Europe Cannot Govern AI by Cloud Region Alone: A Technical Architecture for AI Sovereignty | Zenodo

Skip to main

You are using an outdated browser. Please upgrade your browser to improve your experience.

Published August 16, 2026

| Version v1

Dataset

Open

Europe Cannot Govern AI by Cloud Region Alone: A Technical Architecture for AI Sovereignty

Authors/Creators

Das, Sangam

Description

Problem Space

European AI sovereignty is often approached through where infrastructure is located : where a model is hosted, where data is stored, which cloud region is selected, and which territorial rules apply to the service provider. These controls are important, but they do not fully answer a more difficult execution-time question:

Can a specific AI computation be prevented from starting under unauthorized algorithmic logic, and can a specific AI-generated consequence be prevented from becoming effective in an unauthorized jurisdiction?

Cloud-region selection is principally a deployment decision. An AI workload may execute in an approved European region while its outputs are subsequently transmitted through external APIs, agentic tools, federated services, telecom networks, satellite infrastructure, downstream platforms, or recipient environments outside the intended jurisdictional boundary. IP-based geolocation provides only indirect evidence because routing can be proxied, tunneled, masked, abstracted, or separated from the actual environment in which an AI-generated consequence becomes effective. The underlying disclosure therefore identifies a gap between infrastructure placement and per-invocation jurisdiction enforcement .

The problem becomes more significant as AI systems evolve from passive information generators into systems capable of invoking tools, communicating with external services, controlling workflows, generating legally or commercially consequential outputs, and operating across multiple infrastructure and jurisdictional domains. In such systems, knowing that an AI service was nominally deployed in Europe does not establish that every computation used authorized algorithmic logic or that every resulting output became effective only within an approved jurisdictional environment.

The technical problem is therefore not merely where AI is hosted , but where and under what authority computation and external effect are permitted to occur .

The proposed architecture addresses this problem through two coordinated execution-time boundaries.

At the first boundary, immediately before an external compute plane is irreversibly authorized to begin inference, a protected domain derives or obtains an Algorithmic Logic Fingerprint (ALF) for the proposed model, checkpoint, orchestration configuration, or execution logic and compares it against an approved or pre-bound ALF condition. A computation-start capability is issued only when the required conditions are satisfied; otherwise authorization is withheld in a fail-closed manner.

At the second boundary, immediately before a candidate output becomes externally effective, the architecture evaluates the runtime governance state together with the intended jurisdictional context. Rather than relying solely on IP geolocation, it can use machine-verifiable infrastructure evidence including certificates, remote attestation, sovereign-cloud identifiers, telecom-domain identifiers, satellite-domain credentials, recipient public-key infrastructure, and other infrastructure-linked attributes.

This creates a different model of European AI sovereignty:

sovereignty is not merely declared by infrastructure location; it is technically enforced at the boundaries where computation begins and AI-generated consequences become effective.

Abstract

Europe's pursuit of AI sovereignty raises a technical question that cloud-region selection, data-residency policies, and IP-based geolocation cannot completely answer: how can jurisdictional authority be enforced for a particular AI computation and a particular AI-generated consequence at the moment they become operationally effective?

This paper presents an execution-time, dual-boundary architecture for AI sovereignty . The architecture separates infrastructure placement from machine-enforceable authority and introduces two coordinated control points: a pre-computation authorization boundary immediately before an external compute plane is irreversibly authorized to begin inference, and an output finality boundary immediately before a candidate AI output becomes externally effective.

At the pre-computation boundary, a protected authorized domain, such as a Trusted Execution Environment (TEE), Hardware Security Module (HSM), or functionally equivalent protected environment, derives or verifies an Algorithmic Logic Fingerprint (ALF) representing the proposed model, checkpoint, orchestration configuration, or execution logic. Computation-start authority is released only when the candidate ALF satisfies the required...