Platform and AI Infrastructure: Capacity, Cost and Control
Design a supported developer platform, account for its costs and operate inference workloads. Practise GPU admission, model rollouts and bounded automation in 2026.
Design a supported developer platform, account for its costs and operate inference workloads. Practise GPU admission, model rollouts and bounded automation in 2026.
Work through the modules in order. Each lesson ends with an exercise and a retrieval check; keep your answers before revealing the explanation. The final project brings the earlier decisions together in a design or incident record you can discuss aloud.
One introductory lesson is public, one lesson opens with free sign-in, and ten require Premium. Progress is a personal reading record. Self-checks and project rubrics do not award a professional certification. All numerical scenarios are teaching fixtures, not reported production results. Commands are read-only or run against local fixtures unless a lesson explicitly describes an optional sandbox.
Start with Incident Response and Reliability Engineering if the prerequisites are unfamiliar.
When you finish, you can
- Create a bounded platform exception with ownership, compensating controls and an exit path, avoiding permanent unsupported forks hidden behind self-service.
- Compare full-device allocation, hardware partitioning and time slicing in 2026, with explicit memory and failure-isolation requirements for different tenants.
- Design a model release gate that keeps artifact provenance, task quality, safety checks and serving SLOs separate, including a bounded canary and rollback contract.
- Integrate a support contract, GPU memory calculation, admission policy, cost sensitivity and agent authorization into a final design review with a scored rubric.
BEFORE YOU START · Complete the reliability course. Basic Python and container knowledge; no GPU hardware required.
Syllabus
Build a supported platform
Define a platform path with a support contract. Allocate shared costs without hiding idle capacity. Design an exception that can expire.
Account for capacity
Budget cold start before relying on autoscaling. Calculate GPU memory admission from the request mix. Choose GPU sharing by its isolation contract.
Operate model serving
Separate first-token delay from generation speed. Admit inference work by token and deadline budgets. Roll out a model with quality and serving evidence.
Bound automation and defend the design
Bound an incident agent before granting tools. Compare serving options with failure and cost included. Capstone: defend a platform and inference design.