Release & Changelog Watcher — August 7, 2026

What Happened

• Amazon ECS now supports fractional GPU scheduling on EC2 G6f instances, permitting GPU partitions as small as 1/8 of an NVIDIA L4 (≈3 GB VRAM) by setting GPU=0.125, 0.25 or 0.5 in an ECS task definition; supported on ECS Managed Instances and ECS on EC2 [1].
• AWS Lambda console added console-to-IDE integration for Kiro and Cursor (in addition to VS Code) with guided setup and conversion to AWS SAM templates for local development and IaC [2].
• Amazon EC2 G7 instances (NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs) are available in four Regions, offering up to 8× GPUs, 32 GB/GPU, and large memory/network configurations for accelerated inference and graphics workloads [3].
• Amazon Bedrock AgentCore added “runtime instances” (GA) to run agents on customer EC2 capacity; supports capacity providers, mixed instance types and long-running agent sessions up to 14 days (AgentCore management fee + EC2 cost) [4].
• Amazon ElastiCache added Graviton4-based node families M8g, R8g and C8gn for Redis and Memcached with up to ~47% higher throughput and improved P99 latency vs Graviton3 equivalents; sizes up to 16xlarge and wide Region coverage [5].
• AWS Marketplace sellers can configure net payment terms (Net 30/45/60/90) on private offers; Channel Partner Private Offers (CPPO) allow ISV-set ceilings for partners [6].
• Amazon MSK now delivers Authorizer Logs for Provisioned clusters (Standard and Express) to CloudWatch, S3 or Data Firehose, logging denied authorization attempts with client IP and API attempted at no extra cost (region exceptions apply) [7].
• AWS Backup for S3 supports creating S3 Access Points that expose read-only access to backup/point-in-time recovery points via standard S3 APIs without restoring, and protects recovery points from deletion while an access point is active [8].

Why It Matters to Businesses

• Lower unit cost for model inference and experimentation: Fractional GPUs on G6f reduce the need to provision whole GPUs for small models, lowering per-inference cost and improving density for inference fleets [1].
• Faster developer iteration and IaC adoption: Lambda IDE integrations with Kiro and Cursor streamline local dev and automated template conversion to SAM, speeding serverless CI/CD onboarding [2].
• Performance headroom for production ML and graphics: G7 instances provide higher throughput and GPU memory for demanding inference, streaming and spatial/graphics workloads needing multi‑GPU scale-up [3].
• Operational flexibility for agents: AgentCore runtime instances let you run long-lived or stateful agents on your EC2 fleet while offloading lifecycle management to AgentCore, useful for session-heavy or GPU-bound agents [4].
• Better cache price/performance: Graviton4 ElastiCache nodes deliver measurable throughput and latency improvements that lower cost-per-query for caching layers in front of ML or web services [5].
• Procurement and cashflow control: Configurable net terms in Marketplace private offers improve vendor negotiation and buyer cashflow planning, important for ISV contracts or channel deals [6].
• Improved security and auditability: MSK authorizer logs provide forensic-level detail for failed authorizations; S3 access points for backups enable instant read access to recovery points without restore operations, reducing RTO for investigations [7][8].

Kimbodo Engineering Perspective

We view these updates as practical levers to reduce cost, increase operational density and speed developer workflows — but each introduces trade-offs that need explicit engineering controls.

• Fractional GPUs vs full GPUs: Fractional scheduling is ideal for many small-model inference workloads and dev/test, but expect variability in achievable throughput and potential contention. For large models, multi-GPU training, or tightly-coupled GPU workloads, full GPUs (or G7 multi‑GPU instances) remain necessary [1][3].
• Managed vs DIY agent runtimes: AgentCore runtime instances offload provisioning/patching but add AgentCore management fees and EC2 billing; they suit teams that need long sessions or GPU-backed agents and want to avoid writing agent lifecycle tooling. For ephemeral, latency-sensitive agents prefer microVM serverless runtime [4].
• Graviton4 adoption: Upgrading caching layers to M8g/R8g/C8gn yields measurable latency/throughput gains and cost savings, but you must validate application-specific performance characteristics (JVM flags, memory tuning, network behaviour) before wholesale migration [5].
• Operational observability and cost: Authorizer logs improve security posture but increase log volume and storage needs. Plan retention, S3 lifecycle and SIEM integration to control costs and enable fast investigation [7].
• Backup access points: Instant read-only access removes restore friction for audits and analytics, but increases the attack surface; enforce strict IAM, endpoint controls and monitoring when enabling access points [8].

How We Would Implement It

Fractional GPUs on ECS (G6f)

• Define ECS task definitions with GPU=0.125/0.25/0.5 for small-model containers and attach a G6f capacity provider. Use ECS Managed Instances to get CloudWatch Container Insights GPU metrics and automatic GPU health monitoring [1].
• Autoscale instance groups by GPU partition utilization, and run spot-backed dev fleets for non-critical workloads to reduce cost. Validate throughput with representative p99 latency tests per partition size.

G7 Instances for Production Inference

• Use G7 instances for multi-GPU inference or large-model deployments. Choose instance sizes based on GPU count, GPU memory (32 GB/GPU) and network needs; attach EFA for distributed inference frameworks where required [3].
• Benchmark model latency and throughput vs G6/G6f and use Spot for stateless bursty jobs only.

AgentCore Runtime Instances

• Create an AgentCore capacity provider specifying EC2 families (GPU, memory-optimized) and attach agents; mix instance types for cost/perf balance. Configure allowed session length (up to 14 days) and automated patch/replace policies [4].
• Instrument with host and agent-level telemetry, and centralize logs/metrics in your observability stack. Tag resources for cost attribution (AgentCore fee + EC2).

ElastiCache Graviton4 Migration

• Run canary clusters on M8g/R8g/C8gn, measure throughput and p99 latency with production traffic patterns. Use c8gn for network‑heavy workloads (up to 200 Gbps) and choose sizes up to 16xlarge as needed [5].
• Update parameter groups, test memory/eviction behavior, then use in-place scaling or blue/green cluster swap for cutovers.

MSK Authorizer Logs and Backup Access Points

• Enable MSK Authorizer Log Delivery to CloudWatch/S3/Firehose for Provisioned clusters; centralize into SIEM and set retention/ingestion limits to control cost [7].
• When creating AWS Backup S3 Access Points for recovery points, use access-point policies limiting allowed principals and require VPC-only access where possible; protect recovery points while access point is active and set lifecycle rules for access logs [8].

Lambda IDE Integration and Marketplace Terms

• Adopt Kiro/Cursor option in the Lambda console for teams preferring those IDEs; use automatic SAM conversion to integrate into existing IaC pipelines and CI/CD [2].
• For commercial deals, negotiate private offers with explicit net terms in AWS Marketplace and configure CPPO ceilings for channel partners to align cashflow with revenue recognition [6].

Risks, Costs and Security

• Performance variability and fragmentation: Fractional GPUs increase packing efficiency but can cause noisy-neighbor and scheduling fragmentation. Mitigation: benchmark per-partition performance, reserve full GPUs for latency-critical or large-model workloads, and monitor GPU metrics [1].
• Cost composition: AgentCore runtime adds management fees plus EC2 costs; track AgentCore vs microVM cost per session and prefer serverless runtime for short-lived, high-concurrency use cases [4].
• Logging and storage costs: MSK authorizer logs and S3 access-point reads increase storage and egress; apply retention policies, compression, and targeted ingestion to SIEMs [7].
• Backup access surface: S3 Access Points expose backup data via APIs. Enforce least-privilege IAM, use VPC endpoints and MFA or multi-party approval for cross-account access to minimize exposure [8].
• Compliance and data residency: Region availability varies (e.g., G7 and some features only in specific Regions; MSK authorizer logs excluded from EU sovereign region). Confirm regional availability for regulated workloads before design [3][7].
• Procurement and contractual risk: Using Marketplace net terms changes cashflow and accounting; ensure legal and finance alignment on payment terms and CPPO ceilings [6].
• Security hygiene: Apply encryption at rest/in transit, centralize key management, enable access logging, and enforce IaC guardrails (Terraform/CloudFormation) for any new resource types and capacity providers mentioned above [1][4][5][7][8].

Sources

1. [1] Amazon ECS now supports fractional GPU scheduling with Amazon EC2 G6f instances
2. [2] AWS Lambda console extends console-to-IDE integration to Kiro and Cursor
3. [3] Amazon EC2 G7 instances are now available in the AWS Europe (Spain) Region
4. [4] AgentCore runtime instances are now generally available
5. [5] Amazon ElastiCache now supports Graviton4-based M8g, R8g, and C8gn nodes
6. [6] AWS Marketplace now lets sellers configure net payment terms on private offers
7. [7] Amazon MSK now delivers Kafka Authorizer Logs to customers
8. [8] AWS Backup for Amazon S3 now supports direct access to backup data