Sovereign European AI infrastructure.

Sovereign AI is control of the stack

Where a request runs matters. So do the people operating the machines, the rules governing the data, and the ability to move the workload.

Start with what Eldrim does today

Eldrim's current product is an early-access API and commercial layer for open AI models on European infrastructure. It uses an OpenAI-compatible interface so customers can connect with familiar tools.

The gateway supports account-level residency policies, including Sweden-only and EU execution. It filters provider routes by that policy before execution and, until an answer has started, can select another eligible provider when one fails. The configured provider and residency are recorded with request usage and exposed in response headers.

Those records depend on verified provider information. A routing label cannot establish a legal guarantee by itself: execution location, operator access, logging and contractual terms must be checked for each service.

Control has several dimensions

A server address in Europe answers one question. Customers also need to understand who can operate the service, which dependencies remain, and what happens if they need to leave.

Physical execution location: Where the computation happens, including fallback routes. A residency policy needs to apply when a provider fails as well as during normal operation.
Legal jurisdiction: Which entities and jurisdictions govern the service, its contracts and data handling. Physical location alone does not settle all of these questions.
Infrastructure operation: Who runs the machines, holds administrative access, manages incidents and decides when systems change.
Customer data and logs: Who can access prompts, outputs and operational records; where they are retained; and how they are deleted. These controls must include upstream providers.
Hyperscaler dependency: Whether a workload depends on one cloud's identity, storage, networking or proprietary services, and what replacing those dependencies would require.
Hardware-supplier dependency: Which accelerators, interconnects, firmware and software stacks are required. Owning a server does not remove its supply-chain dependencies.
Software and orchestration: Who controls model versions, deployment, scheduling and the policy that chooses an execution provider.
Workload portability: Whether the model and application can move between suitable providers, with their licences, quality requirements and service levels intact.
Ownership and system design: Eventually, who owns the physical compute and can choose how its memory, networking, power and cooling fit together.

These dimensions should be assessed separately. Control of one layer is useful even while other layers still depend on external suppliers.

Increase sovereignty layer by layer

Sovereignty is something we can increase layer by layer.

European customer relationship and policyToday's focus: the customer interface, commercial relationship and rules for routing requests.
European compute providersThe initial supply model: establish approved providers with explicit execution and data-handling terms.
Eldrim-operated European infrastructureA future step toward control of site operations, provisioning and the physical environment.
Eldrim-owned computeAcquire capacity when measured demand, customer commitments and economics justify it.
Eldrim-designed systemsDesign around the workloads we understand, including memory bandwidth, networking, energy and useful heat.
Potential future custom siliconAn option only if workload and system measurements justify the cost and complexity. No silicon programme is committed.

Each step should fund, validate and reduce the risk of the next. This is a direction with decision points, not a promise that every layer will be built.

Using NVIDIA or AMD hardware leaves dependencies on suppliers, their software ecosystems and a global semiconductor supply chain. European operation or ownership does not make those dependencies disappear. We intend to increase control where it is economically and technically useful, while stating what remains outside our control.

Open models matter

Open-weight models make the model's parameters available for use under its licence. Where the licence and technical requirements allow it, the same model can run on infrastructure chosen by the customer.

Customers can reduce dependence on a single vendor's hosted model service.
Model choice can be separated from the choice of execution provider.
Workloads can move as economics, regulation and hardware change.
Infrastructure policy can constrain where a model runs without requiring a different application interface.

Portability still requires work. Licences differ, models have different memory needs, and behaviour must be evaluated on the customer's task. Open weights do not imply unrestricted use or equivalent quality to any particular proprietary model.

One API, increasing control underneath

The customer-facing contract should remain stable while the infrastructure behind it evolves. Our intention is that moving from purchased compute to Eldrim capacity will not require customers to rewrite their applications.

Customer application

api.eldrim.net

Policy, residency and routing

Today's supply model: approved European providersEarly access; provider onboarding in progress

Later: Eldrim capacity and partner capacitySubject to demand, site feasibility and investment

The final step shows how the supply pool could evolve over time. A request goes to an eligible execution provider; it does not pass through both infrastructure generations.

Provider changes must still satisfy the customer's residency policy, model requirements and service level. A common API reduces integration changes; it does not make different model behaviour, capacity or latency interchangeable. Those differences need testing and explicit terms.

Why Sweden and northern Europe

We are investigating locations where legal, energy and physical conditions can support this design together. Each characteristic needs evidence at the actual site.

European jurisdiction: a basis for the customer relationship and operation, with the full provider and data-handling chain still to be assessed.
Access to low-carbon electricity: a characteristic to investigate alongside grid capacity, supply terms and hourly availability. We have not established an Eldrim site electricity mix or emissions result.
A cold climate: potentially useful for dry cooling, subject to equipment limits and conditions during warmer periods.
District heating: a possible customer for recovered heat when location, temperatures, demand and commercial terms fit.
Geological conditions: some sites may support seasonal thermal storage; boreholes, groundwater and heat losses require site-specific study.
European connectivity: network access must be evaluated for route diversity, capacity and the latency customers require.

Haparanda is the first feasibility case. Available land, grid connection, network paths and a workable heat relationship remain questions to resolve. It is not an announced or committed compute facility.

Network concept showing separate coloured connection routes approaching a modular compute site — Haparanda · network concept **Control includes the connections.** The sketch explores separate network entry routes. Physical route diversity, operators and upstream dependencies still need verification; lines on a plan do not establish independent connections or resilience. View sketch fitted to your screen.

Sovereign and sustainable reinforce each other

Operating and owning more of the infrastructure would let Eldrim control thermal design, water use, heat recovery, energy scheduling and workload placement. Today, many of those physical choices belong to the provider from which capacity is purchased.

That control creates an opportunity, not an environmental result. We would still need to measure useful compute, water consumed and heat delivered, and compare them within a stated system boundary. Our sustainable AI direction explains the designs and measurements we want to pursue.

Sovereign AI means choosing where and how computation happens. Sustainable AI means designing that infrastructure so the energy has more than one useful life. Eldrim is working toward both.