Training costs that track the job

Training is bursty. You warm up data, run for hours, checkpoint, then sit idle while someone debugs a dataloader. If the meter keeps ticking through that idle stretch, cost stops correlating with progress.

Meter while it runs

Chassis charges for GPU capacity while your instance is up. Stop or terminate when the job is done. That sounds obvious; the product work is making start/stop reliable from the console and the API so teams actually do it.

• Pick a SKU that matches the job — overbuying “just in case” is the quiet budget killer.
• Script teardown in CI and notebooks so failed runs do not leave machines warm.
• Watch the org wallet during long runs; top up before you are mid-epoch and out of funds.

Cost is a product surface

We treat pricing honesty as part of the product. Rates live in the catalog next to the hardware you select. There is no separate “surprise” tier for networking jargon you never opted into. If something looks wrong on a statement, support@okeymeta.com.ng will dig in with you.