AI May Consolidate Internet Power, Not Expand Access

The Tech Policy Press piece sets up a tidy duel: abundance versus scarcity, and whoever controls the tech wins the internet. Here's the thing — framing this as a choice between more stuff and less stuff misses the real conflict: who governs the rules of generation, who sets incentives for distribution, and who enforces the boundaries around platform power. The headline is metaphysical; the real battle is legal, architectural, and institutional.

Let’s start with what the article gets right: asking who will “control” the internet after AI is the right instinct. But it treats control as if it’s a single on/off switch. It isn’t. Control fractures into at least three levers: compute and infrastructure; datasets and training pipelines; and the consumer interface — the apps and services people actually use. Each lever channels power in different directions and on different timelines.

Compute concentrates. Build or rent massive model-serving clusters and you don’t just get scale; you get veto power over who else gets to play. Data flows mirror existing gatekeepers; whoever curates and filters at scale shapes what models consider “normal.” And user interfaces quietly govern perception; a search bar or chat box with specific ranking and refusal rules is governance in plain sight.

The article hints at centralization risks, but yeah, no, that’s the softest version of the argument if it isn’t pinned to these levers. Policies aimed only at “abundance” — more models, more APIs, more fine-tuning options — without addressing where the servers sit, what data is packaged, and whose UI patterns dominate will invite concentration back in, just dressed in open-sourcy clothing.

Antitrust gets name-checked, and it should. Breaking up or restraining dominant players is table stakes. But antitrust is necessary, not sufficient. You also need rules for interoperability, for public-interest dataset access, and for auditability of training processes — so we can examine how systems are built, not just how they behave after the fact. The public sector could take a utilities-regulator posture for key pieces of the stack: not controlling content, but ensuring non-discriminatory access to essential compute and core data plumbing. That’s not a five-year plan for Silicon Valley; it’s the kind of boring constraint that keeps any one actor from locking down the stack from silicon to search results.

If this all sounds vaguely familiar, that’s because we’ve run this play before. Think of how the early internet rode on top of common protocols like TCP/IP and HTTP — no one company owned the rails, which is why a couple of students could build Google in a garage. When mobile app stores arrived, we quietly traded protocol-level openness for curated platforms and 30% app taxes. AI risks being the App Store era on steroids unless policymakers separate infrastructure from interface and keep the lowest layers as neutral as possible.

The article also flirts with a seductive line: abundance equals decentralization — more models mean more voices. Tempting story. Funny thing is decentralization requires more than multiplicity; it needs sustained funding, governance structures, and incentives that favor interoperability over winner-take-all lock-in.

Open-source models and community-run infrastructure are vital counterweights, but they are not a self-healing antidote to platform power. They need legal clarity around data collection, access to compute that isn’t buried under punitive pricing or contractual handcuffs, and norms that resist co-option when fast-scaling capital shows up with growth demands that don’t match community values. If we treat open models as an inevitable balancing force, we downplay just how fragile those ecosystems really are.

A second underdeveloped point in the original piece is how uneven “abundance” will be. Policy debates in one jurisdiction don’t automatically propagate across languages and legal regimes. Local content rules, language data scarcity, and different threat models mean that any global AI stack will land asymmetrically. Some regions end up as raw material suppliers for training corpora; others become pure consumers of closed, packaged interfaces, with little say in how those systems are tuned.

When critics argue that centralization is necessary — because big players fund the research, manage safety testing, and deploy reliably at scale — they’re not entirely wrong. Concentration has produced systems that are both useful and, in some cases, safer than the chaos of a thousand ad-hoc deployments. But that doesn’t lock us into a permanent monopoly bargain. Public funding for open foundation models, interoperability requirements baked into regulation, and enforceable rights to move models and data across providers can keep the safety work while chipping away at single points of failure. Safety and competition don’t live on opposite sides of the chessboard; they just depend on intentional institutional design rather than faith that market gravity will fix everything.

One thing the debate often misses is how fast “interfaces” themselves are consolidating. If AI agents end up embedded in operating systems, productivity suites, and browsers, then whoever owns those touchpoints will shape which models matter — no matter how many alternatives technically exist. We’ve seen this with search defaults and mobile browsers; expect the same trench warfare over which AI assistant boots up first. The economics of default choice haven’t changed just because the interface now speaks in paragraphs.

I keep thinking about Ursula Le Guin’s quiet point in “The Dispossessed”: the structure of your institutions quietly teaches people what’s possible. The same goes for AI-era internet governance. If we design for abundance without dislodging the old chokepoints, we’re not escaping scarcity; we’re just reskinning it with better autocomplete.

Tech Policy Press is asking who controls the internet after AI; the more interesting tell will be whose rules show up in the training data, the data-center contracts, and the default prompts long before anyone writes the next headline.