Where all of this runs, reliably and economically.
Infrastructure answers the question that decides whether any of the rest is affordable: where does this execute, at what cost, and what happens when a piece of it fails?
Unit economics is an architecture decision.
A system that reads years of company records, re-reads them when something changes, and runs interpretation across all of it is a high-volume workload, not an occasional query. Whether that is economically sensible is settled by where inference runs, how it is batched, how much capability is packed into the available memory, and how much of the work never needed a model at all.
Owning inference capacity changes those economics directly. It also changes what can be promised about privacy and latency, because a workload served on controlled infrastructure does not have to send its data to an external inference provider.
Capacity, cost, and the assumption of failure.
- Capacity. GPU inference, memory-aware workload allocation and compute routing — deciding what runs where, given finite hardware.
- Throughput. Local serving, batch inference and inference optimization, which is where the difference between a demo and a production workload shows up.
- Persistence. Storage architecture, for evidence that has to remain readable and attributable years after it was ingested.
- Continuity. Containerized deployment, observability, backup and recovery, and the operational discipline that keeps a system running rather than merely built.
GPU inference, memory-aware allocation and compute routing decide what the hardware can hold and where a given job should go. Local model serving, batch inference and inference optimization determine what that hardware actually delivers per hour and per dollar. Storage architecture holds the evidence layer underneath it all. Containerized deployment, observability, backup and recovery, and production reliability are the difference between capability and an operating system somebody depends on.
A backup that has never been restored is an assumption rather than a recovery path, which is why recovery is treated as a tested behaviour here and in Evaluation & Reliability rather than as a checkbox.
Twelve capability pages sit in this category. Availability in the platform is not a claim that every one is used on every engagement.