Every GPU can pass inspection. The cluster can still fail.
Independent inspection of the GPU network: cabling, link health, bandwidth and handover readiness, reported in terms a lender can act on.
Collateral checks stop at the server.
Covered today
- The GPUs and servers exist, with matching serials
- The servers are powered, healthy and in use
- The customer contracts are paying
Not covered
- Whether the cabling matches the design
- Whether the links are healthy, and have been for months
- Whether the cluster delivers its designed bandwidth
- Whether a new operator could run it after a default
A healthy network lets 2,000 GPUs rent as one training cluster. A damaged one forces them to rent in small pieces, at a lower price, while every server still reports healthy.
Checked against the design
The borrower's own cabling plan is the reference, so every finding states exactly where reality differs from what was claimed.
History, not a staged snapshot
The network already records months of link-error counters. That history is far harder to stage than a single inspection day.
Could someone else run it?
We check whether a new operator could take over after a default: topology, configuration, firmware, licences and runbooks.
Nothing to install to start
A first inspection works from standard tool output the operator already has. Access is read-only and never touches customer workloads.
Written for lenders
A scored verdict, closing conditions and covenant-style triggers. No dashboards to interpret.
Five areas, one report.
| Layer | Question | Example finding |
|---|---|---|
| 1 | Is all the network equipment present and on the collateral schedule? | 3 optics in service, not on schedule |
| 2 | Does every cable connect what the design says? | rails 3 and 4 swapped on 7 nodes |
| 3 | Are the links healthy, and have they stayed healthy? | 6 links running at half width |
| 4 | Is the network configured the way the borrower claims? | 2 switches one firmware release behind |
| 5 | Could a new operator take it over? | licence not assignable |
Scope
Agree the cluster, the design files and the access. Read-only, nothing installed.
Inspect
Compare the network against its design and its own error history. One bandwidth test, in an agreed maintenance window.
Report
A scored verdict, findings ranked by severity, and the conditions to meet before closing.
Already verifying servers? We cover what connects them.
Add a network section to the reports you already deliver, without building a network team.
Financing, buying or verifying a GPU cluster?
Tell us about the deal. We reply personally, usually within two business days.