AWS in 2025

AWS Machine Learning in 2025: A Deep Dive into Concrete Updates

AWS’s machine learning landscape continued to evolve in late 2025, with releases focused on practical scalability, cost efficiency, and tighter integration across services. Below is a detailed breakdown of the most significant updates, grounded in facts from AWS re:Invent 2025 and subsequent announcements.

1. Nova 2 Foundation Model Family: Multi-Tiered Capabilities

AWS expanded its Nova series with four distinct models optimized for different workloads:

• Nova 2 Lite: A cost-effective reasoning model for everyday tasks like customer service chatbots, document processing, and business automation. It supports a 1-million-token context window and can be fine-tuned via supervised learning on Bedrock or SageMaker.
• Nova 2 Pro (Preview): Designed for complex, multi-step tasks such as multi-document analysis and software migrations. Early access is available to Nova Forge customers .
• Nova 2 Sonic: A speech-to-speech model enabling real-time, human-like conversational AI. It integrates with Amazon Connect and supports multilingual interactions with expressive voices.
• Nova 2 Omni: The first unified multimodal model that accepts text, images, video, and speech inputs while generating both text and images. It processes up to 750,000 tokens (e.g., entire product catalogs) in a single inference.

All models include three thinking intensity levels (low/medium/high) to balance speed, intelligence, and cost.

2. S3 Vectors: Native Vector Storage & Retrieval

AWS introduced S3 Vectors, a native vector storage solution that eliminates the need for external vector databases. Key features include:

• Semantic search, RAG pipelines, and recommendation systems via AWS-managed APIs.
• Scalability to 2 billion vectors per index (40× the preview limit) and support for 20 trillion vectors per bucket.
• Integration with Amazon Bedrock for structured querying and graph-style modeling.

This reduces complexity for applications requiring vector-based AI workloads.

re:Invent signage

3. Trainium3 UltraServers: Next-Gen AI Training Hardware

AWS launched Trainium3 UltraServers, built around its third-gen AI chips, offering:

• 3–4× performance improvement over prior generations with higher energy efficiency.
• Optimized for large-scale model training, supporting up to 144 Trainium3 chips in a single system.
• 3nm process technology for lower power consumption and reduced costs.

These servers target organizations training billion-parameter models at scale.

4. SageMaker Lakehouse: Unified Data & ML Workflows

The next-gen SageMaker Lakehouse now integrates seamlessly with:

• Amazon S3 Tables (built on Apache Iceberg) for unified access to data lakes, Redshift warehouses, and federated sources like DynamoDB and PostgreSQL.
• Zero-ETL connectors for near-real-time data ingestion from operational databases (e.g., Aurora MySQL, RDS).
• Lake Formation tag-based access control (LF-TBAC) for fine-grained permissions across analytics and ML tools.

The SageMaker Unified Studio IDE centralizes data discovery, processing, and model development, enabling cross-account governance and collaboration.

5. Enhanced Agentic AI Capabilities

AWS strengthened its Bedrock AgentCore with:

• Policy controls to define agent permissions (e.g., tool access, constraints) before any action is taken.
• Built-in evaluation tools monitoring agents for correctness, safety, and compliance.
• Nova Act, a service for building reliable AI agents with 90% success rates on browser-based automation workflows.

These features aim to deploy AI agents at scale with governance and quality assurance.

6. Nova Forge: Custom Model Development

Nova Forge allows enterprises to:

• Access pre-trained model checkpoints and blend proprietary data with AWS-curated datasets.
• Build custom “Novella” models tailored to domain-specific needs.
• Reduce iterative training costs via Trainium3 acceleration.

This addresses enterprise demands for customization without managing infrastructure.

Who Benefits?

• Developers: Nova 2 models and Bedrock’s expanded open-weight catalog (18 new models) simplify building multimodal apps.
• Data Teams: SageMaker Lakehouse unifies analytics and ML on a single copy of data, reducing silos.
• Enterprise IT: Policy controls and LF-TBAC ensure compliant, governed AI deployments.

These updates reflect AWS’s focus on practical, scalable tools rather than hype-driven “breakthroughs.” The integration of hardware (Trainium3), software (Nova 2), and governance (Lake Formation) positions AWS to cater to both foundational research and production-grade AI workloads.