AWS re:Invent 2025 Preview: 5 Key Updates Coming to Bedrock and SageMaker AI

Half a million Trainium2 chips, wired into a single supercomputer. Project Rainier went live in early November, and if that doesn’t tell you what AWS re:Invent 2025 is going to look like, nothing will. The annual cloud conference kicks off December 1 in Las Vegas, and the signals coming out of AWS over the past two months point to some of the most significant Amazon Bedrock and SageMaker updates we’ve seen yet. Here’s what developers and AI practitioners should be watching for.

AWS re:Invent 2025 — The Year of AI Agents

Running December 1 through 5 in Las Vegas, re:Invent 2025 features keynotes from AWS CEO Matt Garman, Peter DeSantis, Swami Sivasubramanian, and CTO Werner Vogels. Over 1,000 technical sessions are planned, and for the second consecutive year, generative AI dominates the agenda.

But this year’s narrative is expected to shift beyond chatbots and AI assistants toward something more ambitious: autonomous AI agents that can work independently for extended periods, make decisions, and take actions on behalf of organizations. The October-November updates to Bedrock and SageMaker all point in this direction.

Amazon Bedrock — Model Ecosystem Expansion and AgentCore

Amazon Bedrock has been on an aggressive update cadence. In October, Stability AI Image Services added four new editing tools — Outpaint, Fast Upscale, Conservative Upscale, and Creative Upscale. In early November, Amazon Nova Web Grounding launched, giving models the ability to perform real-time, citation-based web retrieval. TwelveLabs’ Marengo Embed 3.0 also arrived for video-native multimodal embeddings.

The big re:Invent announcement is expected to be Reinforcement Fine-Tuning (RFT) for Bedrock. Unlike traditional fine-tuning that requires massive labeled datasets, RFT uses feedback-driven training to improve model accuracy — reportedly delivering up to 66% accuracy gains over base models without deep ML expertise. For teams without dedicated ML engineers, this could be transformative.

Bedrock AgentCore is also expected to receive a major expansion. Based on pre-event signals, the additions will likely include policy controls for governing agent behavior, quality evaluations for monitoring agent output, episodic memory enabling agents to learn from past interactions, and bidirectional streaming for natural conversational experiences. This is AWS’s answer to the enterprise question: “How do we deploy AI agents at scale without losing control?”

SageMaker AI — HyperPod’s Checkpointless Training Changes Everything

On the SageMaker side, the anticipated star is HyperPod. Anyone who’s trained large models on clusters with thousands of accelerators knows the pain: when hardware fails (and it will), the training job rolls back to the last checkpoint. Recovery takes hours. Cluster efficiency drops to 60-70%.

The expected announcement of checkpointless training eliminates this entirely. Using peer-to-peer state recovery, training can resume from the exact point of failure in minutes rather than hours. Early reports suggest cluster efficiency could reach 95% — a massive improvement for organizations running multi-million dollar training jobs.

Elastic training is the other feature to watch. When idle accelerators become available, training jobs automatically scale up. When higher-priority workloads need those resources, training scales down without stopping. The training continues with fewer resources rather than crashing. For cost-conscious teams managing shared GPU clusters, this alone could justify the SageMaker investment.

Trainium3 and Project Rainier — AWS’s Infrastructure Ambition

The hardware story is equally compelling. Project Rainier — built in collaboration with Anthropic — went operational in early November with nearly 500,000 Trainium2 chips, making it one of the world’s most powerful AI supercomputers currently in production.

But AWS isn’t stopping there. Trainium3, built on a 3-nanometer process, is widely expected to be announced at re:Invent. The next-generation chip is projected to deliver over 4x the compute performance and energy efficiency of Trainium2, with up to 144 chips packable into a single UltraServer. Training jobs that currently take months could potentially shrink to weeks.

There are also strong signals around AWS AI Factories — a service for deploying AI infrastructure directly in customers’ existing data centers — and deeper NVIDIA collaboration that could yield new P6e-GB300 instances with the latest Blackwell architecture.

5 Things Developers Should Watch For

• Bedrock Reinforcement Fine-Tuning — Model customization without ML expertise. Potentially the most impactful feature for small teams building on foundation models
• AgentCore Enterprise Features — Policy controls, quality monitoring, and episodic memory for deploying AI agents safely at scale
• SageMaker HyperPod Checkpointless Training — Slash failure recovery from hours to minutes. Cluster efficiency approaching 95%
• Trainium3 UltraServers — 4x performance gains on AWS’s next-gen AI silicon, with 144-chip configurations
• Amazon Nova Model Family Expansion — Next-gen foundation models with expanded modalities and reasoning capabilities

Strategic Timing: Why Black Friday Week Matters

AWS didn’t choose the first week of December by accident. By launching new services right after Black Friday and before year-end budget deadlines, AWS ensures enterprise customers factor these capabilities into their 2026 cloud AI investment plans.

The timing is especially critical this year. Google Cloud’s Gemini and TPUs, Microsoft Azure’s AI stack and Copilot ecosystem — the competition for enterprise AI workloads has never been fiercer. AWS is expected to counter with Bedrock’s unmatched model diversity (approaching 100 serverless models) and its custom silicon advantage through Trainium and Graviton.

December 1-5 in Las Vegas will be more than a tech conference — it could set the direction for enterprise AI strategy through 2026. Whether Bedrock’s reinforcement fine-tuning delivers on its promise, whether SageMaker HyperPod truly eliminates checkpoint pain, and what vision Matt Garman lays out for the future of cloud AI — all of it unfolds in just under two weeks. This is one re:Invent you don’t want to miss.