Authors: Arvind Narayanan and Akash Kapur
The goal of this article is to move beyond the debate over whether “AI is in a bubble.” We will do so in two ways: by clearly distinguishing the current financial picture from the long-term question of who will capture value, and by recognizing that these labs are not limited to serving as model providers. They can—and already are—moving upstream in the technology stack. This will likely help them escape the commoditization trap, but it also raises new concerns: customer lock-in and weakened competition.
Akash Kapur is a visiting scholar at Princeton University and a senior fellow at New America. He is not related to Sayash Kapoor.
As leading AI companies continue to invest heavily in computing capacity and race toward blockbuster IPOs, serious questions about their business models remain. How will these companies—and the vast ecosystem of chipmakers, hyperscale cloud providers, and infrastructure partners that depend on them—recoup the 4 to 8 trillion dollars expected to be invested in AI infrastructure by the early 2030s?
The current debate has split into two camps: critics and supporters. Critics point to mounting losses, the gap between capital expenditures and revenue, and reports of enormous cash burn at leading labs. Supporters emphasize rapidly growing revenue, enterprise adoption, and milestones such as Anthropic’s first profitable quarter. Both sides have valid points. But both are looking in the wrong place—at the same quarterly reports and the same short-term perspective on an industry that remains in a state of dramatic flux.
Over the past several months, we have been thinking about the nature and sustainability of the AI business, and we have ultimately reached a conclusion that differs from most existing commentary. Today, a significant share of AI companies’ revenue comes from charging for inference services. But the conditions surrounding inference for frontier models make it an unusually difficult business to sustain. There is little differentiation between models, leading labs have adopted similar capital structures, switching costs are low, and prices can be adjusted freely. Taken together, these factors appear to create a commoditization trap, posing a genuine challenge to building highly profitable or even profitable businesses.
At the same time, we believe the industry is still in a transitional phase, and that its mature structure will look very different. Combining historical evidence with economic theory, we argue that in this equilibrium, competition will likely drive model inference prices down toward the marginal cost of generating tokens, leaving little room for the model layer to earn durable profits. However, this does not mean that the AI business is inherently unviable. The same analysis also points to a path forward for AI labs and leads to the central argument of this article:
The most likely path to durable profitability for labs is not to rely on the foundational layers—chips, data centers, and models—that have attracted the bulk of investment to date. Instead, they will move upward in the technology stack through vertical integration, embedded enterprise deployments, and the deliberate creation of switching costs and other “moats.”
The labs’ strategy of capturing value by moving upward in the technology stack—many elements of which draw on practices from enterprise software—is already under way. Beyond the sustainability of today’s AI ecosystem, these strategies raise broader societal questions involving competition, innovation, and the overall distribution of economic and political power. So far, public discussion of AI has exhibited a rather contradictory dichotomy: on the one hand, there are concerns about monopolistic concentration and unchecked market power; on the other, low switching costs and the relative interchangeability of models seem to run counter to those concerns. But if our assessment is correct—that labs will increasingly move upward in the technology stack—then we should take market concentration and competition seriously now, rather than waiting until lock-in effects begin to emerge before taking action. We will return to these broader questions in the conclusion and explore many of the topics raised in this article in greater depth in an upcoming paper.
Historical analysis: Infrastructure layers rarely capture the value they create
The framework of “AI as Normal Technology” seeks, where applicable, to draw lessons from transformative technologies of the past. We believe AI is subject to many of the same dynamics related to investment, competition, and value capture that shaped previous waves of technological innovation. Accordingly, as part of our research, we adopted a broader historical perspective to examine the challenge of capturing value from AI and the ways labs might respond.
We examined historical cases from six capital-intensive infrastructure industries: railroads, electricity, telecommunications and fiber optics, cloud computing, semiconductor manufacturing, and commercial aviation. We argue that AI today shares many characteristics with infrastructure: enormous capital requirements, low marginal costs, and a commoditized product that is relatively disconnected from the applications that ultimately create value. This makes infrastructure a notoriously difficult industry.
At the same time, AI is also software, and the software industry has historically been highly profitable, with exceptionally high margins. The goal in this industry is to achieve software-like margins. We therefore also examined the value-add and lock-in strategies of software as a service (SaaS), and analyzed whether AI could replicate them. Our argument is that the sustainability and value capture of AI companies will depend to a large extent on whether they can successfully migrate from the first set of infrastructure attributes to the second set of enterprise software attributes.
Overall, our historical analysis yields three instructive lessons for AI.
First, infrastructure providers rarely capture the value they create. In railroads, electricity, telecommunications, and aviation, the companies that built capacity ultimately ended up with thin margins because of competition, regulation, or commoditization. In many cases, they were wiped out entirely. During the telecommunications and fiber-optics construction boom of the late 1990s, infrastructure capacity grew 186,000-fold in seven years, prices collapsed, and approximately $2 trillion in market value was erased. The value created by infrastructure primarily flowed to the industries and applications built on top of it. Commercial aviation destroyed investors’ capital for eight consecutive decades, with typical net margins of just 2%–4%, often below the cost of capital—even though businesses across industries benefited from the globalized economy.
We believe that AI, at least in its current form dominated by hyperscale cloud providers, risks falling into the same commoditization trap that plagued so many previous infrastructure builders. Carlota Perez has identified a paradoxical phenomenon: those who …