What connects the model estate?
AI infrastructure is the collection of services that makes a model useful in an application. A runtime loads the model and generates responses. Other services route requests, schedule work, record experiments, store documents, and connect applications to tools. Understanding these jobs makes an architecture diagram easier to read and helps explain where operational and security decisions belong.
These essays move from the components to their relationships. A vector database matters because of the documents it holds and the applications allowed to retrieve them. A model server matters alongside its callers, workers, storage, and network access. Follow those connections to see which permissions are enforced, which assumptions are shared, and what evidence supports a claimed boundary.
Choose a starting point
-
Learning the landscape
Start with the field guide for an explanation of model runtimes, gateways, serving, orchestration, data, and agents.
A Field Guide to AI Infrastructure -
Building or operating systems
Read the connections essay alongside your own architecture. Identify the caller, operation, credential, and resource on each important connection.
The Tunnels Under the Surface -
Researching security
Use the SGLang case study to examine how a configuration change, a serialization choice, and an actual service boundary relate.
Disabling Pickle Did Not Remove Pickle
Read in sequence
-
Disabling Pickle Did Not Remove Pickle
Work through a SGLang case study that tests an internal service boundary against the running system.
8 min read
Take the question further
Treat a Model Download Like a Software Supply-Chain Decision
Follow the model-download connection from a repository to an approved deployment artifact.