When a pallet of decommissioned servers arrives at our lab, nothing on it is a HygeiaCloud GPU yet. It's a candidate. Roughly speaking, here's the gauntlet between arrival and a slot in the fleet.
Day 0: Intake and triage
Every unit is inventoried, photographed, and given an ID that follows it for life. We pull SMART data and error logs from the previous deployment where available, visually inspect for physical damage, replace thermal interface material, and clean out three years of datacenter dust. Storage from the previous owner never survives intake. Drives are wiped or destroyed before anything else happens.
Days 1–3: The burn-in
Then the abuse starts. For 72 hours, each GPU cycles through sustained full-power compute loads, memory stress patterns designed to surface failing VRAM, and thermal cycling that alternates between full load and idle to shake out marginal solder joints and fans. We're not trying to prove the card works. We're trying to make it fail on our bench instead of on your training run.
Cards that throw a single uncorrectable memory error, thermal-throttle below spec, or drift outside reference clocks are pulled. Some get repaired and re-run the full gauntlet from zero; the rest are responsibly recycled for parts.
Day 3: Benchmark against reference
Survivors run our benchmark suite head-to-head against factory-fresh reference hardware of the same model. To pass, a card has to land within a narrow band of reference performance. A second-life GPU that's 15% slower than it should be isn't a bargain; it's a defect. Passing cards get their results logged against their ID, and only then do they join the fleet.
That's the deal we make with every customer: the hardware had a first life, but the standards start from zero.