The AI Oligopoly Builds Itself

With capital expenditure commitments reaching semi-sovereign levels, the leading AI labs are no longer startups in any meaningful sense. They are infrastructure. And infrastructure, history shows, gets regulated.

By Monexus Staff Writerglobal5-minute read5 May 2026☆ Save ↗ Share ⎙ Print

OpenAI plans to spend $50 billion on computing power this year. Let that number sit for a moment. It exceeds the GDP of more than a hundred sovereign states. It is not a rounding error in a venture pitch deck — it is a capital commitment that defines an industry, forecloses competitors, and invites a structural question that the prevailing cheerleading coverage refuses to ask: what happens to a technology sector when its leading firms stop behaving like companies and start behaving like public utilities in a hurry to lock in?

The question is not hypothetical. It is the logical terminus of a trajectory made concrete this week by multiple signals. OpenAI and Anthropic are each in acquisition talks for AI services firms, according to Reuters, consolidating their positions at both the model layer and the deployment layer simultaneously. Anthropic separately unveiled ten new AI agents designed for banks and insurers — a deliberate pivot toward the kind of regulated, high-margin enterprise customers who sign multi-year contracts and resist displacement. OpenAI, meanwhile, carries a 68 percent implied probability on Polymarket that it launches a social network this year, an escalation from software provider to platform operator that would give it direct access to the feeds and behavioural data that make advertising economics work.

The second-order logic is not complicated. If you are spending $50 billion a year on compute, you need revenue at a scale that consumer subscriptions alone will not deliver. A social network — or a suite of enterprise agents that sit inside the workflows of the world's largest financial institutions — changes the unit economics. It also changes the leverage. Regulators who might have once worried about a chatbot now have to contend with infrastructure embedded in credit decisions, insurance underwriting, and the flows of attention across a social graph.

Anthropic's trajectory points in a related but distinct direction. The company is building agents for regulated industries not because it wants to compete on consumer interfaces, but because the regulated-industry path is more defensible. Banks do not switch core systems on a whim. Insurers do not rip out underwriting models every eighteen months. A model that sits inside a Goldman Sachs or a JPMorgan, embedded in compliance workflows and loan origination processes, is a model that generates recurring revenue and resists the kind of disruptive displacement that upended consumer search. There is a reason Anthropic is reportedly angling for removal of its "supply chain risk" designation in federal procurement frameworks — that designation is a regulatory moat that works both ways, protecting from competition but also limiting access to the government contracts that would accelerate growth.

The venture capital layer reinforces the pattern. Andreessen Horowitz's announcement this week of a $2.2 billion crypto fund explicitly earmarked for projects linking digital assets with AI and traditional finance is not incidental. It is a bet that the financial infrastructure of the AI economy — the settlement layers, the tokenisation of compute, the smart contracts that will govern agent-to-agent transactions — will itself become a major asset class. The fund will not compete with OpenAI or Anthropic at the model layer. It will colonise the plumbing underneath them, collecting fees at every settlement, every computation credit, every cross-border payment mediated by an AI agent.

What this adds up to, structurally, is an industry that is maturing — or calcifying — into something that looks less like a competitive technology market and more like a utilities landscape. In utilities, the pattern is well-documented: the firms that build the physical infrastructure set the terms for everyone downstream. The发电厂 does not negotiate with the factories that depend on it. The railroad sets freight rates; the shipper pays them. The question for AI policy is whether the same gravitational logic will hold — whether the $50 billion compute commitment and the enterprise contracts and the platform plays all converge into a set of firms so capital-intensive and so deeply embedded in critical infrastructure that competition becomes structurally impossible rather than merely difficult.

The counterargument — the one the industry makes and that some analysts accept — is that AI is not a utility in the traditional sense. The marginal cost of an additional query tends toward zero. Models improve and commoditise faster than hardware depreciates. Open-source alternatives constrain pricing power. This counterargument is not wrong, exactly. It is incomplete. It describes the market as it exists today, with current levels of capital commitment, current model architectures, and current regulatory frameworks. It does not describe the market that $50 billion a year in compute investment, combined with enterprise agent deployments inside the world's most risk-averse institutions, is actively constructing.

The stakes are not abstract. If the leading AI labs succeed in locking in enterprise deployments at the scale their capital commitments require, the next generation of AI researchers and builders faces a choice: join one of the incumbents or compete on a playing field where the other team owns the stadium, the referee, and the broadcast rights. A $2.2 billion crypto fund that routes investment toward the infrastructure layer of the AI economy does not widen that field. It deepens the moat. The industry calls this ecosystem investment. The regulatory history of telecommunications, of railroads, of broadband, calls it something else: the moment when an industry stops being a market and starts being a commons with a gate.

That framing will strike many in the industry as unfair. The labs are building genuine capability. The products work. The research contributions are real. But the scale of the capital, the depth of the enterprise integration, and the velocity of the acquisition activity this week are not the behaviours of firms optimising for product-market fit in a competitive market. They are the behaviours of firms building durable structural advantages that, once in place, are extraordinarily difficult to reverse. The policy conversation — one that regulators, competition authorities, and the venture ecosystem itself will eventually have to have — is not about whether these firms have earned their position. They have. The question is what a technology sector looks like when the infrastructure layer and the application layer converge under the same ownership, and who gets to build on top of it.

This piece draws on Reuters reporting on acquisition activity among leading AI labs and Polymarket market data on OpenAI's near-term product and capability projections.

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

http://reut.rs/49gnI4t