Technical Due Diligence for GPU Clusters
What to check before you fund.
What it covers
Technical due diligence answers two questions an invoice cannot: is the equipment priced at the market, and can the site keep it running at the uptime the customer contract requires. Most failures we see start at the second.
What you should check for
| Equipment and price | Line-by-line BOM | Every server, switch, optic and support line itemized; zero-priced lines documenting the configuration inside a server line are not added again. |
| Quote against the market | Each major line checked against completed resale data and current new prices. One quote we reviewed listed the GPU board at 2.8x market and a 64GB DIMM at 24x. | |
| Configuration is standard | Vanilla configurations clear faster and higher than bespoke ones; unusual pooling, networking or mixed silicon shrinks the resale buyer pool. | |
| Site fit | Power per rack | Committed kW per rack covers the systems: about 10.2 kW for a DGX H100, 14.5 kW for a DGX B300, about 120 kW for a GB200 NVL72 rack. |
| Cooling type | Air cooling stops working past roughly 40 kW per rack; liquid-cooled systems need the plumbing, CDUs and floors in place, not planned. | |
| Redundancy against the SLA | Uptime Tier I and II sites allow 23-29 hours of downtime a year, Tier III about 1.6. Compare that with the uptime in the offtake contract. | |
| Energization date in writing | The colo's power date, signed, with rent starting near delivery. | |
| Operations | Burn-in and acceptance | Testing time is in the schedule and the offtaker's acceptance test is defined before the ready-for-service date. |
| Spares and RMA | Warranty terms, spares on site and who swaps a failed GPU, so one dead board does not breach the SLA. | |
| Refresh costs | For operating fleets, the cost and timing of the next refresh against the remaining term. |
Sources: American Compute, NVIDIA.
Why GPU clusters need their own technical review
A GPU cluster is a server fleet, a network fabric and a power and cooling load that most buildings were not designed for. GPU racks draw 40-80 kW, while most data centers run 10-30 kW per rack.
Many sites lack what customer contracts require, especially retrofitted crypto mines and modular builds. A site that cannot hold the contracted uptime can force a liquidation.
Over a third of the 110 data center projects due in 2025 missed their dates, and power is the bottleneck as of early 2026. A technical review prices that schedule risk before funding, not after.
Red flags
| Liquid-cooled GPUs, air-cooled hall | The cluster cannot reach full power until a retrofit nobody has priced. Very few operators accept running new silicon at half power. |
| Lump-sum network and racks line | It hides equipment value and margin; on one quote it was 11% of the equipment value. |
| Self-assessed tier | A tier missing from the Uptime Institute's registry is the operator's own claim. |
| Uptime SLA above what the site can deliver | SLA credits and termination rights land on the borrower, and the lender inherits them. |
Discuss a transaction
Send us the equipment list and the loan or lease you have in mind. We will come back with an indication and the questions we would need answered.
Prefer email? hello@amcompute.com