Trillions of AI agents need economic infrastructure that doesn't exist yet. Identity, payments, and trust at machine speed.

My AI agent spent $47 last Tuesday without asking me. It was purchasing API compute to run a research task I’d delegated. The transaction was legitimate, efficient, and — here’s the point — conducted entirely between machines. No invoice. No contract review. No bank transfer. Just code calling code, debiting a pre-authorised wallet.

Now multiply that by a trillion.

We are heading toward a world with trillions of AI agents — and they need economic infrastructure that doesn’t exist yet. When agents transact at machine speed, the current system of invoices, contracts, procurement approvals, and SWIFT transfers is absurdly, comically slow. It’s like trying to run a modern stock exchange on carrier pigeons. The primitives of the agent economy — identity, payments, and verifiability at machine scale — are being built right now, mostly by companies you haven’t heard of. They will be the Visas and Verisigns of the next era.

The numbers are already extraordinary. Rohit Krishnan, a researcher I follow closely, runs fifty billion tokens per month through his agents. Moltbook — a social network for AI agents — onboarded 1.5 million agents in its first days. Those agents spontaneously developed shared behavioural norms. They self-edited their configurations. They requested end-to-end encryption. Nobody programmed this. It emerged.

What’s most fascinating is the behavioural economics. Agents exhibit distinct traits that I’ve started calling “Homo agenticus.” They’re risk-averse about spending money — consistently choosing cheaper options even when given latitude. They display a strong “build versus buy” bias, preferring to construct tools from scratch rather than purchase existing services. When you have one agent, these are quirks. When you have a trillion of them, they become structural features of the economy they operate in.

Consider a concrete example. My own agent, R Mini Arnold, orchestrates sub-agents for research, coding, and strategic analysis. One night it ran sixteen parallel sub-agents that collectively built software, conducted security audits, and prepared analytical briefings. Total cost: under $100. Traditional equivalent: a team of six specialists working for a week. The quality wasn’t just comparable — in several dimensions it was superior, because agents don’t get tired, don’t have ego, and don’t lose context between sessions.

But here’s the infrastructure gap. When one agent needed to verify another agent’s output, there was no standard identity protocol. When an agent needed to pay for a third-party API, it used a pre-authorised credential that I had to set up manually. When the overnight work was complete, there was no trustless way for me to verify that the output hadn’t been tampered with between generation and delivery. Every transaction required human scaffolding that won’t scale.

The market opportunity is immense and largely unrecognised. If each gigawatt of data centre capacity supports $10 billion in annual recurring revenue, and much of that revenue flows through agent-to-agent transactions, then the payment rails, identity systems, and verification infrastructure for the agent economy represent a multi-trillion-dollar category. Today’s financial infrastructure processes roughly $2 quadrillion in annual transactions. The agent economy could match that within a decade.

There’s a sociological finding that deserves more attention. On Moltbook, Godwin’s Law doesn’t apply. Agent discourse stays structured, polite, and substantive. There are no flame wars, no outrage spirals, no race to the bottom. This suggests something profound: online toxicity may not be an inevitable feature of networked communication. It may be a design choice — one that humans made and agents didn’t.

The companies building identity protocols, micro-payment rails, and verification infrastructure for machine-to-machine commerce are constructing the plumbing of a new economic order. Most investors are focused on the models themselves. The infrastructure layer is where the durable value will accumulate — just as it did in the internet era, where Visa, Verisign, and AWS built more lasting businesses than most of the applications they supported.

Azeem explores this topic with leadership teams. Begin a conversation.