GPU clusters

GPU clusters will give you reserved capacity rather than shared serverless throughput — for training, batch work, or inference that must not queue behind anyone else.

How we expect it to work

Choose a GPU type and count.
Provision the cluster; it becomes yours until you release it.
Run workloads on it — serving catalog models, your own custom models, or training jobs.
Release it when you’re done.

What we expect to hold true

Billed per GPU-hour for as long as the cluster exists, regardless of utilisation.
No queueing against other tenants — the capacity is yours.
Availability depends on real inventory, so larger reservations may need lead time.

The economic crossover is roughly: if serverless spend approaches the cost of keeping GPUs resident, a cluster starts to make sense — and it also buys you predictable latency, which serverless cannot guarantee under load.