GPU Audit Verification · For enterprise AI
Verify what your GPU system delivered.
We commission a GPU cluster independently and show, with evidence, whether it delivers what the contract says.
Taking design partnersDelivery commissioning
NVIDIA H100 cluster, acceptance window
- Delivered assetsMeasured
- Node healthMeasured
- FabricMeasured
- StabilityMeasured
- Thermal and powerMeasured
- Production-workload goodputMeasured
Delivered capacityMeasured per domain
MajorA node power cap below contract. Retest and holdback.
Drawdown certification
NVIDIA H100 cluster, draw milestone
- Serials match the asset listMeasured
- Systems powered and reachableMeasured
- Fabric matches the contracted railsMeasured
- Capacity under sustained loadMeasured
- Stable across test windowsMeasured
- Evidence bundle sealedHashed
Delivery CertificateReady to sign
NoneFinanced capacity delivered as specified.
Hardware inspection
NVIDIA H100 cluster, health and condition
- ECC and error countersMeasured
- NVLink and PCIe linksMeasured
- Thermal soak under loadMeasured
- Power deliveryMeasured
- Firmware and driversMeasured
- Consistency across nodesMeasured
Inspection reportReady for review
AdvisoryAn error counter trending upward. Watch item.
Buyers
Delivery commissioning. Check the cluster in the acceptance window, before you sign off.
Lenders
Drawdown certification. A signed Delivery Certificate at each draw milestone.
Insurers
An inspection report on the health of the hardware you cover.
The missing final level of commissioning.
Facility commissioning stops at power and cooling. We commission the GPU cluster.
Delivered assets
Model, count, serials, NIC rate and rails, reconciled against the contract.
Node health
PCIe link width and generation, ACS, GPUDirect RDMA, NVLink and ECC state.
Fabric
Port state, topology against the contracted rails, and error counters under load.
Stability
Every fabric test run in at least three separate time windows.
Thermal and power
A long soak under full load, to catch clocks that drop over time.
Production-workload goodput
How much useful work your own workload gets done on the cluster.
One verdict, and the evidence behind it.
- Accepted
- Accepted with exceptions
- Not accepted
- Indeterminate
Delivered capacity
The share of contracted achievable compute you got, measured per domain.
Exception register
Each exception graded Critical, Major, Minor or Advisory, with its retest.
Evidence bundle
The exact command, raw output and hash behind every finding.
Limitations
Generated, so a check that could not run is never read as a pass.
Seven steps, from contract to retest.
- 1
Review the contract and the decision
- 2
Agree access and test conditions
- 3
Reconcile the delivered assets
- 4
Run node, fabric and soak tests
- 5
Repeat in separate time windows
- 6
Report the verdict and exceptions
- 7
Review the fix, define the retest
Independent of the seller.
We run read-only and no data leaves your site. Every finding can be reproduced. When a seller pays for the work, we say so.
Accepting, financing or insuring a GPU cluster?
Tell us what you need to check, and when.