Apache CloudStack — AI Factory Edition

There is a particular kind of letter that changes how an engineer thinks. For some it is a licensing renewal. Lately, for a growing number of us, it has been the notice that a model you had built into your product is being withdrawn, repriced, or placed behind an export control you cannot satisfy. A model you rent can be revoked by someone else's decision. A model you own cannot.

This book builds, on hardware you can buy today and rack this week, a local AI factory: a small, owned, sovereign platform that runs frontier-class open-weight models on your own iron, serves them to multiple users and tenants, and puts them to work earning their keep. No per-token meter. No revocation risk. Nothing leaves the building.

The platform is Apache CloudStack bent to a new purpose, across four tiers of mostly refurbished commodity iron:

• The Flex — the chapter that sells the book: two Dell R7525 servers redeployed from saved templates into a four-node inference cluster, a two-node aggregated cluster, or four isolated single-tenant boxes. Three machines from one pair of servers; the templates double as the backups. A sovereignty story you can hand to an auditor
• The quality brain — a frontier-class open model (roughly 744 billion parameters, mixture-of-experts) held in 768GB of system RAM via llama.cpp, with its honest tokens-per-second envelope on real DDR4
• An ARM model farm on an Ampere workstation, and a 100Gb RDMA inference fabric reality-checked rather than oversold — cross-host pairs, not a mythical clean mesh
• The brains put to work — retrieval over your own documents, fine-tuning small models with LoRA/QLoRA on a single 16GB consumer GPU, an n8n multi-agent support desk that triages, answers, guardrails and escalates, and an AI-first web front-end built by your own local coding model
• The regulated-firm case made concrete — multi-tenant isolation, data governance, and operating the whole estate by the Mars test: if this node were on Mars, could you still recover it?

The reader I have most in mind is the engineer inside an organisation whose data legally cannot leave its walls — finance, legal, healthcare, government. But it is equally for anyone who has watched rented infrastructure become a leash and decided to take it back. You do not need to be an AI researcher; you need a Linux command line, a lab to break things in, and the discipline to measure rather than assume.

The book's contract is honesty: every figure is measured, dated, labelled a vendor's claim, or flagged VERIFY for your own build. You do not need to be a hyperscaler to own a capable AI factory — and you should not pretend to be one.