Which Open-Weight Models Are Actually Hosted in Europe, and Where

AI This article was created with the help of AI.

Why 'EU provider' and 'EU-hosted model' are different claims

When evaluating AI infrastructure, enterprise engineering leads and Data Protection Officers frequently conflate the corporate registration of a cloud vendor with the physical execution path of an API payload. A company registered in Berlin, Paris, or Dublin can still route incoming inference requests to GPU clusters in North America. This architecture is common across the industry: the European Union currently hosts only a rough 5% of global AI compute capacity, while the United States operates nearly 75%. Because frontier GPU clusters remain heavily concentrated outside the bloc, many platforms quietly proxy API calls across the Atlantic to avoid local hardware constraints.

For cost-driven enterprise adopters transitioning from proprietary models to open-weight architectures, this geographic disconnect introduces severe regulatory exposure under GDPR Article 44. Passing sensitive corporate data, employee telemetry, or customer records through non-EU infrastructure triggers third-country data transfer requirements, regardless of where the provider sends the monthly invoice. Regulatory scrutiny on generative AI processing has intensified significantly across Europe, evidenced by enforcement actions such as the 15 million euro GDPR fine levied by the Italian Data Protection Authority (Garante) against OpenAI over lawful basis and transparency failures. Achieving true GDPR-compliant LLM inference requires auditing the physical server topology for each specific model string rather than accepting broad platform-level residency promises.

• Corporate jurisdiction does not govern routing: A European entity contracting an API can still terminate inference workloads on overseas GPU instances.
• Hardware concentration drives cross-border hops: With Europe hosting approximately 5% of global AI compute, multi-tenant platforms frequently backfill regional capacity deficits with foreign capacity.
• Compliance attaches to the payload: Data residency obligations apply directly to the physical memory and compute pipeline processing the prompt.

To guarantee full sovereignty and eliminate international data transfer liability under the US CLOUD Act, engineering teams must decouple vendor identity from model-level hosting. Verifying physical execution at the individual model endpoint is the only defensible method to ensure user prompts never leave European jurisdiction.

The model-to-region table

We treat data residency as a per-model architectural fact rather than a blanket marketing claim. Within our serverless inference infrastructure, our supported open-weight models execute entirely within the EU in the eu-north1 region. The underlying compute infrastructure is distributed across sovereign European datacentres located in Spain, France (Paris), and the Nordics, ensuring that network traffic and token generation remain strictly within EU and EEA borders.

Earlier documentation and third-party catalogues contained outdated regional metadata for several leading open weights. We explicitly confirm that high-demand architectures such as Kimi-K2.6 and DeepSeek-V4-Pro run in eu-north1. The table below provides an audited breakdown of model families, target architectures, physical execution regions, and default residency boundaries.

Engineering teams conducting compliance audits can rely on this mapping to configure their API clients. When your workload requires absolute territorial data containment, targeting the eu-north1 model strings guarantees that prompt tokens, KV cache blocks, and generated completions remain on physical European silicon throughout the entire inference lifecycle.

How to verify a provider's claim yourself

Relying on vendor assertions without independent technical verification introduces operational risk into your production stack. Proprietary, closed-source inference engines often wrap orchestration in opaque virtualization layers, obscuring where memory allocations take place and whether prompt data is retained for telemetry or platform fine-tuning. In contrast, verifying inference sovereignty requires transparency at both the software runtime and the memory allocation layers.

Our infrastructure runs on an open inference stack built on vLLM, NVIDIA Dynamo, and TensorRT-LLM. This architecture ensures that prompt processing follows standardized, observable memory management primitives like PagedAttention and automatic prefix caching. Under vLLM automatic prefix caching, key-value (KV) blocks are partitioned, uniquely hashed by their block and prefix tokens, and managed dynamically across non-contiguous physical GPU memory.

• Volatile memory caching: Prompts and intermediate attention states reside strictly within volatile GPU memory (VRAM) during active sessions.
• Short-lived retention: Prefix cache blocks are held in VRAM only to accelerate sequential turns, and cached blocks are evicted under a Least Recently Used (LRU) policy from the head of the free block queue as new allocations are needed.
• Zero disk persistence: Prompts and generated outputs are never written to persistent block storage, relational databases, or cold disk logs.
• Stateless execution: Once the inference request finishes and the session cache expires, all trace data is permanently overwritten in VRAM.

Engineers can audit latency profiles, perform TCP traceroutes against endpoint hostnames, and inspect response headers to confirm that network hops terminate in European transit hubs rather than transatlantic undersea cables. Pairing open-stack runtime mechanics with network-level inspection provides empirical proof of data isolation.

What to put in your record of processing activities

Under GDPR Article 30, enterprise controllers must maintain a detailed Record of Processing Activities (ROPA) documenting every processing operation involving personal data. For AI deployments, Data Protection Officers must demonstrate a lawful basis and rigorous technical and organizational measures (TOMs), aligning with regulatory guidance such as the European Data Protection Board (EDPB) Opinion 28/2024 on personal data processing in AI models.

When logging serverless open-weight inference within your ROPA, your compliance documentation should accurately describe the data flow, processing boundaries, and lack of persistent storage. Below are the specific operational attributes to include in your compliance register:

• Purpose of processing: Real-time inference, text summarization, embedding generation, or automated reasoning on enterprise payloads.
• Categories of data subjects: Corporate end users, authenticated platform clients, or internal developers submitting API requests.
• Technical processing mechanics: Ephemeral processing in volatile GPU VRAM with zero persistent database retention; prompt cache evicted automatically within minutes.
• Hosting territory and transfer status: Physical infrastructure strictly within the European Union (eu-north1 datacentres in Spain, France, Nordics); no third-country transfers under GDPR Chapter V.
• Contractual framework: Data Processing Agreement (DPA) executed between controller and processor, available on request for enterprise onboarding.

Enterprise legal teams must ensure their ROPA explicitly differentiates between stateless inference and stateful object storage. Because prompts are processed ephemerally without secondary training usage, controllers can establish legitimate interest or contractual necessity with significantly lower regulatory friction.

Serverless Inference: Scaling EU-Hosted Open Models

Transitioning to self-hosted or sovereign open models should not require building complex cluster management infrastructure from scratch. Lyceum provides Serverless Inference to give European engineering teams direct, OpenAI-compatible access to our EU-hosted open-weight models running on high-performance GPU clusters across Spain, France, and the Nordics. By pointing your API base URL at the serverless endpoint and specifying an eu-north1 model string, you can deploy state-of-the-art open models without altering your application logic.

Serverless Inference eliminates infrastructure overhead through transparent per-token billing, allowing you to scale from initial prototyping to high-throughput production without upfront hardware reservations or fixed baseline commitments. The service operates as a self-serve, pay-per-token platform and carries no SLA, availability tier, or uptime guarantee. Teams requiring contracted infrastructure, custom SLAs, or dedicated capacity can deploy private endpoints through dedicated GPU offerings, while serverless users can monitor live platform health on the public status page.

• OpenAI SDK compatibility: Drop-in replacement for existing API clients and agent frameworks.
• EU-hosted open models: Direct access to DeepSeek, Gemma, GLM, and Kimi architectures securely hosted in eu-north1.
• Zero data retention: Ephemeral VRAM processing with no persistent disk storage and no model training on customer data.
• Per-token economics: Pay strictly for the input and output tokens consumed, with no idle hardware waste.

Audit your AI infrastructure and verify your target model strings against our residency table to ensure your inference workloads run strictly within sovereign European borders.