AI’s Social Returns Could Be Enormous Even If Returns on AI Infrastructure Are Terrible

AI could become one of the most important productivity technologies of our time without delivering comparable returns to the investors who paid the largest share of the infrastructure bill.

Society measures a technology by the value it creates, while an investor has to compare future cash flows with the capital committed to producing them. If AI capabilities improve while the cost of using those capabilities falls sharply, social returns could rise precisely as infrastructure returns come under pressure.

The railways of the nineteenth century transformed trade, industry, cities and human mobility even though many railway companies failed financially.

When an indebted operator collapsed, the tracks, stations and transport networks did not disappear; they could pass to a new owner at a fraction of the original investment cost.

The technological thesis could therefore be completely correct while the investment thesis was wrong. Society won even when the original investor did not.

Demand growth alone does not determine whether AI infrastructure investments will be profitable, because the price customers are willing to pay for each useful unit of AI output matters just as much.

If AI usage increases tenfold while the price of achieving the same useful outcome falls to a fraction of today’s level, AI could spread throughout the economy without infrastructure revenues growing at anything like the same rate.

That would be excellent news for consumers and business users. It could be much less attractive for the financial model behind a massive data-centre investment.

What matters economically is not only that AI models are improving, but also that the cost of producing the same useful result can continue to fall.

Competition between models, cloud platforms, hardware architectures, open alternatives and local computing is pushing the market towards a world in which a company does not need to own an AI factory to benefit from artificial intelligence.

The analogy is electricity: the user ultimately cares about the useful output, not which individual power plant produced it. As intelligence becomes easier to purchase on demand, value may increasingly shift from owning the infrastructure to using it well.

Data centres and AI capacity are being built today on assumptions about future demand, utilisation rates, pricing and the economic life of the underlying assets.

The problem arises if technology improves faster than the payback period built into those assumptions: the same workload may require much less computing power, or customers may be able to move to cheaper models.

The infrastructure itself could still remain useful and heavily utilised. What turns out to be wrong may not be the infrastructure, but the price originally paid for it.

A large IPO can change where a company gets its capital and who ultimately carries the risk, but it cannot change the underlying economics of the business.

At first, the risk may sit with founders and venture investors; later it can move to large institutions and eventually to a much broader group of public-market investors.

The fundamental question remains unchanged: does the business generate enough cash flow relative to the capital tied up in it? If the answer is no, a listing can pass the rope to the next person, but it does not make the rope any lighter.

AI infrastructure is surrounded by a long chain of model developers, cloud providers, chip manufacturers, data-centre operators, energy companies, property owners and financiers, each with its own assumptions about future growth.

During a rapidly expanding market, the entire chain can look attractive until falling unit prices begin to challenge expected returns. At that point, infrastructure may change hands, debt may be restructured and the next owner may operate with a radically different capital cost. For the user, that process could ultimately mean access to even cheaper AI capacity.

AI does not need every major AI company or infrastructure investment of the current cycle to succeed in order to transform the economy.

The internet did not require every internet company of the 1990s to survive, just as the value of the railway network did not depend on every early railway company remaining solvent.

Capital losses, restructurings and asset repricing can even make existing infrastructure economically more useful to the next owner. A technology can therefore spread because it becomes cheap to use, not because its first owners earn exceptional returns.

That is why the most interesting question may not be which of today’s model companies, chip manufacturers or cloud platforms eventually dominates the market. A more attractive position could belong to the company that can take advantage of constantly improving and increasingly inexpensive AI without making its own business model dependent on a single infrastructure provider winning.

In the railway analogy, the best business was not necessarily owning the railway. It could be building a company whose goods could travel on any train — especially if the price of the ticket kept falling year after year.

We may be heading towards a world in which companies, small businesses and individuals have access to more computational capability than ever before while paying steadily less for it. That would be an enormous success for artificial intelligence even if some of today’s data-centre, GPU and AI investments were made on return expectations that prove far too optimistic.

These outcomes are not contradictory; they may be two sides of the same development. AI can win completely while a large share of today’s AI investors lose — and understanding that requires little more than looking at history and putting two and two together.