Demand growth and capital recovery are different questions.

This is the first article in a four-part weekly series. This series is based on my new 12,000-word research paper, The Economics of AI Infrastructure, and the financial model used in that analysis. It also builds on my recent RCR Wireless retrospective, Fifteen years on, how did that 1,000x mobile broadband forecast fare?, in which I compared my 2011 long-term mobile broadband forecast with outcomes through 2025. My new AI analysis focuses on prospects for financial capital recovery given the vast sums being invested, with various sensitivities including extent of demand growth, AI pricing, utilization rates versus capacity growth and costs in the data center.

AI is generating one of the largest infrastructure investment booms in modern technology history. Data center projects are now measured in gigawatts rather than megawatts. Utilities, grid operators, chipmakers, neoclouds, hyperscalers and governments are all planning around vast increases in compute demand.

Commentary often asks whether AI demand will grow. Of course AI demand will grow—enormously. The harder question is whether investors can recover the capital now being committed. McKinsey estimates that global data centers may require $6.7 trillion of capital expenditure by 2030, including $5.2 trillion for data centers equipped for AI processing loads and $1.5 trillion for traditional IT workloads.

History shows how stakeholder outcomes can differ wildly. Railroads transformed economies while bankrupting many investors. The English Channel Tunnel remains a vital link between Britain and continental Europe even though its original financial structure failed. In the telecom boom around the millennium, transoceanic bandwidth demand grew enormously. Yet many first owners of fiber infrastructure lost money because prices fell too far, too fast, relative to invested capital.

That distinction was central to my own forecasting work. In 2000, I correctly forecast steep and persistent transatlantic fiber bandwidth price declines alongside rapid demand growth. In 2011, I forecast mobile broadband traffic and revenue yield per gigabyte out to 2025. My data traffic exponential growth forecast was right on track for 12 years, before actual growth rates slowed somewhat since then. My pricing forecast there was also a crucial insight: with that volume growth, mobile operators yield orders of magnitude lower revenue per gigabyte of user data.

AI infrastructure may face a similar economic test. A data center, GPU cluster, grid connection or permitted site can remain useful even if the first capital structure fails. The asset does not become worthless. Its cost basis changes. A second owner can earn acceptable returns at prices that would never have supported the first owner’s investment.

This is not a prediction that AI will fail. On the contrary. AI may become one of the most successful technologies in history. Consumers and enterprises may benefit enormously, including from periods of over-investment, particularly if that results in low prices. Some software-platform owners and frontier-model developers may become extraordinarily profitable.

The risk is that some of those benefits will not accrue to the investors financing the first wave of physical infrastructure.

The lesson from fiber is that demand, supply and price must be forecast together so that capital recovery can be properly assessed. Usage growth could be massive while revenue growth is inadequate and returns disappoint.

For AI, that means asking a different set of questions. How much capacity is being built? What proportion of that will be utilized under various demand growth scenarios and expectations? How quickly will hardware become obsolete? How far will inference prices fall? How much of AI value will be captured by model platforms, applications and end users rather than infrastructure owners?

Next week’s article in this series turns to a common limitation in current AI commentary: Implicitly regarding AI service as one business. It is not. Physical AI infrastructure and frontier-model platforms have very different operations, economics, business models and competitive dynamics. 

Keith Mallinson, founder of WiseHarbor, has more than 30 years of experience in the ICT sector as a technical, commercial and financial research analyst, consultant, and testifying expert witness