The situation: You're running production inference on GCP, and you need Hopper-generation GPUs, the A3 instances powered by H100s. 

The problem: On-demand capacity for these is scarce. Google wants a committed use discount (CUD) before it'll guarantee you the silicon. So you're stuck choosing between "commit and get the GPUs" or "stay flexible and hope there's capacity when you need it."

Why this is trickier than a normal commitment: Locking in compute for a year is one thing. Locking in this specific generation of GPU for a year is another. AI hardware moves fast. Blackwell is already rolling out, and whatever comes after that is already on a roadmap somewhere. Commit today, and there's a real chance you're paying for last-gen chips while your workload could be running faster (or cheaper) on something newer.

The trade-off, spelled out:

• Commit now → you get the A3 capacity you need for production inference, today.
• Don't commit → you risk not getting capacity at all, but you keep your options open for next-gen hardware.

Neither option is great on its own. That's the whole tension FinOps and infra teams are wrestling with right now.

Where Archera comes in: A Guaranteed Committed Use Discount (GCUD) is Archera's version of a GCP CUD; same discount, same capacity lock, but backed by a Rebate Guarantee. You still make the commitment Google requires to secure your A3 capacity. But if the commitment stops paying off (say, next-gen instances launch and you'd rather move your workloads over), Archera's Rebate Guarantee kicks in to cash you out of the unfavorable position instead of leaving you stuck holding the bag on outdated hardware.

In plain terms: You get the capacity guarantee you need now, without betting the next 12 months on today's GPU generation staying the best option.

Why "rebate" and not "release": Archera's guarantee shows up differently depending on the cloud. On GCP, it's structured as a Rebate Guarantee. The commitment stays in your account, and after the term, Archera sends a cash rebate if it turns out you overcommitted. That's different from how it works on other clouds, so don't mix up the mechanics. On GCP, GCUDs are a Guaranteed Commitment backed specifically by the Rebate Guarantee.

Bottom line: If your team needs A3/Hopper capacity for production inference, you likely can't avoid committing. A GCUD lets you make that commitment without the fear of being locked into hardware that's obsolete before the term is up. Secure the capacity, keep the hedge.

‍

Get started with Archera on Google Cloud

‍