In March we were two people, an itch, and an empty industrial unit off Aldine Meadows Road. Closing out December, there's a fleet running customer workloads on silicon that was scheduled for a shredder. Not a bad first year for the circular-compute thesis.
What got built
The certification lab came first: benches, burn-in rigs, reference hardware, and the 72-hour gauntlet we've written about. Then the process learned from its first hundred cards. Then the fleet itself: racks of certified A100s and H100s, NVMe storage, 25Gbps networking, and monitoring that watches all of it around the clock.
What got proven
The two claims we needed to prove this year were technical and commercial: that rescued hardware certifies to reference performance, and that someone would trust their workloads to it. Both cleared. Our early customers, research groups and startups who did their diligence hard before signing, are running training and inference on second-life GPUs at a fraction of what the hyperscalers quoted them. The e-waste diverted along the way is tallied on our About page, and watching that number climb has been the most satisfying part of the year.
What 2026 looks like
Demand has outrun us: we ended the year fully subscribed, with a waitlist. The 2026 plan is straightforward: more acquisition partnerships, more lab throughput, more racks, onboarding the waitlist in order. If you want in the queue, tell us what you need.
Thanks to everyone who bet on rescued silicon this year. See you in 2026.