❯ ./book --status coming-soon ▌
/book
// Straight answers on when to run models on your own infrastructure, when to rent, and when to do neither.
Self-Hosting LLMs in Production
// Deciding, sizing, securing and operating open-weight models on-premises, in colocation and on dedicated cloud.
Why this book exists
There is a conversation that keeps repeating itself in enterprise technology, and it always opens in the same way. Someone senior asks, “We want to run our own models. What do we need?” The answer that comes back is a list of hardware. Hardware happens to be the easy part. It is expensive, it is presently in short supply, and it will depreciate faster than any of us would like, but at least it arrives with a datasheet.
What actually kills a self-hosting programme is the set of things that have no datasheet at all: the evaluation harness nobody built, the gateway that was postponed to phase two, the licence clause that went unread until due diligence, the driver and CUDA version tower that consumed a long weekend, the token spend that could not be attributed to any team, and the quantisation change that quietly cost twelve points of extraction accuracy and went unnoticed for six weeks.
This book therefore treats hardware as one part out of six. The other five are the decision itself, the physics of capacity, the software architecture, the operations, and the economics, security and governance that surround all of it.
Hardware is one part of six
// The book is organised around the six concerns that actually decide whether a self-hosting programme succeeds:
- 01The decision — run, rent, or neither
- 02The physics of capacity
- 03The software architecture
- 04The operations
- 05Hardware
- 06Economics, security & governance
❯ The book is in progress — first edition coming 2026.