You track 16 providers. Here is what we actually hold, provider by provider.
GPUAdvisor's 16 publicly listed providers, checked one by one against our public ledger: where we hold device-level evidence, where we only have model-level quotes, and where we have nothing. Where we have nothing, it says nothing. Every figure is aggregated live from the ledger (GET /v1/coverage) — pull it yourself, or download the ledger and recompute it with /cvin_replay.py.
Provider by provider: your 16
| Provider | What we hold | Devices / model IDs | Records | Models | Source of evidence |
|---|---|---|---|---|---|
| AWS | none | — | — | — | platform does not expose individual units |
| Google Cloud | none | — | — | — | platform does not expose individual units |
| Azure | model only | 424 model IDs | 424 | 6 | azure.microsoft.com |
| Lambda | model only | 133 model IDs | 209 | 14 | rentgpu.link · shadeform.com |
| CoreWeave | none | — | — | — | platform does not expose individual units |
| RunPod | model only | 103 model IDs | 103 | 62 | rentgpu.link · www.runpod.io |
| Nebius | model only | 25 model IDs | 34 | 9 | rentgpu.link · shadeform.com |
| Together AI | none | — | — | — | not sold per-device |
| Crusoe | model only | 36 model IDs | 51 | 9 | rentgpu.link · shadeform.com |
| Hyperstack | model only | 39 model IDs | 96 | 11 | rentgpu.link · shadeform.com |
| Thunder Compute | model only | 4 model IDs | 4 | 4 | rentgpu.link |
| Daytona | none | — | — | — | not sold per-device |
| Enverge | none | — | — | — | collectable, not yet wired |
| Lium | none | — | — | — | collectable, not yet wired |
| FluidCore | none | — | — | — | collectable, not yet wired |
| Atlantic.Net | none | — | — | — | collectable, not yet wired |
A "model ID" is not a device: it encodes cloud × model × region × billing mode. Good for price comparison, useless as collateral evidence. See section 03.
What we add: the 37 platforms you do not track
| Platform | Level | Devices | Model IDs | Records |
|---|---|---|---|---|
| vast.ai | device | 4,585 | — | 9,376 |
| clore.ai | device | 2,933 | — | 5,653 |
| jygpu（极云） | model only | — | 2 | 699 |
| Massed Compute | model only | — | 175 | 290 |
| Verda | model only | — | 69 | 179 |
| 智星云 GPUSHARE | model only | — | 6 | 124 |
| Vultr | model only | — | 73 | 75 |
| Prime Intellect | model only | — | 69 | 69 |
| Paperspace | model only | — | 67 | 67 |
| Voltage Park | model only | — | 1 | 64 |
| Excess Supply | model only | — | 34 | 60 |
| Scaleway | model only | — | 25 | 40 |
| AutoDL | model only | — | 11 | 33 |
| IMWT | model only | — | 28 | 28 |
| Akash | model only | — | 23 | 23 |
| Latitude | model only | — | 18 | 22 |
| Denvr | model only | — | 6 | 22 |
| 共绩算力 | model only | — | 18 | 18 |
| DigitalOcean | model only | — | 13 | 15 |
| Boostrun | model only | — | 9 | 12 |
Top 20 shown (device-level first, then by volume); the full list is in GET /v1/coverage. Several of these (jygpu, gpushare, autodl, matpool, compshare) are major China-market suppliers — absent from your price catalogue.
Three kinds of "we do not have it"
Per unit, verifiable
A specific record for this unit: where, at what price, when it appeared, whether a lien was filed. Resolvable and monitorable — the starting point for underwriting. Everything we have collected is L0 listing fingerprint: it proves "this unit appeared at this price", not its serial number (that is L1/L2).
Good for pricing, not for collateral
Only "cloud X lists model Y at $Z/hr". That is your home turf — so we do not duplicate it; device-level only matters when the question becomes "which machine is this".
And the reasons differ
Not exposed: AWS, Google Cloud, CoreWeave — sold through direct contracts, no per-unit data in public catalogues. Nobody can do device-level here. Different product shape: Together AI, Daytona. Collectable, not yet wired: Enverge, Lium, FluidCore, Atlantic.Net — these can be added.
Verify one, live
The most-observed device in the ledger — open it directly:
cvin:nv.rtx5070.pcie.12g:2y0otc4txzm13pi9 · 60 records
- Single device, free:
GET /v1/resolve/{cvin} - No CVIN? Query by serial, platform link or asset tag:
GET /v1/lookup?q=… - Up to 200 at once:
POST /v1/resolve/batch - All figures on this page, machine-readable:
GET /v1/coverage
Where this ends
We are not trying to turn your 16 into our 16 — that is your job, and you do it better. Device-level verification only matters when someone is making a promise about a specific machine: collateral, resale, custody, insurance, dispute. In those settings the machine is not on an AWS invoice — it is on a list, in a seller's hands, on a rack in a colo.
So the most natural first step is: at the exit of your quote-benchmarker / procurement flow, add one check — "has this batch been verified?" One GET does it, no architecture change, and it does not compromise your vendor-neutral position: the registry is public, signed, and recomputable by anyone, so using it cannot be mistaken for being bought.