AI and America's oldest monopoly problem

The debate over power concentration in the artificial intelligence industry has been growing in technical forums for months, but it rarely gets articulated with concrete historical perspective. An article published this week on Substack titled AI Has America's Oldest Monopoly Problem – Part 1 does exactly that: it connects the current structure of the AI market to patterns from industrial monopolies that the United States already had to dismantle more than a century ago.

The piece circulated on May 5th on Hacker News with minimal initial traction (five points, no comments at publication), but the central argument deserves attention beyond voting metrics.

The argument: critical infrastructure in few hands

The text draws a parallel between the major railroad companies of the 19th century, which controlled the infrastructure on which the entire economy depended, and the handful of labs and compute providers that today control access to large language models at scale. The analogy is not new, but the author develops it with considerable rigor: just as railroad access then determined who could compete in physical markets, access to GPUs, training data, and base models today determines who can compete in digital markets.

The companies that emerge are the usual suspects: Microsoft, Google, Amazon, and to a lesser extent, Meta. Anthropic, the company behind Claude, occupies a peculiar position: it depends on Amazon Web Services investment for compute infrastructure, but maintains a Public Benefit Corporation structure that, at least in theory, distances it from pure shareholder maximization. Whether that is enough to escape monopolistic dynamics is a question the article leaves open.

Why it matters now

The timing is not arbitrary. Over the past twelve months, regulatory signals have accumulated across multiple jurisdictions: the European AI Act is already in gradual implementation, the U.S. Department of Justice has opened preliminary investigations into base model licensing practices, and several American states have introduced specific legislation on concentration in generative AI services.

What the Easy Days analysis adds is the historical framing: Gilded Age monopolies did not form overnight, but rather through exclusivity agreements, control of logistical bottlenecks, and progressive elimination of smaller competitors. The author suggests the pattern repeats today with preferential access contracts for data centers, acceptable use clauses that restrict competitive uses, and concentration of talent in a small number of organizations.

Who should care about this discussion

For developers and integrators, the typical ClaudeWave audience, the impact is concrete and already visible: inference pricing, rate limits, terms of service conditions, and model availability depend on decisions made by three or four companies. Anyone building on third-party APIs assumes a platform risk that in other sectors has led to antitrust proceedings.

For teams working with Claude Code, MCP servers, or agents deployed in production, the practical question is how much provider diversification makes sense and what happens if terms change. It is not paranoia; it is dependency management.

The article is the first part of a series, so concrete proposals, if any, will come later. For now it works better as diagnosis than prescription.

---

Editor's note: The historical comparison is useful precisely because it forces specificity about mechanisms, not just about size. If the second part maintains this level of rigor, it will be a worthwhile series to follow. For now, the discussion of AI concentration needs more structural analysis like this and fewer unsubstantiated alarm headlines.