AI sovereignty: own the intelligence, don't rent it

Data, infrastructure, and operational sovereignty are well understood and increasingly available off the shelf. The fourth is the one that decides the others.

Data sovereigntyWhere does the data live, and whose law governs it?

The narrowest and best-understood dimension. Protected health information sits under a jurisdiction, a retention policy, and a set of parties who can lawfully compel access to it. Most enterprise vendors can now answer this question, and answering it has become table stakes rather than a differentiator.

Infrastructure sovereigntyWhere does the compute run, and who controls it?

Which region, which cloud account, which network boundary, whose hardware. Dedicated tenancy and in-region inference resolve most of it. Note that this is a question about location and tenancy — it says nothing about who owns what runs there.

Operational sovereigntyWho holds authority over the lifecycle?

Who may change how a system behaves, who approves that change, who can roll it back, and who can prove after the fact what it did. In healthcare this is where governance stops being a policy document and becomes a runtime: role-based permissions, human approval, audit trails, and pre-deployment evaluation.

Model sovereigntyWho owns the intelligence itself?

The dimension the market has not named, and the one that decides the other three. You can host a rented model in your own datacenter, in your own region, under your own change control, and still not own the intelligence. The workflow knowledge your experts spent years accumulating ends up encoded in weights that belong to someone else — subject to their roadmap, their pricing, their deprecation schedule, and their terms.

In most industries sovereignty is a procurement preference. In healthcare it is a regulatory position. Protected health information is governed by HIPAA, agents that touch care decisions answer to regulators across all 50 states, and organizations operating in Europe face the EU AI Act. Every dimension has to be demonstrable rather than asserted: which data an agent saw, which policy it applied, who approved its behavior, and who owns the model that produced the decision.

There is also a capability argument, and it is the one most vendors skip. Our own χ-BENCH testing put frontier models through long-horizon, policy-rich healthcare workflows drawn from 1,279 handbook policy documents. The best agent completed 28% of them, and none completed prior authorization end to end. A rented frontier model is not a healthcare strategy — the workflow knowledge, the governed harness, and the specialization are where the remaining 72% lives.

That is why sovereignty and capability are the same problem here. The organizations that own their workflow intelligence are the ones that can close the gap at all.

Every time an expert corrects an agent, resolves an edge case, or applies a policy that exists nowhere in writing, something valuable is produced. The question is where it accumulates.

Rent the model and it accumulates upstream: your organization supplies the corrections, and the capability they create belongs to a vendor who licenses it back to you and to everyone else. Own the model and the same evidence compounds inside your boundary, into an asset that gets more specific to your operation every quarter.

ACTAVA's learning loop is built for the second shape. Cura 1T, our one-trillion-parameter healthcare model, acts as the teacher. Your production evidence — agent trajectories, expert corrections, outcomes — trains your specialized models, inside your compliance boundary. Cura 1T is never trained on customer data; it was post-trained through recursive self-improvement, and its role in your deployment is to teach, not to learn from you.

That distinction is the whole of model sovereignty. Read the mechanics on why KORA and the transformation blueprint.

Four products, one loop, each answering a different dimension.

KORAOperational sovereignty

Build, evaluate, deploy, and improve multi-modal agents inside one governed platform, with role-based permissions, human approval workflows, and audit trails on every run.

Χ-BENCHEvidence

A public benchmark of long-horizon healthcare workflows. Our own testing found the best frontier agent completes 28% of them, and none completed prior authorization end to end — which is why a capable rented model is not a strategy on its own.

CHRYSOGovernance you can prove

110+ controls mapped across four framework families, continuous monitoring, and audit-ready evidence — so operational sovereignty is demonstrable to a regulator rather than asserted in a policy document.

CURA 1TModel sovereignty

A one-trillion-parameter healthcare model that acts as the teacher for models the customer owns. Your workflows, policies, and expert decisions become your specialized model, inside your compliance boundary.

Together they close the loop: execute the workflow, prove the behavior, govern the evidence, and train the model you keep.