Apache CloudStack — AI Factory Edition

Introduction — Read This First

1. What this book is
2. Who it is for
3. The thesis: you do not need to be a hyperscaler
4. How to use this book
5. The shape of the journey
6. A method, not a product

Apache CloudStack — The Sovereign AI Factory

1. On owning what you run

The Book in Full — Annotated Contents

1. Part I — Build
2. Part II — The Flex
3. Part III — The Brains
4. Part IV — Make It Pay
5. Front and back matter
6. The expanded edition

Chapter 1 — Why Own an AI Factory

1. The Story: the second letter
2. The thesis: you do not have to be a hyperscaler
3. What this factory is — and what it is not
4. The strongest reason: data that cannot leave the building
5. The live hook: the week a model could be pulled
6. GYOCYO: Grow Your Own, Cook Your Own
7. Return on Research
8. The Mars heuristic
9. The map of the journey ahead
10. The honesty contract

Chapter 2 — The Build of Materials

1. The Story: counting what the building already holds
2. How to read this bill of materials
3. Tier 1 — The responsive tier: two Dell R7525s
4. Tier 2 — The quality brain: two Dell R740XDs, 768GB each
5. Tier 3 — The ARM model farm: one Ampere Ultra
6. The graphics-card pool, counted honestly
7. Tier 4 — The management plane: an R660, or a pair of Raspberry Pis
8. Where these four tiers live: three clusters in one zone
9. Proving the iron: a hardware-readiness pass
10. What is not on the headline list — but you still need
11. The shape of the cost — without a figure

Chapter 3 — Racking, Power and the Off-Grid Interlock

1. The Story: four of a kind
2. The rack as a machine with a shape
3. Power is not a wall socket
4. Measure the draw: a power-and-thermal readiness pass
5. The factory that parks itself
6. The off-grid interlock
7. Every watt on purpose

Chapter 4 — Network Fabric: the Quad-25Gb Front-End, the Back-to-Back 100Gb, and Where CloudStack Ends

1. The Story: the question on the wall
2. Two fabrics, one estate
3. The quad-25Gb front-end: the network CloudStack governs
4. The back-to-back 100Gb: short, switchless, and honest about its shape
5. Where CloudStack ends
6. Bring up the fabric: a network-readiness pass
7. Every wire on purpose

Chapter 5 — The Management Plane: the Conductor of the Estate, Built Two Honest Ways

1. The Story: the conductor they already had
2. What the conductor actually holds
3. The caveat, stated plainly: one node is no HA
4. Option A — the datacentre default: a Dell R660
5. Option B — the low-power showcase: two Raspberry Pi 5s
6. The HA aside: keepalived, HAProxy, and a replicated database
7. A note on management traffic, and a warning about the keys
8. Bring up the conductor — a management-node readiness pass
9. The Mars test, applied to the conductor
10. Built, not yet installed

Chapter 6 — CloudStack on the R7525s: KVM, and the GPU and NIC Made Schedulable

1. The Story: the hosts that waited
2. Installing against the conductor
3. KVM on the R7525s
4. A zone, a pod, a cluster
5. Making the iron schedulable
6. Bring the platform up — an install-and-passthrough readiness pass
7. The Mars test, applied to the install
8. An orchestra, ready to be told what to play

Chapter 7 — The Flex: Profiles A, B and C as Templates

1. The Story: the cluster that left
2. The template is the unit of the flex
3. Switching is destroy-and-redeploy, not live-migrate
4. Profile A — the four-node vLLM cluster
5. Profile B — the two-node aggregated pair
6. Profile C — four independent tenants
7. Placement is your job, not the scheduler’s
8. The flex as code
9. Build the flex — a profiles-and-templates readiness pass
10. Three machines, one rack

Chapter 8 — The RDMA Inference Fabric: RoCEv2 Reality-Checked, vLLM Across the Cluster

1. The Story: the number nobody would say
2. The link is RDMA, not merely a fast port
3. The switchless pair is the easy case — say so
4. The driver question: in-box first
5. GPUDirect RDMA: assume it is not there
6. vLLM across the pair: what distribution actually buys
7. The measured reality: a method, not a number
8. Build the fabric — an RDMA-and-vLLM readiness pass
9. The fast wire, reckoned with

Chapter 9 — The Quality Brain: GLM-5.2 (and Kimi) on the 768GB R740XDs via llama.cpp

1. The Story: the quiet join
2. The tier, and why it is a separate machine
3. Why this model: ownership you cannot have revoked
4. What GLM-5.2 actually is
5. The engine: llama.cpp, and the honest size problem
6. How llama.cpp spends the one GPU
7. Speed, told straight
8. The line-up: one brain, or two, and which ones
9. What this tier is for — and what it is emphatically not
10. Settled, and VERIFY-on-live
11. Build the brain — a quality-brain readiness pass
12. The brain that answers the hard ones

Chapter 10 — The ARM Model Farm: Ampere Ultra, CPU-or-GPU

1. The Story: the odd one in
2. The tier nobody else covers
3. Why ARM, on purpose
4. The multi-architecture zone, made real
5. CPU or GPU: the two ways this box serves
6. The engines that run clean on arm64
7. What this tier is for, and what it is not
8. Settled, and VERIFY-on-live
9. Build the farm — an ARM-farm readiness pass
10. The third tier, in its place

Chapter 11 — Corpus and RAG: Retrieval Over Your Own Documents

1. The Story: the one who asked what the data said
2. Two ways to make a model know something
3. The pipeline, end to end
4. Ingest and chunk: cutting the corpus to size
5. Embed: the second, smaller model
6. Store: Qdrant, the vector index
7. Retrieve and rerank: similarity is not relevance
8. Ground: which brain answers, and from what
9. What RAG is, and what it is not
10. The sovereign half
11. Build the pipeline — a RAG readiness pass
12. Settled, and VERIFY-on-live
13. The half that knows your documents
14. The Story: a promise with a number waiting

Afterword - The estate is yours now

1. About the author
2. Stay in touch

Glossary

Appendix A — Bill of Materials

1. The four tiers at a glance
2. Tier 1 — Responsive / flex compute: Dell PowerEdge R7525 (×2)
3. Tier 2 — Quality brain: Dell PowerEdge R740XD (×2)
4. Tier 3 — ARM model farm: Ampere Ultra workstation (×1)
5. Tier 4 — Management: two documented options
6. The RDMA fabric
7. On power and cooling