capital — is put at 1.2 per cent for 2030 and 1.5 per cent for 2050, against a 2025 baseline of 0.97 per cent. The authors measure that against the IT boom of the late 1990s, which they estimate added roughly 0.65 percentage points to annual TFP growth, and describe the AI-driven acceleration their respondents expect as modest by historical standards.
This is not scepticism about the technology. The same economists assign a 61.4 per cent probability to moderate or rapid AI progress by 2030. Even under the rapid scenario they forecast annualised growth of 3.5 per cent by the late 2040s, where the AI specialists in the sample went to 5.3 per cent. Asked to explain themselves, they wrote about diffusion: new technologies spread unevenly and with long lags, as electrification, the motor car and the personal computer all did. And others named geopolitical, structural and demographic headwinds, or constraints on energy, chips and data-centre construction.
Why economic growth tells investors so little
Whether faster economic growth means better returns for shareholders is a question that predates artificial intelligence by a century.
Jason Hsu, Jay Ritter, Phillip Wool and Yanxiang Zhao took 21 developed markets over the 120 years from 1900 to 2019 and asked whether the countries whose economies grew fastest were the ones whose stock markets paid investors best. The correlation between real stock returns and real per capita GDP growth came out at −0.31: negative in sign, but with a p-value of 0.17, meaning a relationship this weak could comfortably have arisen by chance. And that p-value flatters the result. It assumes each country's experience is independent of the others, which it is not, since neighbouring economies grow together. The true significance is lower still.
Emerging markets over the shorter period from 1988 to 2019 produced a correlation of 0.19 with a p-value of 0.50: positive in sign this time, and no more meaningful.
What those numbers support is a null. Across long histories and many countries, there is no reliable positive relationship between how fast an economy grows and what its shareholders earn. Anyone reasoning from a GDP forecast to a view on future returns has to get past it.
The study appeared in The Journal of Portfolio Management in 2022; the figures above are from the freely available working-paper version. Nor is the finding new. Ritter reported it in the Pacific-Basin Finance Journal in 2005, and Dimson, Marsh and Staunton had noted it in Triumph of the Optimists three years before that.
What AI and stock returns actually depend on
If not GDP growth, then what?
The same researchers tested a different variable. Over the shorter window from 1996 to 2019, for which emerging-market earnings data exist, real stock returns correlated with real growth in earnings per share at 0.54, with a p-value of 0.04. Real per capita GDP growth over the same window and the same markets gave 0.27, with a p-value of 0.34. The earnings relationship clears conventional significance. The growth relationship does not.
GDP measures what an economy produces. Earnings per share measures what reaches the holder of one existing share. Three things sit in between, and a fourth determines how much of the gain a company keeps.
The first is arithmetic: a country's market capitalisation can rise because new companies list, or because listed companies buy private ones with newly issued shares, while earnings per existing share do not move. Competition is the second, and Hsu and his co-authors argue that it passes productivity gains to consumers as lower real prices and to workers as higher real wages, rather than into returns on capital. Third, obsolescence, which renders some existing capital worthless before its time. The fourth is not a mechanism but a condition: Ritter's 2005 paper found that technological change lifts incumbent profits where firms hold lasting protection from competition, and described that as rare.
They put it in a sentence: 'The history of technological change is that most of the benefits accrue to workers and consumers.'
The arithmetic has been measured twice, by William Bernstein and Robert Arnott in the Financial Analysts Journal in 2003. Between the end of 1925 and the end of 2001, they found, the market-capitalisation index for the whole US market grew 5.49 times larger than the price index: for every share in existence in 1926, there were 5.49 by the end. Net of buybacks, that is new share issuance of 2.3 per cent a year.
Their second calculation covered 16 countries over the century to 2000. In the seven not devastated by war, real dividends per share grew 0.7 per cent a year against real GDP growth of 3.0 per cent — the same 2.3-point gap, reached by an entirely different route.
Bernstein and Arnott went further. A rapid rate of technological change, they argued, can make dilution worse rather than better: accelerating obsolescence destroys the value of plant and equipment much as war does, forcing recapitalisation, and the dilution that follows can arrive faster than any growth the technology delivers. They aimed that at investors convinced that internet and telecommunications companies were revolutionary, in 2003. It is the mechanism AI capital spending is now testing.
The clearest route from AI-led growth to higher corporate profits runs back through Karger and his colleagues, who asked their respondents about labour's share of economic output. The economists put it at 55.48 per cent in 2025, falling to 50.0 per cent by 2050 unconditionally and to 45.0 per cent under rapid AI progress; the AI industry and policy experts went to 40.0 per cent. A falling labour share is a rising capital share, and shareholders hold claims on some of that capital.
Only some. Much of it sits in private and unlisted companies, which a public-equity investor reaches only indirectly, if at all. And the gains are expected decades out, which means anyone buying today pays for them in advance, at a price set by other people who expect them too.
The price already in the market
Vanguard holds both views at once. Its economic and market outlook for 2026, published in December 2025 and subtitled AI exuberance: economic upside, stock market downside, assigns up to a 60 per cent probability that the US economy reaches 3 per cent real GDP growth, and a 25 to 30 per cent probability that AI disappoints and produces no improvement in growth at all.
An update to Vanguard's capital markets model in January 2026, run on data to the end of December, projected annualised US equity returns of about 3.9 to 5.9 per cent over the following ten years. Vanguard states that this muted long-term projection is entirely consistent with its more bullish view of an AI-led US economic boom. The case for lower future returns does not rest on AI alone.
Jessica Wachter and Jonathan Wachter came at the same question through the spending data. Their paper, an NBER working paper issued in June 2026 and not peer reviewed, begins with a figure: the five largest US technology firms spent $380 billion on capital expenditure in 2025 and are forecast to spend roughly double that in 2026. They ask what beliefs about future productivity would make that rational, then what those beliefs imply for asset prices. In their calibrated model, the risk-free rate rises by approximately half a percentage point and the equity premium by approximately three.
But the second figure runs the opposite way to intuition. The equity premium is the extra return investors require for holding shares rather than safe assets. If a transition to AI raises it, then for any given stream of future cash flows the price an investor should pay today is lower, not higher. Higher expected returns are compensation for higher risk. They are not a windfall. It is a model, though, not a record of what investors earned.
None of which has been much help in 2026. On Vanguard's own calculations, AI-related companies outside the largest cloud providers returned 100 per cent in the first six months of the year, outperforming both the Magnificent Seven and the wider S&P 500. Being right about what prices imply and being right about what prices do next are different problems.
What happened to the railways
The record of new technologies and shareholder returns does not run all one way. Dimson, Marsh and Staunton set it out in the 2015 Credit Suisse Global Investment Returns Yearbook.
British railway mania peaked in 1846, when Parliament approved 272 new lines. Railway share prices had more than doubled in 1835 before falling back almost to where they started, doubled again by 1845, then fell by two-thirds by 1849. The index behind those moves, constructed by Rostow and Schwartz, excludes dividends. Across the quarter-century the chapter covers, investors earned an annualised capital gain of 3 per cent, plus whatever dividends they received. The railways got built.
Then the industry declined. Railways were 63 per cent of the US stock market in 1900 and almost half the UK market; by 2015 they were under 1 per cent in the US and close to nothing in Britain. And from 1900 to 2014 they beat the US market as a whole. A dollar invested in railway shares at the start of 1900 grew to $62,716, against $39,134 in the market. Airlines, the newest of the transport technologies, managed $7,090.
The dot-com record points the same way. US technology stocks rose nine-fold in the five years to March 2000, then fell 82 per cent over the following two and a half years. Turning points like that are notoriously difficult to identify in advance. Over the full period from 1995 to 2014 the sector returned 10.5 per cent a year against 9.9 per cent for the US market. By the end of 2014, only investors who had bought between January and September 2000 were still behind.
Dimson, Marsh and Staunton conclude that investors should shun neither new industries nor old ones. Prices in new industries can reflect over-enthusiasm, and investors can become too pessimistic about declining ones, but it is dangerous to assume the errors run persistently in one direction: investors may underestimate a new technology as readily as they overestimate the survival prospects of a dying one.
Where that leaves investors
At the start of 2015, Dimson, Marsh and Staunton recorded technology as 1 per cent of the UK market against 17 per cent of the American one. The London market has never had much technology in it. The label understates things, of course: banks, insurers and retailers adopt artificial intelligence without being reclassified, and UK-listed companies earn their revenues worldwide.
That does not leave a British investor outside it. Whatever they conclude, they are unlikely to be starting from no exposure at all: a global tracker or a US equity allocation already carries it, in proportions nobody chose. The weight arrived through market capitalisation, which is other investors' expectations made concrete. How long American market leadership lasts is a question the AI debate does not settle. What a UK portfolio holds in AI was inherited rather than chosen, and the only decision on the table is whether to add to it.
The economists were asked what artificial intelligence will do to the American economy.
They answered, carefully and at length. Nothing in the answer was about what shareholders will earn, and a century of evidence suggests that was the right place to stop.
Resources
Bernstein, W. J., & Arnott, R. D. (2003). Earnings growth: The two percent dilution. Financial Analysts Journal, 59(5), 47–55.
Dimson, E., Marsh, P., & Staunton, M. (2015). Industries: Their rise and fall. In Credit Suisse Global Investment Returns Yearbook 2015 (pp. 5–15). Credit Suisse Research Institute.
Hsu, J., Ritter, J. R., Wool, P., & Zhao, Y. (2022). What matters more for emerging markets investors: Economic growth or EPS growth? The Journal of Portfolio Management, 48(8), 11–19.
Karger, E., Kuusela, O., Abaluck, J., Bryan, K. A., Halperin, B., Jones, T. R., Murphy, C., Trammell, P., Reynolds, M., Mayland, D., Viswanathan, R., Mittal, A., Ceppas de Castro, R., Rosenberg, J., & Tetlock, P. (2026). Forecasting the economic effects of AI (NBER Working Paper No. 35046). National Bureau of Economic Research.
Ritter, J. R. (2005). Economic growth and equity returns. Pacific-Basin Finance Journal, 13(5), 489–503.
Vanguard Investment Strategy Group. (2025). Vanguard economic and market outlook for 2026: AI exuberance — Economic upside, stock market downside. Vanguard Research.
Wachter, J. A., & Wachter, J. (2026). What investment data implies about the AI transition (NBER Working Paper No. 35290). National Bureau of Economic Research.
Investing without a forecast
For readers who would rather work through questions like these with a professional, TEBI's Find an Adviser directory lists advisers who have publicly committed to evidence-based investing, the approach the research in this piece describes.
Readers who would rather work through it themselves will find the case set out at book length in How to Fund the Life You Want by Robin Powell and Jonathan Hollow, published by Bloomsbury in a second edition and written for a UK audience. It is available on Amazon.
