Home / Technology / AWS Trains AI Smarter, Not Harder
AWS Trains AI Smarter, Not Harder
23 Oct
Summary
- AWS SageMaker offers new AI training for flexible model customization.
- Users control training logic; AWS manages expensive GPU infrastructure.
- New token-based billing model impacts cost calculations.
Amazon Web Services (AWS) has launched general availability for new training sessions within its SageMaker AI platform, enhancing the customization of machine learning models. This update allows developers to control training procedures and success metrics, while AWS handles the underlying GPU infrastructure, reducing operational complexity and costs for enterprises. The service enables teams to experiment with custom feedback mechanisms without building their own hardware.
Previously, SageMaker offered fine-tuning and serverless customization. The new training sessions combine granular programming control with a serverless approach, where developers direct individual operations via API calls, and AWS executes them on GPUs. This method supports agents requiring multi-step actions, like internal support bots. AWS also supports Low-Rank Adaptation (LoRA) and full fine-tuning within this offering.
Early adopters like ImagineSoftware and Karini AI have utilized the service for specialized tasks, citing benefits in performance control and reduced setup work. AWS has specified supported models such as gpt-oss-20b and gemma-4-31B-it, available in regions including Northern Virginia, Oregon, and Mumbai.
Billing for training sessions is now based on processed tokens, with the meter pausing during customer-defined scoring. This approach differs from reserving GPU capacity, though other associated costs for computing, storage, and inference still apply. AWS is entering a competitive market with this offering, emphasizing integration with existing AWS enterprise infrastructure and security controls.
