What past technological revolutions can teach us about where value may move next.
The first AI winners were relatively easy to understand. AI needed advanced semiconductors and computing power, while data centres created demand for electricity, cooling, transformers and cables.
Investors could follow the spending.
The scale is now extraordinary. The Bank for International Settlements describes the AI build-out as one of the largest technology-driven investment booms in US history.
Yet the product of that investment is becoming dramatically cheaper. According to the Stanford AI Index, the cost of querying a model performing at roughly GPT-3.5 level fell from $20 per million tokens in November 2022 to $0.07 by October 2024 — more than 280 times cheaper.
Meanwhile, AI adoption is spreading rapidly. The 2026 Stanford AI Index reports that 88% of surveyed organisations used AI in at least one business function in 2025.
This creates an unusual combination: enormous investment, falling costs and widespread adoption.
But there is an important distinction between technological progress and usable capacity. A more powerful GPU does not immediately create more compute that customers can actually access. It still needs to be installed, powered, cooled and connected, often within infrastructure designed years earlier. New generations can therefore become dramatically more efficient while existing hardware retains economic value because immediately available compute remains scarce.
Technology can become cheaper much faster than the physical capacity required to deliver it becomes abundant.
The first winners supplied what AI needed. Finding the next AI winners may require looking somewhere else: at what happens when intelligence becomes cheap enough to change the economics of other businesses.
History offers some useful clues.
What Previous Technology Booms Can Teach Us
Railways: A Great Technology Can Still Be a Poor Investment
Britain’s Railway Mania transformed transportation, but correctly identifying that revolution did not guarantee attractive returns. Capital flooded into the sector, competing lines were financed and capacity expanded.
The parallel with AI is worth considering. In 2026, the Bank for International Settlements modelled the current AI investment race. Under its baseline assumptions, investment reaches around 50% above the economically efficient level. With less elastic demand, it can approach three times that level.
When companies believe a technology could determine future market leadership, not investing can appear more dangerous than investing too much. Rational decisions by individual companies can therefore create overcapacity across an industry.
The lesson is not that AI infrastructure is a poor investment. It is that being right about a technological revolution is not enough; investors must also be right about the return on the capital used to build it.
I explored the physical side of this investment cycle in How to Invest in the AI Boom: The Bottlenecks Markets May Still Be Underestimating.
Electrification: The Value Moved Beyond Electricity
Electricity offers a different lesson.
Factories did not capture its full potential simply by replacing steam engines with electric motors. Larger gains appeared when production was reorganised around electric power.
An NBER study of US manufacturing between 1890 and 1940 found that electrification produced rapid and persistent productivity gains, accompanied by capital deepening and organisational change.
Eventually, access to electricity stopped differentiating industrial companies. What mattered was how effectively they used it.
AI may follow a similar path. The cost of GPT-3.5-level inference fell more than 280-fold in less than two years.
But the transition may not happen as quickly as the improvement in the technology itself. Electricity became more useful long before generation, transmission and distribution became abundant everywhere. AI could follow the same pattern: intelligence can become cheaper while the physical infrastructure required to deliver it remains constrained.
That changes the investment question from who supplies intelligence? to what becomes more valuable when intelligence becomes cheap?
The Internet: Someone Else’s Capex Can Become Your Cheap Input
The internet offers perhaps the most relevant precedent.
During the 1990s, investors correctly anticipated huge demand for networks, fibre and computing infrastructure. Yet the dot-com collapse destroyed enormous amounts of shareholder capital.
The infrastructure survived and became a cheaper input for another generation of businesses. Search, e-commerce, digital advertising, social networks and cloud computing could build on investments already made by others.
Overinvestment in one layer can improve the economics of the layer above it.
If today’s AI investment race eventually produces abundant usable compute and cheaper models, some future winners may therefore be companies buying cheap intelligence rather than selling expensive intelligence.
Eventually matters. Technological capability can improve much faster than physical capacity can be deployed, allowing falling intelligence costs and infrastructure scarcity to coexist. The emergence of new productivity winners therefore does not require today’s bottlenecks to disappear first. The two phases can overlap, potentially for years.
These examples do not predict what will happen with AI. They suggest a mechanism worth testing:
Scarcity attracts capital → usable capacity expands → bottlenecks migrate → delivered technology becomes cheaper → the profit pool can move.
The question is where it moves — and how long the existing scarcity lasts before it does.
AI Is Creating Value. Who Is Capturing It?
AI is already producing measurable economic value.
The Stanford AI Index estimates that US consumer surplus from generative AI reached an annualised $172 billion in early 2026, up 54% in one year.
Productivity gains are also visible. An NBER study of 5,179 customer-support agents found that generative AI increased productivity by around 14% on average, with gains of roughly 34% among novice and lower-skilled workers.
However, value creation and shareholder value are not the same thing.
Recent research illustrates the gap:
| AI adoption and impact | Share |
|---|---|
| Organisations using AI | 88% |
| Reporting productivity benefits | ~80% |
| Reporting some EBIT impact | 37% |
| McKinsey AI “high performers” | ~6% |
The first figure comes from Stanford’s 2026 AI Index. The other three are reported in McKinsey’s 2026 State of AI survey. They should therefore not be interpreted as a single conversion funnel.
The contrast is nevertheless striking. AI use is becoming widespread, while measurable financial outperformance remains much more concentrated.
PwC finds a similar pattern. In its study of 1,217 organisations across 25 sectors, the top 20% captured 74% of measured AI-driven returns.
More importantly, those leaders are not simply using more AI. PwC finds that they are more likely to redesign workflows and use AI to pursue growth and rethink their business models.
This echoes electrification: adopting the technology is not the same as reorganising a business around it.
Where Does the Productivity Go?
Imagine a company selling a service for £120 that costs £100 to provide. AI reduces the cost to £80, potentially doubling profit from £20 to £40.
But competitors can use similar technology. If competition pushes the price down to £105, profit rises to only £25. Most of the productivity gain has gone to customers.
This is no longer purely theoretical. The Bank of England’s July 2026 survey of business conditions reports productivity improvements from AI across software development, finance, administration, customer service and professional services. But it also finds that lower unit costs are creating pressure to reduce prices for routine tasks that can increasingly be automated.
Research from the Federal Reserve Bank of San Francisco points to another possible effect. Large incumbents can initially gain market share as they adopt AI, but that advantage can weaken as the technology spreads to smaller companies and new entrants.
AI can therefore create enormous economic value without all of it reaching shareholders.
For investors, the central question is not who benefits from AI, but who keeps enough of the benefit to change shareholder returns.
How I Would Look for the Next AI Winners
Four questions follow from this analysis.
Can cheaper intelligence change the economics?
The question is not how much AI a company uses. It is whether cheaper intelligence can materially affect revenue, margins, asset utilisation, capital efficiency or important decisions.
The strongest opportunities may therefore emerge in businesses where intelligence is currently expensive, scarce or embedded in labour-intensive processes.
Can we see the benefit?
AI announcements are easy. Economic impact is harder.
Over time, the benefit should appear in revenue growth, margins, free cash flow, return on invested capital or market share.
This distinction matters because markets price expectations rather than simply rewarding good companies. I explored that idea in Why Markets Reward Surprises… Not the Best Companies.
Does the company already have an advantage?
The strongest candidates should control something difficult to reproduce, such as proprietary data, distribution, customer relationships, network effects, regulatory rights, trusted brands, embedded workflows or scarce physical assets.
AI may not be the moat. It can amplify the moat.
For physical bottlenecks, however, I would now add another dimension: time.
A scarcity does not need to be permanent to create substantial economic value. What matters is whether supply can respond before the owner of the scarce capacity has earned attractive returns on the capital committed to it.
That means asking not simply whether a bottleneck will disappear, but how long it will take to disappear.
This distinction matters because high prices attract investment and technological progress can increase the amount of useful output produced from the same physical infrastructure. Both eventually work against scarcity rents.
Who keeps the gain?
If AI saves a company $1 billion but competition returns $900 million to customers through lower prices, the technology has created economic value without creating the same amount of shareholder value.
The investment sequence therefore becomes:
Economic impact → Incremental ROIC → Ability to retain the gain → Duration → Supply response → Reinvestment → Valuation → Expected shareholder return
For infrastructure businesses, duration and supply response become particularly important. The relevant question is not whether today’s scarcity lasts forever, but whether it lasts long enough to generate attractive returns before additional capacity or technological efficiency erodes it.
A great company can still be a poor investment if the market has already priced in an even better future.
So Where Are the Next AI Winners?
History cannot tell us which companies will win. But it can suggest where to look.
I would look beyond the obvious AI suppliers towards businesses where intelligence is an important cost or constraint, where cheaper AI can materially improve the economics, and where an existing advantage allows the company to keep part of that gain.
But I would no longer frame this as a simple transition from infrastructure winners to productivity winners.
The two can coexist.
Physical infrastructure can continue capturing value while usable compute remains scarce, even as falling intelligence costs create new opportunities elsewhere. The balance changes as capacity expands, bottlenecks migrate and competition redistributes the gains.
This could include businesses where better intelligence improves pricing, design, asset utilisation, productivity, customer conversion or high-value decisions. The sector matters less than the economic sensitivity to cheaper intelligence and the ability to retain the resulting benefit.
The first phase of AI investing was largely about understanding what AI needed.
The next may be about understanding what AI changes — while recognising that the transition between the two may take much longer than the technology cycle itself suggests.
And the most interesting winner may not look like an AI company at all.
This article is published for general informational and educational purposes and reflects my personal research and opinions. References to companies, sectors or securities are intended to illustrate economic concepts and should not be interpreted as investment recommendations or an assessment of suitability for any investor. Information is based on publicly available sources and may become outdated. Investing involves risk, including loss of capital. Any relevant interests in securities discussed should be disclosed separately.
Laisser un commentaire