Updated 2026-09-10 · re-ranked weekly
The Disclosure Leaderboard
Who publishes enough about their fabric to be underwritten. Ranked by disclosed training-class offerings (Grade A/B per ASSAY-1), then by disclosure. This ranks what sellers state in writing — not measured performance. No fleet holds a measured Verified badge yet. Harness v0.1 has shipped — run it on your own metal; the first Verified fleets will be announced in the weekly report. Measured performance that does exist in public (inference benchmarks, source-classed by independence) is on the measured tier. Operators: get graded.
| # | Provider | Tracked offerings | Grade A/B | Disclosure | Cheapest A/B $/GPU-hr | Undisclosed | Incidents (90d, self-reported) | Self-serve |
|---|---|---|---|---|---|---|---|---|
| 1 | AWS | 15 | 14 | 93% | $6.88 | 1 | no feed | ✓ |
| 2 | Microsoft Azure | 9 | 9 | 100% | $10.60 | 0 | 0 | partial |
| 3 | Nebius | 6 | 6 | 100% | $3.85 | 0 | 14 | ✓ |
| 4 | Lambda | 10 | 6 | 100% | — | 0 | 12 | ✓ |
| 5 | CoreWeave | 6 | 5 | 83% | $6.16 | 1 | no feed | partial |
| 6 | Crusoe | 4 | 3 | 100% | $2.30 | 0 | 6 | ✓ |
| 7 | Oracle Cloud (OCI) | 4 | 3 | 100% | $10.00 | 0 | no feed | ✓ |
| 8 | Together AI (GPU Clusters) | 12 | 2 | 67% | $8.19 | 4 | 1 | partial |
| 9 | Google Cloud | 12 | 0 | 100% | — | 0 | 4 | partial |
| 10 | TensorDock | 5 | 0 | 100% | — | 0 | no feed | ✓ |
| 11 | Hyperstack | 10 | 0 | 100% | — | 0 | 4 | ✓ |
| 12 | OVHcloud | 4 | 0 | 100% | — | 0 | 50+ | ✓ |
| 13 | Voltage Park | 1 | 0 | 100% | — | 0 | no feed | ✓ |
| 14 | DataCrunch (rebranded to Verda, verda.com) | 14 | 0 | 93% | — | 1 | 4 | — |
| 15 | Scaleway | 5 | 0 | 80% | — | 1 | 50+ | ✓ |
| 16 | RunPod | 6 | 0 | 67% | — | 2 | 24 | ✓ |
| 17 | Vast.ai (live API) | 5 | 0 | 40% | — | 3 | no feed | ✓ |
| 18 | Prime Intellect | 5 | 0 | 0% | — | 5 | 5 | ✓ |
Incident counts come from each provider's own public status feed (last 90 days, collected weekly) — a floor, not a ceiling, and granular reporters (per-rack incidents) look worse than silent ones. "No feed" means the provider publishes no machine-readable incident history at all; full history on each provider page.
How to move up
Publish your interconnect: fabric type, per-GPU bandwidth, topology, and the largest contiguous allocation you will actually sell. Grades are recomputed from public sources weekly — no submission required, no fee, no exceptions. Disagree with a grade? The methodology is open; write to us with a source and we will correct in public.
The leaderboard moves when disclosure moves
The weekly report tracks who rose, who fell, and why.
Subscribe free