Serve inference when traffic spikes
Add capacity for the hours you need it and stop paying when the spike passes. Nothing continues to bill after the container stops.
Traffic arrives in peaks. Capacity bought for the peak is idle for everything that is not the peak.
Why it is hard
The peak sets the bill
Provisioning for the worst hour of the week means paying for that hour's capacity during the other hundred and sixty seven.
Scaling down is slower than scaling up
Adding capacity when traffic arrives is the easy half. Most of the money is lost in how long it takes to give it back.
Idle still costs
A reserved instance serving nothing costs the same as one serving everything, which is what makes over-provisioning quietly expensive.
What answers it
Add capacity for the hours you need it
Launch against live listings when the spike starts. No reservation to negotiate and nothing to plan a quarter ahead.
Stopping is immediate
The meter ends with the container. Giving capacity back is the same action as stopping it, and it takes effect at once.
Idle costs nothing because it does not exist
There is no reserved pool sitting underneath. You hold hardware while you are using it and not a second longer.
How billing works
per second, to the microeuro · no minimum · no egress fee · idle costs nothing
Metering starts when the container does and stops when it stops. The rate is frozen the moment you match, so a provider re-pricing cannot change a running rental, and an aborted run costs what it used.
What you use from Kracht
Related jobs
- ComputeTrain a model without buying hardwareTake an H100 for the length of a training run and give it back. No reservation, no minimum, and no card sitting idle between experiments.Read the job
- ComputeFine-tune on a scheduleA job that runs nightly needs a GPU for the hour it runs, not for the day around it. Per-second metering makes the difference the whole business case.Read the job
- ComputeRun a lab with no capital budgetTurn a hardware purchase into a line item you can stop. Useful when the grant is annual and the experiment is not.Read the job