Amazon SageMaker AI now supports NVIDIA Multi-Instance GPU partitioning, so one GPU can host a different model in each isolated partition. It covers P4d (A100), P5 (H100), G7 and G7e instances, with up to 7 partitions on A100 and H100 GPUs. It only works with endpoints that use inference components, not single-model or serverless endpoints, and the post reports roughly 17 seconds of extra provisioning time. Teams running several small models on large GPUs may get better utilisation.
