Nvidia secures 45,000 Grace GPUs for Kazakhstan; AWS launches serverless SageMaker training

AWS made its “training sessions” feature in SageMaker AI generally available. The serverless offering lets users write custom training loops, reward calculations and other procedures for four supported models—gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B and Qwen3.8-27B—while SageMaker handles GPU allocation, distributed execution, failure recovery and checkpointing. Pricing is token‑based, billing pauses during reward scoring, and no capacity reservation is required. User data and custom weights are not reused to train other models.