Methodology & known limitations
This tool reports a relative availability signal, not absolute instance counts. AWS does not expose how many instances exist or are free, and there is no free API that tests whether a launch would succeed.
What is measured
- Spot Placement Score (1–10) — AWS's own likelihood that a spot request for the instance type would be fulfilled in a given zone, sampled over time. This is the primary signal.
- Offering presence — whether an instance type is offered in a zone at all (a boolean, not a quantity).
GPU capacity blocks (purchasable ML/GPU capacity windows) are a planned signal that is currently deferred, so they are not collected or shown yet.
Why not instance counts?
Estimating real On-Demand counts would require creating Capacity Reservations or launching instances — both cost money and actually consume capacity during the probe. This project is intentionally free-signal-only, so it does not do that.
Known limitations (accepted)
- Spot Placement Score is a recommendation, not a guarantee, and varies per AWS account.
- AWS returns only the top-scored zones per query; unscored zones show as gaps.
- Coverage: offering presence is tracked for the full GPU/accelerator catalog (NVIDIA, AMD, Inferentia, Trainium, Gaudi, Qualcomm) across all enabled regions. The Spot Placement Score signal covers a curated subset of those types, because AWS caps the number of distinct score configurations per account per day.
- Availability visible to one customer may differ from another; large accounts may see more.
- Availability-zone identity is shown by stable
AZ-ID(e.g.use1-az1) because zone names are randomized per account.
Cadence
A probe samples all tracked instance types across all regions and zones every 15 minutes, building the history shown on the dashboard.