Index v0
Capacity Index
Launches since Sep 21, 2026 · as of Oct 11, 2026, 17:34 UTC · 439 launches · 261 ready · 80 No machine started · 98 machine did not become ready
Most launches in this data are Kilawatt's own tests.
- Internal
- 416
- Sponsored
- 15
- Outside paying wallets
- 8
- unclassified
- 0
All launches
| GPU type | Launches | Ready | No machine started | Machine did not become ready | Ready rate |
|---|---|---|---|---|---|
| RTX A4000 | 279 | 157 | 44 | 78 | 56.3% |
| A40 | 44 | 38 | 4 | 2 | 86.4% |
| RTX 3090 | 31 | 7 | 22 | 2 | 22.6% |
| A10 | 27 | 23 | 1 | 3 | 85.2% |
| A100 | 21 | 15 | 0 | 6 | 71.4% |
Not enough data yet (fewer than 20 launches): RTX 4090, RTX A5000, H100, A6000, B200, B300, A5000, L40S
Normal routing · 110 launches
| GPU type | Launches | Ready | No machine started | Machine did not become ready | Ready rate |
|---|---|---|---|---|---|
| RTX A4000 | 58 | 29 | 5 | 24 | 50.0% |
| A100 | 21 | 15 | 0 | 6 | 71.4% |
Not enough data yet (fewer than 20 launches): RTX A5000, A40, H100, A10, RTX 4090, B200, B300, RTX 3090, A5000, L40S
Counts launches on real production paths, including test traffic. Not a customer success rate. Suppliers are not named. Time-to-ready and failover counts are not published yet because the data is not tracked reliably.
x402 demand log began collecting on Oct 11, 2026. A demand summary will appear here once there is enough data.
Methodology
- Launch: one GPU machine placed on a real production path (console, API key or x402), including test traffic.
- Ready: a launch whose machine passed a health check at least once (a healthy provider check or a healthy in-machine agent report). Acceptance by the underlying infrastructure alone does not count as ready.
- Never ready: a launch that never passed a health check. Since Sep 27, 2026, any charge on these launches is returned automatically as account credit.
- Data note: before Sep 27, 2026 the automatic refund did not exist. 24 early launches by internal test accounts were charged 30 cents in total and were not refunded.
- No machine started: the order could not be placed, so no machine was ever created. Most of these are a supplier reporting no stock or no capacity; a few were expired offers or placement errors on our side.
- Machine did not become ready: a machine was started but never passed a health check before it ended.
- Never ready is the sum of those two kinds.
- Normal routing: launches placed by the normal router. Forced test runs, where placement was pinned to one supplier or a temporary test exemption was active, are excluded from this view but included in All launches.
- Ready rate: ready launches divided by all launches for that GPU type.
- Window: launches created from Sep 21, 2026 (UTC) up to the as-of time.
- Normalization: GPU type names are lowercased, trimmed, spaces become hyphens and a leading 'nvidia-' is removed before grouping, so 'NVIDIA RTX A4000' and 'rtx-a4000' are one row.
- Minimum sample: a GPU type with fewer than 20 launches shows 'Not enough data yet' instead of a rate.
- Owner classes: internal (Kilawatt test accounts and wallets), sponsored (wallets classified as sponsored), outside (other accounts and paying wallets), unclassified (wallets not yet reviewed).
- Refresh: computed live from the production database and cached for 5 minutes.
Machine-readable: /api/public/capacity-index.json · See also Super Intelligence reports