# Amazon SageMaker Inference Component

> Inference components carry no meter of their own; the endpoint's instances bill per instance-hour, and components divide that paid capacity among models with individual compute allocations. ZopNight discovers components on the 6-hour cycle, attributes cost to the parent endpoint's spend, and recommends packing improvements where allocations waste instance capacity.

Source: https://zop.dev/integrations/aws/sagemaker-inference-component
Updated: 2026-08-19

---

An inference component packs a model with its own compute allocation onto a shared SageMaker endpoint, letting multiple models share instances. Component-level visibility shows which models actually consume the endpoint capacity you pay for.

## Slicing an instance-hour among models

Inference components change the accounting inside an endpoint without changing its meter. The endpoint's instances still bill per instance-hour; components carve that capacity into per-model allocations of accelerators, CPU, and memory, each with its own copy count that can scale independently, including to zero. Done well, this is the cost-efficient way to serve many models: a dozen small models sharing GPU instances that would otherwise each demand their own endpoint. Done carelessly, it reproduces the idle-endpoint problem one layer down, with allocations reserving slices of paid hardware for models nobody invokes.

## Packing efficiency as a finding

ZopNight discovers inference components automatically on the 6-hour cycle and attributes their cost through the parent endpoint's instance spend. The packing and utilization recommendations look at the division itself: components whose allocations sum to a fraction of the fleet's capacity (stranded headroom that might permit fewer instances), components whose invocation counts are zero while their allocation holds accelerator share, and copy counts pinned above what traffic justifies. The unit of action is sometimes the component (shrink, scale to zero, delete) and sometimes the endpoint (fewer or smaller instances once the packing is honest).

## Multi-model serving gone slack

The waste patterns mirror multi-tenant servers everywhere: every team overstates its allocation to be safe, so the endpoint scales out on reservations rather than load; deprecated model versions keep their components (and their slices) after traffic moved to the successor; and a component scaled up for a launch never scales back down, holding accelerator share as a souvenir.

## Component lists and their allocations

The SageMaker console shows inference components per endpoint with their allocations and copy counts, alongside per-component invocation metrics. Allocation against invocations, component by component, is the packing audit. The stranded difference is measured in the instance-hours the endpoint's fleet bills regardless.