The AI Dividend: Who will capture the value?
Artificial intelligence may well enrich the economy that uses it. History suggests that is a different matter from enriching the people who pay to build it.
At a glance
Being right about AI is not the same as profiting from it. Britain was right about the railways and the late-1990s carriers were right about the internet, yet many of the people who financed both lost money. The question for a family's capital is who captures the value, and whether today's prices already assume the answer.
The cost is concentrated; the benefits may not be. Since 2024, companies enthusiastic about AI have accounted for all net growth in US listed-company investment, most of it from Amazon, Alphabet, Microsoft and Meta. Their balance sheets make a disorderly collapse less likely. They do not guarantee that the next dollar earns its cost of capital, or that competition won't hand the gains to customers.
Expectations have risen quickly. Share prices imply AI will lift US output by around 3.6% through software engineering alone, and that expected gain more than doubled in the first half of 2026. With real yields up more than 70 basis points this year, a price that assumes large gains leaves little room for delivery that falls short.
Deployment, not capability, sets the timetable. Graduate unemployment shows no clear AI effect yet, which says little either way about adoption. A better gauge is the gap between the roughly one-third of tasks in computer and mathematical occupations that AI usage covers and the 94% theoretically within its reach. How fast organisations close that gap will decide when today's infrastructure earns its keep.
People accept help more readily than they hand over responsibility. Two-thirds of workers worldwide expect AI to make their jobs easier, and nearly one in five also see it as a threat. Fear is strongest in roles that carry authority over people's lives, such as judging cases, which bears on which businesses can keep the gains.
For families, concentration hides behind labels. Chip makers, a data-centre fund, private credit to a power developer and a private engineering business may all depend on the same few capital budgets. We test each holding on its net benefit from AI, whether it can keep that benefit, the return on any extra capital, and the price. We stay open to the possibility that the users of AI, not its builders, prove the better investments
Right about the railways
In the autumn of 1845, it was hard to open a British newspaper without finding a prospectus for a new railway. The country was right about the railways; within a generation they had remade its economy. Many of the people who financed them lost money, some of them everything. A century and a half later, telecoms companies laid far more fibre-optic cable than anyone could use, on the correct view that the internet would transform commerce. It did, but the benefit flowed mostly to the customers of that cheap capacity rather than to the carriers that built it.
We do not think artificial intelligence is a passing enthusiasm. But the question that matters for a family's capital is narrower than whether AI will change the world. It is who gets paid when it does, and whether today's prices have already settled that question in the seller's favour. It is quite possible to be right about a technology and still make a poor investment in it: competition can hand the gains to customers, growth can come with thin returns on capital, and an excellent business bought at the wrong price can disappoint for a decade. For a family with listed investments, private funds and perhaps an operating business, there is a further complication: the same AI spending cycle may appear in several places under different names.
But the question that matters for a family's capital is narrower than whether AI will change the world. It is who gets paid when it does, and whether today's prices have already settled that question in the seller's favour.
A small group of recent academic papers, from the Federal Reserve, the National Bureau of Economic Research and a team of psychologists, helps separate these questions. None of them settles the argument, but together they draw a more useful picture than the euphoria or the alarm that fills most of what is written on the subject.
The bill arrives first
Two Federal Reserve economists, Aakash Kalyani and Huiyu Li, scored the earnings calls of America's listed companies for how warmly managements talked about AI, a measure of mood rather than of audited spending. By late 2025 about a quarter counted as enthusiasts. Since the start of 2024 those enthusiasts have accounted for all of the net growth in capital spending and research budgets among listed firms, and most of it came from four companies: Amazon, Alphabet, Microsoft and Meta. Spending on information-processing equipment, software and data centres reached a third of US business investment in the third quarter of 2025, the highest share since the series began in 1947. UBS now expects global AI-related capital spending to reach around $900bn this year and $1.2tn in 2027.
The result is an asymmetry. The cost of the infrastructure sits on a handful of balance sheets, while the benefits may be scattered across thousands of companies and millions of customers. Where the value settles depends on whether the builders keep their pricing power or competition passes the gains to those renting their capacity, and on whether demand arrives fast enough to justify the cheques already written. The Fed authors name both dangers: disappointing demand would land on a few large firms, while too much pricing power could slow the very adoption they are counting on.
These are not the profitless start-ups of 1999, a point academics Satya Subrahmanyam and Richard Beainy, among others, make forcefully. A company that can fund tens of billions of dollars a quarter from its own operations is a far sturdier proposition than Pets.com, and that lowers the risk of a disorderly collapse. It does not tell us whether the next dollar will earn its cost of capital. Nor is borrowing, on its own, a sign of strain: J.P. Morgan expects the hyperscalers to issue around $280bn of bonds this year and $315bn next, but cheap debt can be a choice rather than a necessity. The test worth watching is whether capital spending begins to outrun the cash these businesses generate.
Chart 1: Capital expenditure against operating cash flow for the four largest spenders - Alphabet, Amazon, Meta and Microsoft, Q3 2022 - Q2 2026 (actual), Q3 2026 - Q1 2027 (estimate).
Source: Bloomberg
What the price already assumes
A new NBER paper by Alex Blumenfeld, Jonathon Hazell, Chen Lian and Andreas Schaab tries to read the market's mind. The authors take companies that use a lot of software engineering without selling software or chips, such as eBay, Expedia and Airbnb, and find that their shares respond more strongly to AI news the more of their payroll goes to engineers. Run through an economic model, this implies the market expects AI to raise the level of US output by about 3.6% through software engineering alone, based on news to the end of 2025. Alternative calibrations put the figure between 1.8% and 6.5%.
The detail we keep returning to is that in the first half of 2026, as coding agents spread, the AI news priced into these shares exceeded that of the previous three years combined. The market's expected gain more than doubled in six months. That cuts both ways. The implied gains are not fanciful; they sit within the range of serious academic estimates. But they are expectations rather than measurements of output, and a price that already assumes large gains leaves little room for delivery that falls short of them. Work the authors cite also finds that big gains in coding activity turn into much smaller gains in finished projects and releases, a reminder that software is only one link in the chain.
The timing is awkward. The 10-year US real yield, a benchmark for the cost of capital, has risen by more than 70 basis points this year, to around 2.7% by this month. Equities have held up regardless, a resilience we attribute largely to strong earnings – but when the discount rate and expectations rise together, the margin for disappointment narrows from both sides.
The view from the job market
The labour market offers a different kind of check. Robert Fairlie and Jane Wu of UCLA studied the class of 2026, widely expected to be the first to feel AI in its job prospects, and found nothing yet that can be told apart from noise. Unemployment among 22-to-25-year-old graduates averaged 7.3% this summer, against 7.2% a year earlier and 7.8% in 2024. A broader measure, which counts those who want work but have stopped looking, reached the highest of the five summers studied, though not by a statistically significant margin.
That tells us less than it might seem to. An absence of visible displacement is consistent with AI spreading slowly, but just as consistent with AI helping people do their jobs rather than taking them, which would be good news for productivity. The authors themselves warn that later graduating classes may fare worse.
A better guide to the pace of deployment is the gap between what AI could do and what it is used for. Anthropic's usage-based measure covers about a third of tasks in computer and mathematical occupations, far below the 94% it estimates to be theoretically within AI's reach. Spending suggests the gap is narrowing: among firms tracked by the corporate-card company Ramp, a sample tilted towards young, venture-backed businesses, median AI spending per employee more than doubled between December 2025 and July 2026. How quickly it narrows across the wider economy matters a great deal to investors, because the revenue that justifies today's infrastructure depends on organisations changing how they work.
Help is welcome. Handing over is another matter
Adoption depends on what people will let the technology do. A survey of more than 64,000 adults in 32 countries, by economists at the Bank for International Settlements and several universities, found that about two-thirds of workers expect AI to make their jobs easier and about 30% see a threat. Nearly one in five expect both, which may be the most human finding of all. We noted that workers in finance and insurance felt markedly more threatened than those in manufacturing, the study's benchmark.
A second study, in American Psychologist, asked 10,000 people in 20 countries how they would feel about AI as a doctor, judge, manager, care worker, journalist or member of the clergy. Fear tracked the gap between the qualities a role seems to demand, such as warmth or fairness, and those people believe a machine could have. An AI judge was the most or second-most feared role in every country, and more feared than the authors' model predicted. We suspect the issue is accountability: when a decision carries authority over someone's life, people want to know who will answer for it.
Neither study proves that a human will command a permanent premium, and attitudes tend to soften with familiarity. But they suggest a question every business should be testing: where do customers welcome assistance, and where do they still refuse to delegate responsibility? For professional services the answer may decide whether automation deepens a relationship or removes the reason for paying for it. We ask that question of our own work, too.
An AI judge was the most or second-most feared role in every country. We suspect the issue is accountability: when a decision carries authority over someone's life, people want to know who will answer for it.
What this means for the families we serve
None of this makes us pessimistic about artificial intelligence. It makes us particular about how we own it.
First, we look through labels. A family might own chip makers through an index, a data centre fund in a private markets allocation, a loan to a power developer through private credit and a stake in a private firm selling cooling systems. That is four asset classes on a statement and, in practice, perhaps one customer: the capital budgets of a few very large companies.
Second, we ask the same questions of every holding that claims to benefit. Does AI improve its economics once implementation and supervision are paid for? Can it keep the gain, or will competitors with the same tools force it to pass the gain to customers? Does the extra capital earn more than it costs? And how much of the answer is already in the price?
Third, we keep an open mind about where the value will settle. The British railways' greatest beneficiaries were often their users, such as the brewers of Burton upon Trent, whose pale ales reached the whole country once the line arrived. Companies that use AI well, without financing its costliest infrastructure, may likewise prove the better investments, provided their advantage rests on what cheap tools cannot copy: proprietary data, distribution, and the trust that persuades a customer to hand over responsibility.
Britain got its railways and was immeasurably richer for them, yet a good part of the money invested simply vanished in the gap between what investors hoped for and what the lines could earn. AI need not fail for an investment in it to disappoint. Our task is to work out, holding by holding, who will turn its capabilities into durable cash flows, and what will be left for the owner once today's price has been paid.
Sources
Blumenfeld, A., Hazell, J., Lian, C. and Schaab, A. (2026). ‘The Macroeconomic Effect of AI: Sizing the Software Engineering Channel.’ NBER Working Paper, No. 35793, September.
Bruschi, C., D’Acunto, F., De Fiore, F., Sandri, D. and Weber, M. (2026). ‘Will AI Help or Hurt Work and the Economy? What People Around the World Expect.’ CESifo Working Paper, No. 13006, September.
Dong, M., Conway, J. R., Bonnefon, J.-F., Shariff, A. and Rahwan, I. (2026). ‘Fears About Artificial Intelligence Across 20 Countries and Six Domains of Application.’ American Psychologist, 81(1).
Fairlie, R. W. and Wu, J. (2026). ‘The Early Impacts of AI on Employment among Recent College Graduates.’ NBER Working Paper, No. 35796, September.
Kalyani, A. and Li, H. (2026). ‘Is Optimism for Artificial Intelligence Boosting Investment?’ FRBSF Economic Letter, 2026-13, 18 May.
Massenkoff, M. and McCrory, P. (2026). ‘Labor Market Impacts of AI.’ Anthropic Research, March.
Subrahmanyam, S. and Beainy, R. H. (2026). ‘The Real Value of Artificial Intelligence.’ International Maronite Journal, 2(26).
About the author
Nick Perryman is Vice Chairman and Partner at Clarus Global Capital and has 25 years of investment and wealth management experience. He was previously a managing director at UBS where he was head of its international wealth business in London, and before that head of investment products and services which included $18bn under discretionary investment management. He is co-author of the book, Leadership in Wealth: Mastering the opportunities of wealth in your family, firm and society. Email Nick
Important notice
This article is provided for general information, discussion and educational purposes only. It reflects the views of the author at the date of publication and is not intended to constitute, and should not be relied upon as, investment, financial, legal, tax, accounting or other professional advice. It does not constitute an offer, solicitation, recommendation or invitation to buy, sell or hold any investment, financial instrument or service, nor should it be regarded as a personal recommendation or as taking account of the objectives, financial circumstances or needs of any particular person.
