AWS arrived in New York this week with a dense set of announcements, while Azure quietly published something platform engineers planning a storage migration will actually want to read. Cutting across both is a broader theme: the parts of the infrastructure stack that were previously manual, periodic, or reactive are being pushed toward continuous and autonomous. That is a shift worth taking seriously.
EC2 G7 Lands with NVIDIA Blackwell and 700 Gbps EFA
Amazon EC2 G7 instances are now generally available, making AWS the first major cloud provider to offer NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs in production. The instances come in seven sizes from g7.2xlarge (1 GPU, 8 vCPUs) to g7.48xlarge (8 GPUs, 192 vCPUs, 768 GiB RAM), with the metal variant listed as coming soon. The headline figures are 4.6x AI inference throughput and 2.1x graphics performance over G6, with GPU memory up to 32 GB per card and EFA-enabled networking at 700 Gbps on the largest SKU.
Why it matters: Teams running GPU-accelerated inference, VDI fleets, or video transcoding workloads on G6 now have a clear upgrade path with substantially more memory headroom and network throughput. The 700 Gbps EFA and GPUDirect RDMA support also opens up multi-node GPU workloads that would have been impractical at G6 bandwidth limits. Initial availability is US East (Ohio) and US West (Oregon) only.
ECS Auto Scaling Gets 20-Second Metrics
Amazon ECS service auto scaling now supports high-resolution 20-second CloudWatch metrics, replacing the default 60-second resolution for teams that enable the feature. AWS’s own benchmarking puts scale-out trigger time at 86 seconds versus the previous 363 seconds (a 4.2x improvement), with total time from trigger to provisioned tasks dropping from 386 seconds to 109 seconds. The feature is available across Fargate, ECS Managed Instances, and EC2 launch types and is enabled by selecting the new ECSServiceAverageCPUUtilizationHighResolution or ECSServiceAverageMemoryUtilizationHighResolution metrics in a target tracking policy.
Why it matters: The practical payoff is twofold: faster reaction to spikes means you can run leaner baseline task counts without taking latency hits during traffic surges, and the improved target tracking behaviour eliminates the need for more complex step-scaling configurations that many teams have built as workarounds. The feature itself has no cost; high-resolution CloudWatch metrics do carry an additional charge, so factor that into your CloudWatch spend.
AWS Transform Starts Filing Its Own Pull Requests
Launched at AWS Summit New York, AWS Transform – continuous modernisation (preview) adds ongoing autonomous tech debt analysis and remediation on top of Transform’s existing migration tooling. You connect your source control system, define policies (end-of-life dependencies, deprecated frameworks, internal coding standards), and the service scans your repositories against those baselines in hours rather than weeks. When findings exceed your threshold, it raises pull requests against the affected repositories automatically, notifying the owning team with the context needed to review and merge. Integration with AWS Security Agent means security vulnerabilities flow into the same prioritised findings list.
Why it matters: This targets a real gap: most organisations run periodic manual audits that produce reports nobody acts on before the next cycle starts. The pull-request-per-repo model shifts the friction from detection to decision, which is where it belongs. Platform teams enforcing internal library standards or Java version baselines across hundreds of repos will find the custom policy capability directly relevant. It is a preview, so treat the integrations and UI as subject to change before GA.
Azure Storage Migration Gets a Planning Layer
Microsoft has published updated guidance and tooling for storage migrations to Azure, centred on Azure Migrate, Azure Storage Mover, and Azure Data Box. The notable addition is the Azure Copilot Migration Agent, now in preview, which extends Azure Migrate with AI-assisted storage tool selection: it reads your existing Azure Migrate project data and recommends whether to use Storage Mover, Data Box, or another approach based on data volume, network capacity, and downtime tolerance. Azure Storage Mover itself remains free for the managed online transfer and synchronisation capability, with standard storage and networking costs applying. Azure Data Box 120 and 525 now carry no service or Microsoft-managed shipping fees.
Why it matters: The main value here is the planning layer, not the individual tools. Organisations approaching a datacentre exit or a bulk AWS S3 to Azure Blob migration typically reach tool selection before completing the dependency and sequencing work that determines which approach is actually appropriate. The Copilot Migration Agent is attempting to enforce that sequence rather than letting teams jump straight to execution. Whether the preview AI guidance is accurate enough to trust without verification is something to probe early.
Claude Design Tightens the Loop with Claude Code
Anthropic overhauled Claude Design this week, moving it out of isolated preview and into a shared token pool with Claude Code, chat, and Cowork. The headline capability is bidirectional Design-Code integration: designers can run /design-sync inside Claude Code to pull existing component libraries into Claude Design, and developers can use /design to create or edit design projects without leaving the terminal. Brand consistency is now a first-class feature, with design systems importable from GitHub repos or file uploads and automatically inherited by new projects. Connectors for Adobe, Canva, Figma, Lovable, Vercel, and others have been added.
Why it matters: The shared token pool addresses the main complaint from early users, who found that running Claude Design alongside Claude Code chewed through separate usage limits simultaneously. The /design terminal command is the more interesting addition for engineering teams: pulling a design spec into Claude Code context without context switching is a meaningful workflow improvement for frontend engineers reviewing designs before implementation. The caveat from practitioners quoted in The New Stack’s coverage is worth noting: token burn rates remain high, and letting Claude handle the entire design-to-code cycle autonomously is not yet reliably more efficient than a hybrid approach where humans lead and Claude assists.
A thread connecting several stories this week is the shift from tools that surface information to tools that act on it. AWS Transform raises pull requests; ECS auto scaling scales without intervention at 20-second intervals; Azure’s Copilot Migration Agent makes a tool recommendation rather than presenting a matrix. That is a meaningful change in what cloud tooling is expected to do, and it raises a question platform teams should be thinking about: where in this autonomous stack do humans need to remain in the loop, and what does the accountability model look like when a tool that generated a PR also triggered the pipeline that merged it?
