Sovereign AI — GPU Fleet Sizing

How many GPUs each model needs, what throughput you get, and how to share a fleet across tiers and teams. Default example: a large enterprise (for example a Nordic bank) with 80 H200s and 5,000 developers.

Loading Python… Results ↓

Headline

Starting the Python engine (first load takes a few seconds)…

    Models add any model; every row is sized for the default workload

    Serving pools workloads shift: give each pool its own model, context and GPUs

    Fleet sharing structure tiers for the primary model

    Throughput vs GPUs aggregate output tok/s per model, target line dashed

    Scenarios comparison each model on the whole inference share

    Model sizing detail (primary model)

    Cost & power

    Sources & assumptions

      Planning estimates, not benchmarks. Validate MBU/MFU on your own hardware before committing to a purchase or SLA. The calculation engine is plain Python (gpu_sizing.py) running in your browser with Pyodide; nothing you type leaves this page.