The index, recomputed at every hourly read of its sources.
The published price is still the 11:00 UTC fixing, computed from one collection a day. The line below is the same calculation, under the same methodology version, replayed over reads taken through the day. It is research beside the index: it is not a print, it is never stored as one, and it is not a reference price for any financial instrument.
TCI-CRI-H100 through the day
USD per GPU-hour, times in UTC. A break in a line is a read that gapped, a sweep that did not run, or for a constituent a read it was not part of; nothing is drawn across it. Each constituent is drawn at its own price in the index at that read.
Last 24 hours
Last 7 days
Recomputed at each read · 11:00 UTC fixing as published
- digitalocean$4.41 · 24h +$0.00 (+0.00%)
- lambdalabs$3.99 · 24h +$0.00 (+0.00%)
- nebius$3.85 · 24h +$0.00 (+0.00%)
- runpod$3.49 · 24h +$0.00 (+0.00%)
- scaleway$3.64 · 24h +$0.00 (+0.00%)
- seeweb$2.16 · 24h +$0.00 (+0.00%)
- vast.ai$2.27 · 24h +$0.00 (+0.00%)
- verda$3.48 · 24h +$0.00 (+0.00%)
- datacrunch (grey: no colour slot left)not in the index in the last 24 hours
Table view
| Read | From | USD per GPU-hour | Providers | Gap reason |
|---|---|---|---|---|
| 26 Sep 2026, 11:17 UTC | fixing reads | 3.49 | 8 | |
| 26 Sep 2026, 10:13 UTC | hourly sweep | 3.49 | 8 | |
| 26 Sep 2026, 09:14 UTC | hourly sweep | 3.49 | 8 | |
| 26 Sep 2026, 08:18 UTC | hourly sweep | 3.49 | 8 | |
| 26 Sep 2026, 07:14 UTC | hourly sweep | 3.49 | 8 | |
| 26 Sep 2026, 06:23 UTC | hourly sweep | 3.49 | 8 | |
| 26 Sep 2026, 05:14 UTC | hourly sweep | 3.49 | 8 |
Constituents at each read
| Read | digitalocean | lambdalabs | nebius | runpod | scaleway | seeweb | vast.ai | verda | datacrunch |
|---|---|---|---|---|---|---|---|---|---|
| 26 Sep 2026, 11:17 UTC | 4.41 | 3.99 | 3.85 | 3.49 | 3.64 | 2.16 | 2.27 | 3.48 | — |
| 26 Sep 2026, 10:13 UTC | 4.41 | 3.99 | 3.85 | 3.49 | 3.64 | 2.16 | 2.27 | 3.48 | — |
| 26 Sep 2026, 09:14 UTC | 4.41 | 3.99 | 3.85 | 3.49 | 3.64 | 2.16 | 2.27 | 3.48 | — |
| 26 Sep 2026, 08:18 UTC | 4.41 | 3.99 | 3.85 | 3.49 | 3.64 | 2.16 | 2.27 | 3.48 | — |
| 26 Sep 2026, 07:14 UTC | 4.41 | 3.99 | 3.85 | 3.49 | 3.64 | 2.16 | 2.27 | 3.48 | — |
| 26 Sep 2026, 06:23 UTC | 4.41 | 3.99 | 3.85 | 3.49 | 3.64 | 2.16 | 2.27 | 3.48 | — |
| 26 Sep 2026, 05:14 UTC | 4.41 | 3.99 | 3.85 | 3.49 | 3.64 | 2.16 | 2.27 | 3.48 | — |
The rest of the family, last 7 days
Only series with at least one value in the window are drawn. A series that gapped at every read is left out rather than drawn flat.
TCI-CRI-H100-NC
- digitalocean$4.41 · 24h +$0.00 (+0.00%)
- lambdalabs$3.99 · 24h +$0.00 (+0.00%)
- nebius$3.85 · 24h +$0.00 (+0.00%)
- scaleway$3.64 · 24h +$0.00 (+0.00%)
- seeweb$2.16 · 24h +$0.00 (+0.00%)
- verda$3.48 · 24h +$0.00 (+0.00%)
A settlement window, against the fixing
The mean of every reconstructed TCI-CRI-H100 value from 08:00 to 11:00 UTC, beside the 11:00 fixing as published. At least 2 values are required; with fewer the window has no value, and no earlier one is carried in. Shown to measure what averaging would change. The published price remains the fixing.
| Date | Window | Reads | Fixing | Window minus fixing | As % of fixing | Why empty |
|---|---|---|---|---|---|---|
| 26 Sep 2026 | 3.49 | 3 | 3.49 | +0.00 | +0.00% |
Sources at the latest read
How old each source's read was at 26 Sep 2026, 11:17 UTC. A source past its limit drops out of the reconstruction, which can then gap exactly as the fixing would with that source missing. A source that keeps failing is asked less often, the wait doubling with each miss, until one read succeeds.
| Source | Read | Counts for up to | Age at latest read | Failing |
|---|---|---|---|---|
| azure_retail | every 12 h | 25 h | 0 min | |
| civo | every 6 h | 13 h | 0 min | |
| coreweave | every 6 h | 13 h | 0 min | |
| crusoe | every 6 h | 13 h | 0 min | |
| digitalocean | every 6 h | 13 h | 0 min | |
| gpuhunt | every 1 h | 3 h | 3 min | |
| hyperstack | every 1 h | 3 h | 0 min | |
| lambda_pricing | every 6 h | 13 h | 0 min | |
| latitude | every 6 h | 13 h | 0 min | |
| ovh | every 6 h | 13 h | 0 min | |
| runpod | every 1 h | 3 h | 3 min | |
| scaleway | every 6 h | 13 h | 3 min | |
| seeweb | every 6 h | 13 h | 0 min | |
| static_yaml | every 1 h | 3 h | 3 min | |
| vast_ai | every 1 h | 3 h | 3 min | |
| voltagepark | every 6 h | 13 h | 0 min |
How the line is made
Each source is read at the rate its price can move. Marketplace books and stock-driven feeds are read every hour; rate cards on pricing pages every six hours. The fixing's own 11:00 read counts as a read, so the hour after it asks nobody for the same page twice.
At every sweep, the newest read of each source is taken, provided it is younger than that source's limit. Those rows go through the fixing's normalisation and estimator unchanged, under the methodology version in effect on the date, with the panel, the node-size floor, the provider gate and the trim that version sets. A read below the provider gate is a gap here as it would be at 11:00.
The reads are kept in an append-only store in the repository, one file per UTC day, so every value on this page can be recomputed from what was read. Nothing in that store reaches a published print.