Why Competition?

This book examines competition dynamics and power concentration in artificial intelligence, with a focus on computation and the concentration of computational resources.

Overview

Introduction covers:

• Extreme concentration of value: AI is characterized by extreme concentration of value. 35 percent of S&P 500 (and 16 percent of worldwide capitalization) is made up of its top 7 companies (21.5 trillions as of 12/17/25), which are all heavily invested in AI and/or compute. (Tesla is a bit of an exception but it is tied to embodied AI and xAI through Elon Musk, Apple is invested in personal rather than Cloud compute, NVIDIA compute more than AI services, then Alphabet both, Amazon both, Meta both, Microsoft both)

• Market value growth: Over the last 10 years, the total market value of these companies has grown by a factor of 10, compared to a factor of 2 for the total capitalization of stock exchanges around the world. A linear evolution along these lines would see the "magnificent seven" account for half of the world's total market capitalization by 2031.

• AI ubiquity: While this concentration is driven by a complex combination of market, societal, and regulatory evolutions over the last two decades, the years since the release of ChatGPT have seen the concept of "Artificial Intelligence" become increasingly ubiquitous in these three areas - especially driven by the companies under consideration.

• Defining AI through scale: "Artificial Intelligence" is a very diffuse concept, but increasingly defined by scale of data and compute use. In particular, AI companies and other proponents of scaling use the necessary high cost of training so-called "frontier" models and the running costs of models without amortization as a reason why the market would inherently need to be dominated by a few actors.

• Components of AI performance: The components of AI performance are typically advertised as "compute, data, and talent". Of these three, compute is the most concentrated. Data is ubiquitous, and while the largest tech companies do centralize access, it remains created at the level of the users. Finally, talent is both mobile and accessible in less-resourced organization, as evidenced by the breadth and quality of open models.

• Research questions: Thus, in order to assess whether AI is bound to remain an inherent driver of further market concentration, we dive further into the role of compute in this book. If indeed concentrated access to compute is a prerequisite to development of commercially viable versions of the technology, we must accept some form of concentration within a few actors, and must tailor economic and market governance to mitigating the negative effects of this state. If however compute is less of an inherently concentrating factor than advertised by these companies, we may instead be able to push policies that help diversify the entire AI ecosystem.

• Market dynamics: In particular, not all versions of AI are as computationally intensive; and one would assume that in a fluid market demand would shift to either proprietary systems developed and maintained by their users or cheaper alternatives. The current domination of the market by a few companies raises the question of whether they are acting as oligopoly actors pushing in a common direction at the expense of others, or whether they are supported in their position by their control of a resource or mechanisms other than compute.

• Research approach: We make progress on these questions by investigating:

  • BACKGROUND: the role in resource monopolization in competition analysis, including alternative drivers of concentration that may better explain both the concentrated space of the market (vertical integration) and the unbalanced state of the price of AI (common ownership of drivers of computationally intensive AI by a few actors, assuming a balanced price would be driven way down by the less intensive options)
  • the cost of AI systems, both training and running, under different metrics (market cost of hosting infrastructure, subscriptions, token cost for hosted solutions) and technology types (large models, fine-tuned small and tiny models)
  • the demand for and concerns about AI expressed by companies across the entire S&P 500 in their earning calls and 10K filings up to FY2024. In particular, investigating whether current demands account for the scale of investment in compute resources, and whether companies have resources to develop their own AI solutions without relying on supply from the largest actors; which would help them escape monopoly prices imposed by the larger actors.

• Preliminary findings: We find (from preliminary numbers so far) that while the interest for and concerns about AI have sharply risen across S&P 500 companies, these same companies do have the resources to meet their own needs without paying inflated prices to the centralizing actors.