Intelligence may become abundant. Control of the systems around it may not.

AGI may already be here under some definitions. It may arrive soon under others. Or it may never arrive as one coherent technical threshold. This article does not depend on choosing among those futures.

Many threshold-centred arguments about artificial general intelligence place too much explanatory weight on capability. They treat AGI as the point after which the important questions answer themselves.

One side expects abundance, scientific breakthroughs and explosive economic growth. Another points to brittle reasoning, weak benchmarks and the distance between a fluent model and a reliable worker. Capability forecasts can matter without resolving who will control deployment, which institutions will grant authority, or where the value will go.

Capability progress is real. The illusion concerns what that progress settles.

OpenAI’s Charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work, while leaving the timing uncertain. That is a useful definition of capability. It says little about the system around the capability.

A model can become dramatically better without gaining permission to move money, access a medical record, approve a loan or change production infrastructure. Those permissions come from somewhere else.

Capability is the first handoff

AI commentary often compresses four separate events into one.

A model acquires a capability. A product makes that capability available. An institution permits the product to act. Someone captures the resulting value.

Each handoff can fail.

The measurement problem appears at the first step. Apple’s 2025 research on reasoning models found sharp performance collapse beyond certain complexity levels in controlled puzzles. The result does not prove that AGI is impossible. It shows that fluent intermediate reasoning and success on easier tasks can hide a brittle edge.

METR’s time-horizon research points in the other direction. Models have improved quickly on longer software, machine-learning and cybersecurity tasks. METR also warns that its results do not imply the automation of whole jobs and that its longest estimates carry substantial uncertainty.

Both findings can be true. Models can improve rapidly while remaining uneven. A benchmark can reveal a capability without showing that an organisation can rely on it.

Deployment adds data quality, legacy systems, institutional approval, responsibility and the cost of being wrong. Authority adds identity, permissions, policy and law. Value adds bargaining power.

AGI, if it arrives, changes the first variable most directly. The other three remain institutional questions.

The infrastructure already has owners

The infrastructure around advanced models is forming before any agreed AGI threshold.

Competition authorities in the United States, United Kingdom and European Union warned in a 2024 joint statement that specialised chips, compute, data and expertise could place bottlenecks under the control of a small number of companies. They did not claim that one permanent cartel had formed. They identified the structure that makes concentration possible.

The UK Competition and Markets Authority reached a similarly concrete finding in cloud services. Its 2025 market investigation identified Amazon Web Services and Microsoft as the two largest UK providers and recommended considering further strategic-market-status investigations.

Cloud and model development are connected. The US Federal Trade Commission’s study of partnerships documented cloud-spending commitments, consultation or exclusivity rights, access to sensitive information and technical switching costs across several large provider-lab relationships.

This is the present contest. Compute funds the model. Cloud agreements shape where it runs. Distribution determines who can reach users. Switching costs decide how credible the threat of leaving really is.

A smarter model can lower some costs. It can also make the infrastructure serving it more valuable.

Concentration is not fate. Open models, specialised hardware, public compute and interoperability rules can change the balance. But that change requires architecture and policy. Capability alone does not produce it.

Intelligence needs an operating system

A model generates an answer. An agent acts through tools, credentials and services.

That gap has produced a new layer of infrastructure. The Model Context Protocol gives applications a common way to connect models to tools and context. The Agent2Agent protocol addresses discovery and communication between agents. NIST launched an AI Agent Standards Initiative in 2026 to work on interoperability, security, authentication and authorization.

These efforts are early. None is a universal standard, and a shared protocol does not guarantee shared trust.

An organisation still needs to know which agent is acting, whom it represents, what it may do, and how long that authority lasts. It needs hard rules for actions that should never depend on a model’s judgment. It needs an evidence trail that separates a recommendation from permission, execution and final record state.

I work on this governance layer, so I have a stake in the argument. The argument does not depend on my work. Identity, delegated authority and runtime evidence exist as requirements whenever a probabilistic system can cause a real-world change.

The distinction becomes clearest at the moment of action. A model may propose a payment. A policy service decides whether the amount, purpose and recipient fall inside delegated authority. A payment network attempts the transfer. A bank’s ledger determines whether it settled. Four systems can report four different truths unless the architecture keeps them separate.

That is why runtime is the moment of consequence. Governance documents can describe an organisation’s intentions. Operational controls decide what the system is actually allowed to do.

Infrastructure loses the attention contest

Capability produces a clean demonstration. A model writes a program, solves a proof or controls a computer. The result fits in a video.

Infrastructure appears as a collection of less dramatic decisions. Which identity standard is accepted? Who issues the credential? Which log becomes evidence? Can a customer move its operating history to another provider? What happens after a timeout when an external action may already have succeeded?

Those questions are distributed across standards bodies, regulators, cloud providers, security teams and commercial-selection functions. No single announcement captures the change.

AGI also offers a simpler story. A threshold divides before from after. Infrastructure develops through partial adoption, competing standards and legacy systems that refuse to disappear. It is harder to describe because it looks like ordinary institutional work.

The ordinary work accumulates. Once identity, data and operating history settle around a provider, changing the model endpoint may do little to change the underlying dependency.

The strongest countercase

The infrastructure thesis can be overstated.

A large enough capability jump could lower the cost of writing software, designing chips and operating complex systems. Open models could diffuse useful capability faster than incumbent firms can contain it. Governments could fund shared compute. New protocols could make switching easier. Today’s bottlenecks may weaken.

Those possibilities deserve more than a disclaimer. They are the mechanisms by which intelligence abundance could become broadly distributed power.

They also need evidence. An open model running on scarce accelerators still depends on the compute layer. An interoperable agent still needs an accepted identity and a party willing to honour its authority. A public compute programme still relies on energy, hardware and operational competence.

The thesis should therefore be tested layer by layer. Which bottleneck became cheaper? Which switching cost fell? Which authority moved from a provider to the user? Which part of the stack became meaningfully contestable?

Fast capability progress raises the stakes of those questions. It does not retire them.

What the threshold leaves behind

AGI may already be here, may arrive soon or may never resolve into one clean public moment. The infrastructure decisions are happening under every version of that future.

The important divide is already visible. Intelligence can become cheaper while compute remains concentrated. Models can become more capable while permission stays bounded. Agents can become more autonomous while identity and accountability remain unsettled.

The AGI debate asks when machines cross a line. The harder question is who owns the ground on both sides of it.