"Make it work like ChatGPT, but nothing leaves the building."
That is the brief from every ministry, central bank and telecom operator whose regulator has said the data stays in the country. Most AI tutorials assume the internet: public registries, model hubs, SaaS dashboards and someone else's API. None of that works in a data centre with no route out.
Sovereign AI Stacks is the reference design that does. It takes you from an empty rack to a multi-tenant LLM platform that runs on hardware you own, with no connection to any vendor, and gives you the evidence an auditor or procurement panel will ask for.
What you will buildA threat model and the technical controls behind Tanzania's, Kenya's, the UAE's, Saudi Arabia's and the EU's data-protection rulesGPU sizing you can defend line by line, from model memory to data-centre powerA signed, verified supply chain that carries container images and 300 GB model weights across the air gapAn offline Kubernetes 1.37 cluster on Talos Linux, with Cilium default-deny networking and GPU nodesvLLM serving open-weight models behind one metered, multi-tenant AI gatewayKeycloak identity, Kyverno admission policy and tenant-isolated RAG over sensitive documentsSelf-hosted metrics, logs and traces, a redacted audit trail and runtime securityAcceptance tests in Swahili, Arabic and English that block a model upgrade when one language regressesBackups, disaster-recovery drills, zero-downtime upgrades and a handover checklistBuilt from production, not from slidesEvery chapter opens with a production failure: the chat UI that phoned home, the 300 GB transfer that arrived corrupted, the cache that served a confidential answer to the wrong user. Each one ends with the fix and a hands-on lab. All the code is in a public companion repository, linted and tested, with every version pinned.
Who should buy this bookInfrastructure and platform architects at ministries, regulators, banks and telecom operatorsSystems integrators bidding on on-premises AI tendersSenior DevOps engineers and SREs handed the mandate above and a rack of GPUsYou should be comfortable with Linux, Kubernetes and YAML. No machine-learning background is required.