Document-heavy claims processing

Context

An insurer runs document-heavy claims through AI — long inputs, moderate volume, but bursty (spikes after weather events). Self-host GPUs sit idle between surges.

The decision

Here utilization, not volume, decides. At 40% utilization the idle capacity keeps API ahead; drive utilization up (batch the backlog) and the cliff appears.

What most miss

Vendors quote self-host at 90% utilization. Bursty workloads run at 40% — the idle GPUs are the cost the pitch omits, and they move the cliff by years.

Stakes

Committing to self-host for a bursty workload can lock in idle-capacity cost that dwarfs the API bill it replaced.