AWS Launches Graviton5, Blackwell GPUs, and Bedrock AgentCore at New York Summit

AWS used its New York Summit on June 17–18 to ship the broadest AI infrastructure update in its history. The headline hardware, the GA agent orchestration platform, and a new S3 capability are all aimed at the same goal: making AWS the default place to run production AI agents.

Graviton5 and EC2 M9g. The new M9g instances on Graviton5 deliver 25% faster compute performance compared to the prior generation, with a 5× larger L3 cache and DDR5-8800 memory support. Graviton chips are the backbone compute that makes everything else cheaper to operate — not the flashy GPU instances, but the ones processing millions of inference calls for price-sensitive workloads.

EC2 G7 with NVIDIA RTX PRO 4500 Blackwell. The GPU story is architectural, not incremental. G7 instances deliver 4.6× AI inference performance improvement over the G6 generation and 2.1× better graphics. Blackwell is a generational shift in throughput, and AWS landing it in G7 instances makes it production-accessible at cloud pricing rather than through custom cluster procurement.

Amazon Bedrock AgentCore GA. This is the most consequential announcement for developers building real AI systems. AgentCore coordinates autonomous agents, manages their execution state, and integrates with AWS access controls — without requiring teams to build their own orchestration infrastructure. It pairs with Amazon Quick, an autonomous agent layer that runs continuously across enterprise apps (Salesforce, Slack, Snowflake, email) even when users are offline. Agents describe what they need in natural language; Quick handles execution with configurable autonomy levels from step-by-step approval to broad goal-based completion.

Amazon Bedrock Managed Knowledge Base. Fully managed RAG pipeline that handles ingestion, chunking, and retrieval from S3, SharePoint, Confluence, and Google Drive — without managing vector stores manually. It removes the plumbing work that has been the primary source of friction in enterprise RAG deployments.

S3 Annotations. Developers can now attach up to 1 GB of queryable context — JSON, XML, or YAML — directly to S3 objects, discoverable by AI agents via a natural-language S3 Tables MCP server. Purpose-built for agentic workflows where agents need to understand enterprise data without maintaining external indexing systems.

AWS Continuum. A security service that continuously discovers, prioritizes, validates, and remediates code vulnerabilities using AI agents, with a supervised “learn mode” that requires explicit permission expansion before autonomous action.

The throughline across all of it: AWS is not building tools for AI demos. They are building the infrastructure stack that enterprises need to run agents that have real access controls, real audit trails, and real execution capability. That positioning is explicit, and the summit was the clearest statement of it yet.

Sources

Related reading