This week’s AWS Summit NYC dropped a pile of engineering-relevant releases. If you were debugging a tent pole instead of watching the keynote, here’s what matters for your data pipelines and ML workloads.

**Graviton5 General Availability** The M9g and M9gd instances are now GA, powered by Graviton5 processors and the sixth-gen Nitro System. Expect up to 25% better compute performance over Graviton4, with 35% gains for web apps and ML inference, and 30% for databases. Graviton5 introduces PCIe Gen6 and DDR5-8800 memory, plus a 5x larger L3 cache. Network bandwidth is up 15%, EBS bandwidth up 20% across sizes. The M9gd adds up to 11.4 TB of NVMe SSD with 30% higher IOPS. Notably, the Nitro Isolation Engine now uses formal verification for mathematically proven VM isolation — the first formally verified cloud hypervisor. Both instance types support Instance Bandwidth Configuration, letting you shift up to 25% of bandwidth between EBS and VPC networking.

**FinOps Agent Now in Preview** The AWS FinOps Agent answers cost queries, surfaces optimization opportunities, and investigates anomalies. It can auto-generate Jira tickets from Cost Optimization Hub and Compute Optimizer recommendations, and post root-cause findings to Slack when anomalies trigger. For teams managing large-scale infrastructure costs, this reduces manual triage overhead significantly.

**Claude Fable 5: Launched, Then Revoked** Anthropic released Claude Fable 5 on Bedrock on June 9 with extended async execution and advanced vision. On June 12, Anthropic asked AWS to revoke access for all users to comply with a US export control directive. Opus 4.8 and other models remain unaffected. If you were depending on Fable 5, you’ll need to pivot.

**Gemma 4 on Bedrock** Google DeepMind’s Gemma 4 family is now available in three variants: a 31B dense model (256K context, reasoning/coding), a 26B-A4B MoE (cost-sensitive workloads), and the E2B (low-latency interactive). All support function calling, structured output, streaming, and multimodal input across text, image, video, and audio in 35+ languages.

**OpenSearch MCP Apps for Observability** OpenSearch Service now supports MCP Apps for agentic IDEs like Claude Desktop and VS Code. Agents can investigate incidents using logs, traces, and metrics, returning both a text summary and an interactive visualization in the same thread. Tools cover log/metrics/trace investigation, topology, cluster health, and instrumentation scoring.

**Other Engineering Notes** - AWS CLI v1 enters maintenance mode; botocore and s3transfer will be vendored directly. Migrate to v2. - The new Workload Credentials Provider issues short-term credentials without long-term keys, useful for on-prem or third-party workloads. - Kiro Pro Max launched with higher usage limits and access to frontier models.

Next week: AWS Summit NYC (June 17), Hong Kong (June 17), and Shanghai (June 23-24).

Source: AWS