A small nation is usually told its AI future is a choice of allegiance — the American stack or the Chinese stack. Read instead as a risk decision, three options appear, and they fail differently. This deck prices them: aligning with either superpower buys capability at the cost of a dependency you cannot reverse on your own timetable; federating accepts a capacity ceiling and coordination overhead, but carries no irreversible tail. A companion paper, “The Map Has No Node for Legitimacy”, makes the underlying case — that sovereignty runs together capacity and rightful authority. This deck takes that as given and asks the practical question: if a small nation will never win the capacity race, what is the lowest-risk way to hold the authority it can?

Strip away the allegiance framing and three options remain, each with a different risk profile. The point is not that one is virtuous and the others are not — it is that they fail differently, and one of them fails in a way you can walk back. The companion paper “The Map Has No Node for Legitimacy” distinguishes sovereignty-as-capacity (the de facto ability to build and run AI) from sovereignty-as-authority (the de jure standing to decide how a system behaves on the people it acts on). This argument assumes that and asks the operational question it leaves open.

You get capability now, and you accept a set of standing exposures: your public data flows through infrastructure governed by another country’s law, including its extraterritorial reach; your dependency deepens with every integration, so the cost of leaving rises over time; and your continuity is hostage to commercial decisions and policy weather you do not control — a model deprecated, a price changed, an export rule rewritten. None of these is catastrophic on any given day. The risk is that they are not reversible on your timetable.

The People’s Republic of China’s 2026 Implementation Guidelines for Intelligent Agents describe a coherent and, on its own terms, impressive architecture: a national intelligent internet with a central agent-registration platform that gives every agent a queryable identity; categorised and tiered governance, in which the state and its industry authorities decide which classes of application get direct scrutiny; and rule-embedding with behavioural fencing. It is sovereignty as capacity, executed well. But it is built around a single substrate that authority flows down through, and it has no exit. For a community that is not the state operating that fabric, the exposure is total: the registry that makes your agents legible is the registry that makes them governable by someone else.

Hold your own data and your own models, and reach other communities through bilateral, consent-bound, instantly-revocable channels. You accept two real costs: a capacity ceiling — your models are smaller, your substrate is borrowed — and coordination overhead, because federation is many agreements rather than one platform. What you do not accept is an irreversible tail. There is no central point whose compromise reaches you, no dependency you cannot withdraw from, no authority above you that cannot be refused. The safe option is not the powerful one. It is the only one of the three whose worst case you can walk back.

Two installations exchange information only through a signed envelope: consent-bound, scope-limited, addressed to a named recipient, and carrying the cryptographic provenance of every record inside it. There is no shared platform that a third party pivots through, and no central authority that can revoke a participant’s standing. A national agent platform is efficient precisely because it is one thing — which is also why its compromise, capture, or policy reversal reaches everyone attached to it at once. A federation has no such centre to seize. The cost of that resilience is coordination overhead; the benefit is that there is no single key whose loss is everyone’s loss.

The move that matters for AI is that the same envelope carries inference. A community’s situated language model — trained on that community’s own content, under its own authority — can reach a peer community’s corpus through the envelope, scoped to a fixed query shape, and answer against it without that corpus ever leaving the peer’s control or its provenance being stripped. This is federated inference: capability assembled by composition across consenting holders, rather than concentrated in one model everyone must query and therefore trust. Authority over how a model behaves on your data is not a permission the platform grants and could revoke; it is a property of the envelope and the signed record, enforced by refusal. When a community withdraws its steering, the model falls back to silence on that domain, not to a vendor default.

It is worth being exact about what federation does and does not buy, because the honesty is the credibility. It does not close the substrate gap: the model weights, accelerators, and compute sit inside exactly the foreign-controlled levers the capacity model maps. What federation buys is the separability of governance sovereignty from substrate sovereignty. The first — authority over data, provenance, steering, and the behaviour of a model acting in your name — can be held today, in full, by actors who will never hold the second. Capability at this layer is gained by composition: a federation of modest models, each sovereign over its own domain, can do at national scale what no one of them could do alone — without any surrendering authority to a centre.

There is an older word for declining to resolve a bipolar contest by joining one side of it. In the twentieth century, nations that refused to be sorted into the American or Soviet bloc called themselves non-aligned — not neutral, not passive, but organised around the refusal itself, and stronger together for it. A non-aligned layer for AI is the same posture rendered in architecture: small nations and indigenous polities that decline the US-or-China sorting, gain capability by federating with each other rather than aligning above themselves, and hold authority at the layer where authority can actually be held. The point of the name is to mark that this is a recognised category of strategic behaviour with a track record, not a novelty — and that it is available now.

Make it concrete in one small country. Aotearoa New Zealand has roughly five million people and no hyperscaler; on the capacity axis it does not register and never will. On the authority axis it has something most large states lack: a constitutional settlement that already assumes plural, co-equal centres of authority. Picture the mesh rather than the platform — an iwi Village, a hauora service, a kura or wānanga, a research group, te-reo infrastructure like Papa Reo, projects like Kete AI, Māori Lab, kahu.code — each holding its own records and a model trained to its kaupapa, each federating with the others through revocable envelopes. The aggregate is national-scale capability with no national registry to capture and no foreign substrate in the production path. The governance form is polycentric, in Ostrom’s sense — and the trust is carried in the record itself: kaitiakitanga as a cryptographically signed field that survives federation and cannot be quietly altered.

Federation does not win the capacity race; it refuses to enter it. It leaves substrate dependence exactly where it is and does not pretend otherwise. It is more work than buying a platform, and it asks communities to hold authority rather than delegate it. What it offers in return is the one property the two alignment doors cannot: no catastrophic, unrecoverable bet. A federated community can lose a partner, a model, even a substrate provider, and recover — because it never handed anyone the single key. The capacity contest was never one most of the world could win. The authority layer was always open. The federated mesh is where a great deal of the world will end up living — and it is already running.

The Village platform and the Tractatus framework are an attempt to make AI sovereignty achievable for actors who will never win the capacity race, by relocating sovereignty to the layer where rightful authority can actually be held — and by letting those actors federate that authority rather than surrender it. Companion paper: “The Map Has No Node for Legitimacy.” This work is licensed under CC BY 4.0 — share and adapt with attribution.