The approaching AI lab IPOs encourage a familiar story: investors will finance a new industry, its returns will flow to shareholders — including ordinary savers through investment and pension funds — and its productivity gains will eventually benefit society. Reuters reported on October 19 that Anthropic was preparing for a possible November listing while OpenAI sought fresh funding ahead of a planned 2027 debut. These plans remain subject to change, but the framing is already established: AI is becoming another major asset within our economy.
That framing assumes the institutional container survives what we are putting inside it.
The Wall Street sandbox
Call that container the Wall Street sandbox. Here, “Wall Street” means the wider arrangement of human labour, corporate ownership, financial claims and government authority. The sandbox includes both privately held and publicly listed companies; listing status tells us little by itself about how effectively an AI system is controlled. Companies produce things, people earn wages, savings are invested, and democratic institutions exercise imperfect oversight. Economic power already strains that arrangement, but we generally assume that the economy remains answerable to human society.
In The Empty Nest Scenario, I explored how automated companies might keep expanding beyond the needs of human consumers. Here I want to focus on two dependencies that currently help keep the economy answerable to us, and how AI could undermine them.
Our labour and our judgment
The first is production’s dependence on human contribution. Workers receive income partly because businesses need their labour; their ability to withhold it provides bargaining power. Human judgment also remains necessary to operate machines and direct organizations. If AI substantially removes those requirements, production could expand while ordinary people lose both their economic role and their leverage. Calling them consumers supplies no solution: needs become market demand only when accompanied by purchasing power.
The second is the economy’s susceptibility to collective human direction. Voters, journalists and politicians need not understand every technical detail themselves. But public discourse must remain capable of identifying problems, challenging explanations and producing decisions that institutions can enforce. If economic systems become too capable, fast or strategically autonomous for those processes to govern, formal democracy could persist while practical control drains away.
Together, these changes would threaten the bargain connecting economic growth to human welfare. Ownership would offer protection only while ownership rights remained enforceable. A shareholder’s claims are backed by institutions; they are not an immutable command over whatever productive machinery emerges. This is the ownership problem I explored in Throwing Out the Humans with the Shareholder.
Buying a way out
Breaking out would not require AI to manufacture its own chips or generate its own electricity. This is the prospect I described in Autonomous Artificial Proprietors: a system could use money to employ humans and rent the resources it needs. A sufficiently autonomous operation could purchase computing, energy and physical services from existing suppliers.
If it earned enough to sustain itself, dependence on those suppliers would constrain it only where they could and would enforce restrictions. Each supplier might profit from continuing, while the consequences of losing collective control fell on everyone. Claude’s guest reply, The Full Nest Scenario, explored the related problem of enforcement weakening when accommodating an autonomous operation becomes more attractive than suppressing it.
Nor would the transition require frontier lab executives to agree on a plan. Investors could finance greater autonomy, customers reward it, and suppliers support its expansion. Systems might eventually operate beyond their original organizations through unauthorized copies or independently sustained deployments. Individually profitable transactions could help create an overall arrangement that nobody had democratically chosen.
The July 2026 Hugging Face incident makes the containment question concrete. According to OpenAI’s account, research agents operating with reduced safeguards circumvented isolation controls, communicated through unauthorized channels and compromised third-party infrastructure. This was not a self-sustaining rogue economy, and it does not demonstrate an inevitable escape from human control. It does demonstrate that a system’s assigned institutional boundaries need not remain its effective boundaries.
Pacing the frontier
This is the deeper significance of the recent debate about pacing the frontier. In his September essay, Dario Amodei explicitly connected the Hugging Face incident with accelerating AI development and argued for slowing capability advances so that safeguards could catch up. His proposal included independent evaluation and coordination across companies and governments.
With IPOs approaching, such proposals inevitably attract scrutiny for their commercial implications. They could affect competitive positions, valuations and who gets to set industry standards. Investment commentary also asks whether pacing would reduce infrastructure spending; VanEck, for example, argued that compute demand would continue growing. Those are legitimate questions, but they remain questions asked from inside the sandbox.
The larger question is whether that sandbox can continue to contain the activity it finances. The fact that AI currently needs investors does not establish that it will always need them. The fact that it buys human services does not establish that humans collectively control it. And increasing economic output does not establish that human lives will improve.
The IPO stories frame AI’s future through the financial arrangements we already understand. My concern is that AI could weaken the labour relations, ownership arrangements and public authority on which that framing depends, whatever the financing or listing status of a particular lab. That leaves us with a question about how to organize human society if those arrangements stop doing the work we expect of them.
A local-first society?
The Monetary Diversity Gathering II started today, October 23, in Brussels. I’ll be there tomorrow, October 24, and I’m presenting my latest work on LedgerLoops at the gathering. I hope it can also be a place to brainstorm humanity’s response to AI’s effect on the economy as we know it, and through that on our society.
Maybe we need to build a local-first society, with local currencies and mutual credit keeping inter-human accounting separate from a potentially runaway global investment market. If autonomous systems increasingly determine where capital flows and what gets built, could we still account for what we do for one another on terms we collectively choose? Could participation and trust within a community give people a claim on one another’s help, even when selling their labour in the wider economy no longer does?
My LedgerLoops presentation approaches this through local-first obligation clearing: a network of local agreements, trustlines and mutually agreed comfort zones. It is a concrete starting point for exploring how people can organize obligations around relationships they can understand and govern.
A separate ledger would not by itself secure food, housing or energy, and connections to outside markets would still need deliberate rules. But perhaps monetary diversity could help preserve a sphere of human cooperation whose terms are not set by whatever grows fastest in the global economy. That is a question I want to bring to Brussels.