Europe – October 27, 2026 -- Self-hosted LLM infrastructure tends to become the cheaper option above roughly 50 million tokens processed per day, while commercial APIs usually stay more economical below 5 million, according to AppsTek's analysis of when enterprises need a sovereign LLM. The figures are illustrative and should be checked against current provider pricing and actual workload patterns.
AppsTek's third piece in its sovereign AI series argues that the model layer is now a governance decision, not only a technology choice, as generative AI moves into regulated, revenue-bearing workflows.
EU AI Act deadlines turn model choice into a compliance question
Prohibited AI practices became enforceable in February 2025. Obligations for general-purpose AI model providers, including transparency and training-data documentation, took effect in August 2025. High-risk system obligations, including conformity assessments, reach full enforcement in August 2026.
An LLM used in employment screening, credit decisions, healthcare triage or critical infrastructure can trigger high-risk classification wherever it runs. Such systems require documentation, audit trails and conformity assessments that many general-purpose API services are not built to expose in full.
CLOUD Act exposure means data location alone does not settle legal risk
The US CLOUD Act can compel a provider to disclose data to a foreign government regardless of where it is stored. The Schrems II ruling found that conflict structurally difficult to reconcile with EU data protection law. An in-region hyperscaler offering may therefore meet territorial requirements while remaining subject to a foreign parent company's legal obligations.
Gartner forecasts sovereign cloud IaaS spending near $80 billion in 2026
Gartner projects the spending at approximately $80 billion in 2026, up 35.6% on 2025. Governments and regulated industries such as energy, utilities and telecommunications are the leading buyers. NVIDIA reports that sovereign AI revenue roughly tripled to more than $30 billion in fiscal 2026. In a McKinsey survey of roughly 300 executives, investors and government officials, 71% called sovereign AI an existential concern or a strategic imperative.
Five criteria place each workload on the deployment spectrum
Options run from a public multi-tenant API to regional APIs, dedicated tenancy and fully self-hosted open-weight models, including air-gapped on-premises setups. AppsTek names five decision factors:
- Data classification: regulated, proprietary or classified data can narrow the choice to a single self-hosted or air-gapped tier.
- Regulatory classification: EU AI Act high-risk status or sector rules in financial services and healthcare set documentation, auditability and human-oversight requirements.
- Token volume: between 5 million and 50 million tokens per day, the outcome depends heavily on sustained utilization.
- Licensing: Meta's Llama 4 community license excludes companies domiciled in the EU from its multimodal capabilities.
- Operational readiness: industry guidance points to budgeting at least half to one full-time engineer in year one, a cost that often exceeds the hardware budget for mid-sized deployments.
GPU memory, not compute, sets the hardware bill for self-hosting
Required VRAM equals model weights plus key-value cache plus operational overhead, with a 20% to 30% safety margin added. A 70-billion-parameter model quantized to INT4 needs approximately 39 GB for weights alone and can fit on a single high-end workstation GPU. The same model at FP16 needs roughly 140 GB, typically spread across multiple data center-class GPUs. Quantization to INT8 or INT4 cuts memory needs substantially, at a measurable but often acceptable cost to output quality.
Production deployments also need an inference serving layer. Common choices are vLLM, SGLang, TensorRT-LLM (tuned for NVIDIA hardware) and Text Generation Inference (tied to the Hugging Face ecosystem). Ollama and llama.cpp are common for proof-of-concept work but are generally not the production standard at enterprise scale.
Self-hosting also removes the content filtering, abuse prevention and safety tooling that commercial API providers typically bundle into their service.