Radio began as an experimental medium. By the late 1920s, networks were linking stations, advertisers were funding national programming and regulators were assigning scarce spectrum.

NBC formed in 1926. By early 1927 it operated the Red and Blue networks. The Radio Act of 1927 created the Federal Radio Commission after interference had made the existing system difficult to sustain. Regulation did more than constrain a market. It helped define who could broadcast inside it.

The pattern appears across communications industries. A period of technical experimentation produces new entrants. Distribution develops bottlenecks. Scale improves economics. Policy then formalises, restrains or sometimes breaks apart the structure that emerged.

This is recurrence, not destiny.

Radio is the clearest parallel

The useful comparison between radio and AI concerns the transition from experimental capability to networked distribution.

Early radio had many experimenters, local stations and competing technical practices. Networks changed the economics. A station connected to national programming could offer audiences and advertisers something a local operator could not reproduce alone. Ownership of distribution and programming became mutually reinforcing.

By 1939, the Library of Congress records NBC supplying programming to 171 stations. That reach created value. The FCC’s chain-broadcasting rules later led to the Blue network divestiture.

The analogy stops at spectrum. Radio frequencies are physically scarce and mutually interfering. AI software can be copied, models can be downloaded and applications can reach users over an existing global network. AI’s scarcity sits further upstream in chips, energy, cloud capacity, capital and some forms of expertise.

The common mechanism is a bottleneck that improves the economics of everything attached to it.

Concentration takes different forms

Telephony concentrated through a vertically integrated network. Control of the local loop affected long-distance and equipment competition. The antitrust decree required AT&T to divest its local operating companies into seven regional groups.

The breakup expanded long-distance competition. It did not end concentration everywhere. Local incumbents retained power, and Congress returned to the problem in the Telecommunications Act of 1996.

Television inherited much of radio’s network structure, spectrum licensing and corporate base. Cable added local franchises, infrastructure costs, programming leverage and economies of scale. These were related histories, not independent proof that every medium follows one law.

Policy can also change market structure without changing the apparent number of outlets. Ownership rules affect who controls distribution, not merely how many stations, channels or imprints remain visible.

Publishing shows a narrower version. It is inaccurate to say five firms control all publishing. Markets differ by format, genre and rights. A US court blocked Penguin Random House’s proposed acquisition of Simon & Schuster after a Justice Department challenge focused on bargaining power over anticipated top-selling books.

The bottleneck there was not printing. It was bargaining power over a particular class of authors and books.

The historical lesson is specific. Concentration forms where scale, distribution or control of a scarce input changes the terms for everyone else.

The open internet was institutionally unusual

The internet did not begin as a normal commercial network.

Government-funded research supported its foundational development. Universities and research institutions implemented protocols whose specifications were publicly available. The NSFNET backbone connected roughly 2,000 computers in 1986 and more than two million by 1993 before commercial services took over and the backbone was decommissioned in 1995.

Open protocols let independent teams build compatible systems. Publicly available code made working implementations easier to inspect and extend. The RFC process turned argument and implementation experience into shared technical records.

None of this made the internet free of power. Carriers controlled infrastructure. Platforms later controlled discovery, identity, advertising and mobile distribution. States shaped access and surveillance. The early commercial internet still gave a small operator an unusual ability to publish or offer a service without negotiating with one central network owner.

That low-entry phase was an institutional achievement. Public funding, academic practice, open specifications and commercial competition happened to reinforce one another.

Calling it an accident understates the work. Calling it inevitable erases the choices.

AI begins from a different position

Modern AI grew from decades of public and academic research, but its frontier deployment became commercial early. Stanford’s 2026 AI Index tracks industry as the source of most recent notable models. Advanced compute and cloud capacity already sit in concentrated markets.

That gives AI a more centralised starting core than the early internet protocol layer.

Its edge is less settled. Open-weight models create alternatives at the model layer. Open protocols can reduce integration friction. Public compute programmes can widen access. Domain applications can compete through workflow knowledge rather than frontier training.

AI therefore resembles radio only in part. Networks are forming around expensive capability and distribution. The software remains more replicable, and the standards have not finished setting.

This is why “consolidation has already won” is as weak as “open source will automatically prevent it.” Different layers can settle differently.

Markets can reopen

Communications history contains reversals as well as concentration.

Antitrust separated the Bell System and forced NBC to divest a network. New technical standards created room for new devices and services. Public broadcasting added institutions with different incentives. The Web let publishers reach readers directly before platforms consolidated much of that discovery again.

Reopening rarely comes from one intervention. It usually combines a technical alternative with institutional support and an economic model that can survive.

For AI, that means asking concrete questions. Can workloads move between providers without losing evaluations and operating history? Are protocols implemented by multiple independent teams? Can public or smaller institutions obtain enough compute to remain credible? Do customers retain authority over identity, policy and evidence?

The answer will vary by layer. Chip fabrication may remain concentrated while model serving fragments. Model access may become cheap while distribution stays controlled. Open interfaces may coexist with closed identity systems.

The pattern is a cycle, not a final state.

The norm can be changed

Consolidation recurs because scale and bottlenecks recur. It becomes durable when the surrounding institutions accept the resulting dependency as natural.

The internet’s open phase shows another possibility. Public investment, open standards, independent implementations and competition policy can make alternatives credible before a market closes around them.

AI will not repeat radio, telephony or the Web exactly. It will inherit the same question.

Which bottlenecks are unavoidable, and which ones have simply gone unchallenged?