One hundred certifications in, our process has stopped being a plan and started being a body of evidence. Some of what we expected held up. Some didn't. In the spirit of building this company in the open, here's the honest version.
Most enterprise hardware really is fine
The core bet, that scheduled decommissioning discards mostly-healthy hardware, survived contact with reality. The large majority of cards that arrive at our lab pass the full gauntlet and benchmark within our reference band. Enterprises run conservative thermals and clean power; it shows.
The failures cluster where you'd least enjoy finding them
The cards that fail rarely fail loudly. The classic pattern is a card that boots fine, runs fine for hours, then throws a single memory error deep into a sustained load: exactly the kind of fault that would silently corrupt a training run. This is why the burn-in is 72 hours and not a quick smoke test. The expensive failures are the patient ones.
Dust is the real enemy
Our most consistent finding is unglamorous: thermal performance on arrival correlates with dust more than with age. Three years of datacenter particulate plus tired thermal paste costs real headroom, and a teardown, clean, and repaste brings most cards back within a degree or two of reference. A shocking amount of "degraded" hardware is just dirty hardware.
The next hundred will teach us more. If you want the fleet those hundred cards built, the plans are here.