\cleardoublepage \part{Infrastructure} \setbookpart{Infrastructure}

Part III: Infrastructure

Part II described what runs on the platform; Part III decides where it runs. For a European organisation that question is never purely technical, because the most capable models and largest pools of accelerators sit largely outside European jurisdiction, data protection law constrains where personal data may be processed, and the cost of accelerators makes every placement a budget decision.

Chapter 13 sets out the options from hyperscale cloud to on-premises clusters and the sovereignty arguments on each. Chapter 14 shows that consuming AI as a cloud service and running models in-house differ in the kind of problems they create: one shifts risk into contracts and supplier assurance, the other into capacity planning, patching and engineering competence. Chapter 15 closes with the hybrid pattern most large organisations will adopt: provider-neutral agentic work tools whose sensitive traffic a gateway, not the user, routes to self-hosted models.

Model origin and supplier geopolitics are treated in AI Governance (Chapters 6 and 7); this part concentrates on the engineering.

What you will learn in this part:

• How to place AI workloads, balancing capability against European digital sovereignty
• How cloud AI services and self-hosted models differ in security, operations and contracts
• How to record the responsibility split with a cloud provider as code
• How hybrid routing behind one gateway keeps sensitive information on self-hosted models