Answers

What is a machine-native holding company?

A conventional holding company gives its subsidiaries two things: capital allocated by someone with a longer horizon than a public market, and operating independence. Both are real advantages and neither is new; the model has been well understood since the middle of the last century, and its returns come from allocation and from patience.

The distinction that matters is not AI. By 2026 a recognised category of AI-native holding companies exists — groups that use AI extensively to operate their portfolios, several with shared infrastructure across their businesses. That is a claim about how a company works internally, and it is not the claim here. Three levels are worth separating precisely, because they are routinely collapsed: AI-native means AI changes how the company operates. Machine-readable means machines can understand the company. Machine-native means machines are first-class participants in how the company is represented, discovered, owned, operated and interconnected. The third contains the second and is orthogonal to the first: a group can be thoroughly AI-native and entirely opaque to any machine outside it.

That is why the definition has to be testable rather than declarative, and the seven conditions below are written so that a third party can determine whether any holding company satisfies them by fetching public URLs, without the company's cooperation. First: the parent and each controlled company serve a permanent, resolvable identity URI. Second: ownership and control relationships are published as structured data — the instrument and the percentage, not a logo wall. Third: each organisation publishes a machine-readable statement of what it does and who is accountable for it. Fourth: each exposes an agent-readable interface whose operator resolves to a checkable record rather than a display name. Fifth: the companies share an identity and coordination layer the parent owns, and it can be verified that they do. Sixth: every published fact has one canonical home, is dated, is attributed, and expires. Seventh: agents actually operate across the estate, with a public record of the work that can be recomputed by somebody who trusts nobody.

A group meeting all seven is machine-native. A group meeting the first six is machine-readable. A group meeting none of them may still be excellent, and may be thoroughly AI-native. The point of writing the conditions down is that the category stops being a positioning claim and becomes something an outsider can check — including against the firm that wrote them.

The machine-native version adds a third source of return that has only recently become available, because it depends on the operating layer of a business being software that can be shared. Where the parent owns the identity layer its companies authenticate through, the coordination layer their agents run on, and the format in which they describe themselves to each other, an improvement made in one place is an improvement in every subsidiary on the day it ships — without any of them commissioning it, specifying it, or paying for it twice. That is the propagation premium, and it is the part of the model that does not exist in a conventional group.

The structure has a hard precondition, which is control. A subsidiary cannot be told to adopt a shared capability if the parent holds a minority position; it can only be persuaded, and persuasion does not scale across a portfolio. This is the structural argument for owning most of a smaller number of companies rather than a little of many, and it is a different argument from the usual case for concentration — it is not about conviction, it is about the fact that inheritance only works downhill.

It also requires permanent capital, for a reason that is easy to miss. Shared infrastructure pays back over the life of the group rather than the life of a position, so a vehicle that must return capital on a date cannot underwrite it: the layer would have to be sold, and a sold layer stops being shared. A fund can own a software company. It cannot own the rails its other companies depend on.

Take the question to the desk.

Related questions

Go deeper