AI in the Data Center

A trained model only creates value once it's served. Inference plays by different rules than training: millions of short requests, latency sensitivity, cost per request to minimize. This module covers serving and batching (dynamic, continuous), quantization and the KV cache that cut cost, then multi-tenant autoscaling and observability that keep a service reliable and profitable.

Building a cluster is one thing; running it sustainably is another. This module covers the real economics of AI (costs, energy, PUE), workload orchestration via Kubernetes and Slurm, fleet-wide observability, and data sovereignty under GDPR for French data centers. It closes with a guided five-step capstone: sizing a cluster end-to-end for a real-world case.