When people picture the climate cost of AI, they picture electricity: humming racks, spinning meters, power plants somewhere upstream. That's real, but it's only half the story, and for modern accelerators it may not even be the bigger half.
A meaningful share of a server's lifetime emissions is embodied: the mining, refining, fabrication, assembly, and shipping that happen before the machine is powered on for the first time. Semiconductor fabrication in particular is extraordinarily resource-intensive, ultra-pure water, specialty chemicals, and energy-hungry cleanrooms, concentrated in a handful of fabs running flat out.
Why this changes the second-life math
Here's the implication people miss: when you extend the life of an existing GPU, you don't just delay a purchase. You avoid re-paying an embodied-carbon bill that has already been paid. A rescued A100 arrives with its manufacturing footprint fully sunk. Every additional GPU-hour it delivers amortizes that footprint further, while a new card starts the meter from zero.
That's why "reuse before recycle" isn't a slogan for us. It's an ordering of operations. Recycling recovers materials; reuse recovers the whole manufactured object, footprint and all. The greenest GPU is, overwhelmingly, the one that already exists.
We're building our fleet on that arithmetic. If your organization carries embodied-carbon targets and wants compute that helps rather than hurts, let's talk.