It's the first question every skeptical engineer asks us, and it deserves a real answer rather than a vibe: does a GPU that spent three years in someone else's datacenter perform like one fresh out of the box?
How we measure
Every certified card runs our benchmark suite (sustained training throughput, inference latency under load, memory bandwidth, and thermal behavior over multi-hour runs) head-to-head against factory-fresh reference hardware of the same model that we keep specifically for comparison. Not spec sheets: our reference cards, on our benches, same power and cooling.
What the data shows
Silicon doesn't get slower with age: it either works or it doesn't. Cards that pass our gauntlet cluster tightly around reference performance, which is exactly what the physics predicts: transistors don't wear like tires. The components that do age (fans, thermal interfaces, capacitors on the power stages) are precisely the ones we inspect, service, or replace at intake.
The honest caveat: an older architecture is an older architecture. A second-life A100 performs like an A100, not like this year's flagship. Our claim was never "old cards magically match new ones." It's that a certified A100 at a fraction of hyperscaler price is the best value in compute for a huge class of workloads. The benchmarks back that up.
Skeptical? Good. That's the right default in this market. Ask us for the numbers on the tier you're considering.