SageMaker AI adds SOCI to reduce container start-up delays
SageMaker AI supports SOCI-indexed container images for training and processing jobs. AWS places the improvement in the initial download, not in the overall cost of training a model.
News summary
On 29 October, AWS announced support for images indexed using Seekable OCI, or SOCI, in SageMaker AI training and processing jobs. Lazy loading retrieves the files needed to start first and streams the rest in the background. AWS says it can reduce GPU idle time during image downloads by up to 90%; this is not a 90% reduction in total training costs.
The capability has been announced for regions where SageMaker AI is available and allows indexed images in private Amazon ECR repositories to be used. For teams running frequent jobs with large images, start-up can account for a significant share of the waiting time. The actual benefit depends on the files used, the container size and the duration of each job.
Source: AWS — 29 October 2026
What to review in your business
Measure image download, start-up and model execution separately. An optimisation that reduces the initial wait may have little impact on a job that runs for many hours.
Validate the behaviour of the indexed image and its dependencies before rolling out the change more widely. Maintain controls over container provenance and updates.
At Seintec, we help you review your cloud infrastructure and its dependencies to make informed plans for upcoming changes. Contact our team.
Contact SeintecRelated service
Cloud Services
Cloud services to scale your business.