Don't train a base model. Own the evidence layer.

Every AI startup right now says it is fine-tuning its own model. For EU-sovereign compliance work, that is the wrong question.

The problem is not model quality. The problem is the artifact-legitimacy gap: technical readiness and legitimacy do not run on the same clock. A frontier model produces an artifact in hours. Binding it to regulatory reality — peer review, independent implementation, regulator approval, per-jurisdiction legal review — takes weeks or months, and it does not get shorter. Owning your own model does not close that gap. It moves it.

And physical AI and compliance evidence are two different optimization problems. A robot's perception→action in real time is a different job from producing a court-defensible chain of reasoning. One model for both is a compromise in both directions.

Do this instead

Adopt a ready sovereign component for the physical layer. NVIDIA Cosmos 3 Edge is an open world model that runs on the device — on-prem, on the box you already have. You don't train it. You deploy it.

Run compliance reasoning on an open LLM through NIM. That is non-sovereign work: benchmarking and drafting, not the core of customer data.

Invest in what you cannot buy: an owned evidence loop. Private evals, an RL environment, adaptation on your own fleet-data, KYA governance, Article 12 evidence. The base model commoditizes every month. The loop does not. It is the one layer a competitor cannot copy, because it is built on top of your operations.

Name the risk too. Don't hide it.

The NVIDIA stack is dominant: silicon, model, and serving from one vendor. That is real concentration. But because the weights are open, portability is real. Keep one open model and one workload runnable on a non-NVIDIA accelerator (AMD MI300, Intel Gaudi) at testbed level. Then the exit cost is auditable, not theoretical.

How this shows up in the product line

One product, three tiers — not four disconnected products.

DWS IQ 7 Aegis is a two-layer product: physical perception→action (an adopted world model of the Cosmos 3 Edge class) and compliance evidence (a separate LLM layer), bound together by the KYA loop. Its tiers scale sovereignty: Sovereign Infra splits the fleet/edge physical layer from cloud compliance, Connect is the global tier, without sovereign constraints. And Firehorse Production Engine runs operator-independent workflows on computer vision where operator perception is replaced — construction forensics, not a new model.

What is still missing

Because the artifact-legitimacy gap applies to us too. The fallback plan for NVIDIA supply risk has to be written into a visible risk register, per product where DGX, Cosmos, or NIM is critical. It is not enough that an exit is possible; it has to be rehearsed. This is the question an investor and an enterprise buyer ask the moment you mention NVIDIA dependency. We say it out loud before they ask.

The principle is universal

The moat is not in the model. It is in how the model's output is bound to regulatory reality. You can buy the components. You cannot buy the evidence layer.

More on architecture and products: DWS IQ 7 Aegis.

Risto Anton Päärni
Founder, CEO · Lifetime Oy · the Agent Outcome Company