AI may be wildly successful while parts of the infrastructure investment cycle disappoint.
This is the final 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. The first article AI may succeed spectacularly while infrastructure investments fail separated demand growth from capital recovery. The second article AI is not one business. It is two. distinguished AI infrastructure from frontier-model platforms. The third article last week The most important AI infrastructure metric is the utilization rate explained why utilization and investment feedback loops matter. This week’s article asks who captures the value and profits.
AI may become one of the most successful technologies in history. That does not mean all investors in the AI supply chain will make satisfactory returns, or even that all parts of the ecosystem will do so overall in the medium term of 5 to 10 years.
Semiconductor suppliers have already captured enormous value. NVIDIA is the obvious example. Memory, networking, construction, cooling and power-equipment suppliers are also benefiting substantially because these are also the inputs to the current wave infrastructure capacity construction.
Frontier-model platforms could become highly profitable too. It seems most likely that at least a small number of them will, or might instead cash-out by being acquired. If an agentic AI assistant becomes the default consumer or enterprise interface for search, productivity, coding, shopping, travel, customer service or workflow automation, the owner of that platform may influence enormous commercial activity. It might even provide the end-user service cheaply or free while monetizing attention, enterprise integration, cloud usage, advertising, commerce or subscriptions.
That is software-platform business model economics.
Physical infrastructure is different. AI data centers are essential and can be valuable, but the financial upside of a facility is bounded by its capacity, utilization and market pricing. A five-gigawatt campus is impressive, but it is still a capacity-constrained asset. If too much comparable capacity is built, or if prices fall toward marginal cost, first-owner financial returns may disappoint.
This is the central distinction in my recent research paper and financial model. AI infrastructure and frontier-model platforms sit close and are highly interdependent in the same ecosystem, but they do not have the same business models, economics and competitive dynamics.
Chinese and open-weight models complicate the picture further. They may accelerate AI adoption while placing severe downward pressure on pricing. Cheaper models can be excellent for many users and application developers. They can also reduce the scarcity value of proprietary models for some tasks and lower willingness and need to pay for hosted inference.
That is why financial stress testing matters. A model price that appears comfortably above infrastructure cost today may not remain there if competition intensifies, routing improves, open-weight models become good enough and enterprises push back on AI spending.
The important point is not that AI is a bubble. That is too simplistic. AI clearly works, is already very useful and is being extensively adopted. Its capabilities are improving rapidly. Agentic AI is new, and even LLMs have only been in widespread use for a couple of years. Adoption will grow enormously.
The important point is that technological success, user benefits, platform value and infrastructure returns can be very different.
Telecom history illustrates this well. The internet succeeded. Bandwidth demand exploded. Consumers and enterprises benefited enormously. Yet many original infrastructure investors in the boom around the millennium lost money because prices fell too far relative to their invested capital with excess supply and many competitors.
Mobile broadband from 2020 to 2025 provides a more nuanced lesson. Traffic grew massively and revenue yield per gigabyte collapsed, broadly as I forecast in 2011. But market structure was more concentrated than long-haul fiber, with spectrum scarcity limiting the number of nationwide operators. Returns were challenged, but capital destruction was less extreme than in many fiber markets.
AI will have elements of both stories. Some parts of the ecosystem may have software-platform-like economics. Other parts may resemble high-fixed-cost infrastructure. Some data center assets may be enhanced by power generation and grid interconnection scarcity, water access, permitting, superior latency, sovereignty and long-term contracts. Others may be more exposed to competitors, utilization risk and technology obsolescence.
The industry therefore needs sound infrastructure economics. Rampant demand is not sufficient.
Forecasting AI usage over the medium and long terms of 5 years or more is hard and highly speculative at this stage. Forecasting prices, utilization, capital recovery and value capture is harder. But those are the questions that determine who ultimately earns adequate or exceptional returns.
My objective in this series was to bring infrastructure forecasting discipline to AI. My recent RCR Wireless retrospective showed why long-horizon traffic forecasting must be paired with pricing and revenue-yield analysis. The same discipline is now needed for AI data centers, model platforms and the wider ecosystem.
The conclusion is simple.
AI may succeed spectacularly. Some companies may build extraordinary businesses. Consumers and enterprises may obtain cheap or free AI of great utility.
But some investors financing the first wave of AI infrastructure may still be greatly disappointed.
Those outcomes are not contradictory. They would be two sides of the same story.
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.
