The market for sovereign intelligence.

1Intelligence became infrastructure — and it's borrowed.

A decade ago, "AI" was a feature. Today it is becoming load-bearing: support, underwriting, claims, clinical triage, code, decisions. The capability is extraordinary — and almost none of it is sovereign.

The models economies increasingly run on are produced, hosted and governed by a small number of firms, in a small number of jurisdictions. That looks like a new kind of dependency, with the shape of every dependency before it — energy, payments, cloud: invisible until it isn't, and hard to unwind once load-bearing. The question we keep hearing boards and treasuries start to ask is the same one: what happens to us if the supply changes its mind? We're interested in whether you hear it too.

2One vendor's bad day is now your bad day.

Build your operations on a single AI provider and that vendor's every bad day becomes yours: an outage, a sudden price increase, a policy or model change that breaks a workflow, a quiet shift in where data is processed. Concentration turns ordinary supplier risk into correlated risk — when a whole economy leans on the same foreign vendor, the loss is simultaneous and hard to insure at scale.

And it isn't only vendors that fail. The 2024 Red Sea cable cuts made the lesson physical: routes fail, not just providers. A continuity story worth anything has to survive losing a vendor, a region, or a path — automatically, without the customer wiring the failover themselves.

3Why now: four forces we think are converging.

• Regulation is drawing the line between layers. The rules taking shape already seem to treat the model layer and the application layer as distinct things with distinct duties. That line needs a control to sit on.
• Residency is becoming non-optional. Data-localisation and sovereignty requirements are spreading across regulated industry and government. "Where does the inference happen, and under whose law?" is turning into a procurement question.
• The insurance precedent. Cyber-insurance once paid out on any breach; then it began requiring controls — MFA, EDR, backups — before it would cover you. We suspect AI dependency follows the same path: "what is your AI fallback?" becomes an underwriting question.
• Commoditisation makes vendors swappable. The OpenAI-compatible interface is a de-facto standard. When the interface is shared, the vendor behind it becomes a routing decision — which is the seam a sovereign layer would occupy.

4The shape of the market: an energy market that clears globally.

Intelligence is now produced from energy and silicon, metered by the token, and shipped over fibre. That makes it behave like an energy market in the ways that matter operationally: supply has to meet demand in real time — capacity must be live when the request lands — it splits into firm and floating tiers, it dispatches merit-order (cheapest capacity first), and it prices locationally. If you've run a grid, this is familiar.

But one difference rewrites the geography of the market. Electricity is regional because electrons leak — resistive losses, distance limits and interconnection caps mean you can't economically move power across the world, so power markets clear inside a grid. Intelligence has no such loss. A token is information; it crosses the planet over fibre at near light-speed and arrives identical, its delivery cost a few milliseconds of latency against an inference that already takes hundreds. So the market for intelligence is global by default — supply anywhere can serve demand anywhere.

Which means the only thing that draws a border around intelligence is the law, not physics. Data-residency and sovereignty rules are the artificial grid lines — and they're exactly the lines a routing layer has to know. A market that is global by physics but partitioned by jurisdiction is precisely what sovrgn would operate: take the cheapest lossless supply anywhere the law allows, and pin to in-country silicon when it doesn't. The two-sidedness follows — every country and firm is a buyer of inference (short the cognition it can't make at home) or a seller (long on cheap power and spare silicon) — and because supply is lossless and global, the time-zone hedge is real: your overnight silicon serves their daytime peak, and theirs serves yours.

Two frictions remain, and we'd test how much they bite: latency still rewards nearby supply for real-time work, and — unlike a kilowatt-hour — not every token is fungible; a frontier model's output isn't a weak one's. Capability tiers (smart / fast / sovereign) are how we'd make supply comparable within a tier — whether that's enough to clear a true market is an open question.

5How big — we don't fully know yet.

We won't pretend to a number we haven't earned. The first-order intuition: a nation's demand for inference tracks its knowledge workers × adoption × tokens per worker per day. At deliberately conservative, illustrative rates that already implies very large daily volumes per economy — and that's today's adoption, not the asymptote. Most jurisdictions look structurally short of their own supply. What we can argue is the shape of the opportunity; whether it's venture-scale, and how much margin a routing layer actually captures, is genuinely one of the things we want help pressure-testing.

6Why a neutral routing layer might win — and why it might not.

The position we find most defensible isn't another model or a reseller locked to one stack — it's the neutral layer in between. You ask for a capability — smart, fast, or sovereign — and the layer resolves the right provider under your jurisdiction's rules, with automatic failover, a circuit breaker, and fail-closed residency. Because you code against capabilities, not vendors, the layer makes the vendors swappable — which is what could make the layer, not the vendor,the durable position. The terminal node of every in-jurisdiction chain would be a national sovereign-silicon backstop; AloomU is Australia's. It's also low-friction to try: the protocol speaks plain OpenAI, so adoption is a one-line base_url change — and so is leaving.

The honest counter-argument: the hyperscalers already route across models (Bedrock, Azure) and could fold "sovereign failover" into their own stacks. So the open question isn't whether routing is useful — it plainly is — but whether a neutral layer, owned by no provider and accountable to the jurisdiction, is a business or a feature. We think neutrality is exactly the thing a cloud can't credibly sell. We'd like to be argued with on that.

7What would confirm this — and what would kill it.

So we can be wrong out loud, here's what we'd treat as evidence either way.

Signals it's real

• A buyer chooses a neutral layer over a single cloud specifically for residency or continuity — and pays for it.
• An insurer or regulator names "AI fallback / failover" as an expected control.
• A second sovereign instance (beyond Australia) gets stood up because a government wants the aggregation, not just the tech.

Signals it's a feature, not a market

• Buyers are happy with their cloud's built-in multi-model routing and don't value neutrality.
• Residency turns out to be satisfiable by in-region cloud regions alone, with no demand for a cross-vendor control.
• The routing layer can't hold enough margin to be more than a thin toll.

8Where we're least sure — tell us.

If you read one section, make it this one. These are the questions we'd most like challenged:

• Is sovereignty a buying criterion today, or a 2–3-year one? Who pays for it now, and for which workloads?
• Is the neutral layer defensible, or do the clouds absorb it? What's the moat beyond "we're not a provider"?
• Is the seller side real near-term? Or is supply just the hyperscalers for now, making the "exchange" framing premature?
• Nation-first or business-first? Is the wedge selling aggregation to governments, or continuity to enterprises?
• Does a routing layer capture enough value to be a venture-scale business, or is it infrastructure that wants to be cheap?
• What are we not seeing? The risk, the competitor, or the reason this is harder than it looks.

This is a discussion draft circulated for comment — not an offer, a forecast, or investment advice. Figures are illustrative and labelled as such. We expect to be wrong about parts of this; surfacing which parts is the point of sending it. Energy-market framing references public grid-operator data (AEMO, Elexon, ENTSO-E, US ISOs); the 2024 Red Sea submarine-cable disruption is public record. © sovrgn, 2026.