How we appraise, and what the portal estimates are
MethodologyA Safewaterline appraisal is made unit by unit. The figures on this portal are something else: generic estimates by model and age, built from public data. This page explains both at a high level.
1. How we appraise
Each unit is appraised on its own evidence, not from a table. For every unit we confirm what it is and who owns it, read how it has been used, look at what it earns, find what comparable units sell for, and check what already has a claim on it.
- Identity. Serials matched to invoices, the site and the owner of record.
- Condition and use. Telemetry, hour meters, service logs and warranty.
- Earning power. Contracts, charters and utilization.
- Resale evidence. Listings, auction results and broker sales for the same model and age.
- Legal position. Lien searches, title and flag registry.
The result is an orderly liquidation value per unit and an indicative range for the whole pool. A person reads every appraisal before it goes out. While a loan is outstanding, the same readings are taken again each quarter. The checks for each asset class are listed on the home page.
2. The portal estimates
The figures on this portal are generic estimates for a model and age, built from public data. They give a market level picture for lenders and owners. They are not an appraisal of any unit, and they are not how a Safewaterline appraisal is made.
Fair value comes from a simple income model: what a unit of that model and age can still earn at today's public rental rate over the rest of its useful life, after running costs. Orderly and forced liquidation values allow for the cost and time of a broker sale over about 90 days or an auction over about 30 days. The advance figure is an indicative starting point. Every lender sets its own.
3. Status of a number
Each figure carries one of four statuses. The status says where it came from and how much to trust it.
| Status | Meaning | Example |
|---|---|---|
| listing | A public price read from a provider's page on a date, quoted and linked. | H100 at $2.99 an hour on a pricing page |
| model | A generic estimate from the portal model. | H100 fair value at 48 months |
| comp | An observed sale, auction or broker print, with date and channel. None published yet. | a lot of 256 H100 sold at auction |
| filings | Read from a public SEC filing, with the verbatim quote checked against the document. | a facility size in an 8-K exhibit |
A number with no status is not published. "verify" marks a desk estimate that is waiting on a document, such as a street price waiting on an invoice.
4. The rental input
The on demand rate per model is the median of recent public listings for that model. A listing is a price read from a provider's own pricing page, an archived copy of it, or a published survey, with the date, the quote and the link kept.
5. The risk score
The desk scores each class of asset or fleet on four factors at a quarterly review and publishes the reason beside each score. A score describes a class of asset or fleet, never a borrower.
- Obsolescence. How fast the next generation cuts this asset's earning power.
- Resale depth. How many buyers and channels there are, and how quickly a unit sells.
- Enforcement. How hard it is to find, recover, move and sell the asset after a default.
- Counterparty. Who pays the borrower, for how long, and whether that contract travels with the asset.
6. The register
The register is read from SEC EDGAR filings, one row per facility, with the verbatim quote kept for every cell. A cell whose quote cannot be found in the filing is left blank, and rows that cannot be verified are not shown. Amendments are folded under the original facility.
7. Limits
- No comps yet. Every portal value is a model. The curve becomes a benchmark as observed sales arrive.
- The portal model ignores the capacity contract. A unit under a long contract with a strong offtaker is worth more than the curve says. An appraisal reads the contract.
- One rate per model. Region, service level, cluster size and interconnect all move the real rate.
- Rack scale systems, NVL72 and later, are not priced per GPU on the open market. The B200 and B300 curves are the weakest on the page.
- No robot, equipment or ship curves. Those classes have a framework and a risk score only.
8. Change log
| Date | Change |
|---|---|
| 2026-10-30 | Rewritten as an overview: how we appraise, and what the portal estimates are. |
| 2026-10-28 | First publication. Six GPU models, portal model, liquidation values, risk score framework, register snapshot. |