On 16 October 2026 AWS Transform made Azure-to-AWS code conversion generally available: the .NET modernization agent now finds Azure SDK calls in your source and rewrites them as AWS SDK calls, which moves the Azure migration conversation from "move the VMs" to "change the code". The same week Lambda gained a second form factor, Lambda Web Functions, billed on active CPU rather than wall-clock time; ECS shipped a task-count metric that counts draining tasks; and ARC Region switch learned to check standby-Region quotas before you need them.

Transform converts Azure SDK code, not just Azure VMs

The 16 October announcement is specific about scope. Transform discovers the Azure SDK in submitted .NET source, presents conversion options in the assessment report, lets you accept or adjust them, then converts the calls and validates the result. Generally available: Azure Functions to AWS Lambda, Azure Blob Storage to Amazon S3, Azure Queue Storage to Amazon SQS. Azure Cosmos DB to Amazon DynamoDB is in preview. AWS Transform runs in eight Regions: N. Virginia, Mumbai, Tokyo, Seoul, Sydney, Canada Central, Frankfurt and London.

Why this matters more than another VMware story: until now the agentic Azure path was infrastructure. You could replicate Azure VMs, but a Functions app talking to Blob Storage was still a rewrite someone had to schedule. The GA pairs are the three services that make up most of a cloud-native .NET back end: a trigger, an object store, a queue. The preview pair is the honest one. Cosmos DB to DynamoDB is a data-model conversion, not an SDK swap; partition keys, consistency levels and query shapes do not translate by regex, so expect generated DynamoDB code to compile before it is correct. Run the assessment even if you are not moving this quarter: how much of the estate sits in the GA pairs versus Cosmos, Service Bus or Key Vault is the number every migration business case has been guessing at.

Three more Transform items: Transform for migrations reached Asia Pacific (Singapore) (15 October). Full-stack Windows modernization now deploys converted Aurora PostgreSQL clusters to a different Region and is available in seven more (15 October): upload SQL Server DDL, set a target Region, and the dependent .NET code is made PostgreSQL-compatible in the same run. On the mainframe side, a joint post with Cognition (12 October) has Transform recovering business rules from COBOL, PL/I and JCL with source-line traceability, and Devin turning the approved requirements into specs, backlog, code and tests, each as a pull request a human merges. No percent-automation figure, no time-saved claim. That restraint is the most credible thing about it.

Lambda Web Functions: a port, not a handler

On 12 October AWS announced Lambda Web Functions. Your code listens on the host and port in AWS_LAMBDA_HTTP_ENDPOINT and handles HTTP like any server; Express, Hono, Fastify and Next.js run without adapters, and existing functions are renamed Lambda Event Functions. One warm environment serves up to 128 concurrent requests (default 64, via maxConcurrencyPerEnvironment); requests above the limit get a 429, not a queue. Each environment has 2 vCPUs and 2 GB, the runtime is Node.js 24 on Arm64, and it is live in N. Virginia, Ohio, Oregon, Tokyo and Ireland. aws lambda-web deploy --name app --code ./dist --regions us-east-1,eu-west-1 returns the URL. The same day CloudFront added Web Functions as an origin with Origin Access Control, which is how you get a custom domain and WAF in front of one.

The billing model is the real announcement. The Lambda pricing page lists Web Functions in N. Virginia on Arm at $0.0000276944 per active vCPU-second, $0.0000036667 per GB-second of memory held, and $0.30 per million requests, against $0.20 per million for Event Functions and $0.0000166667 per GB-second at the x86 Event Function rate. Take the two extremes. A 2 vCPU, 2 GB environment at full CPU costs about $0.000063 per second; a 2 GB Event Function about $0.000033. An environment mostly waiting on a database costs about $0.0000073 per second for memory and almost nothing for CPU. The model wins when the CPU is idle, the normal web tier, and loses when it is busy, image processing and heavy SSR. Benchmark your p95 CPU share before moving a workload; the pricing page shows no Web Functions free tier.

Failure mode worth stating: 128 concurrent requests in one Node.js process share one event loop, so one slow synchronous path degrades 127 neighbours. And since excess requests get a 429, your client retry policy is part of the architecture.

ECS finally counts the tasks that are still running

On 13 October Amazon ECS added LiveTaskCount and removed the charge for RunningTaskCount. LiveTaskCount reports tasks across ACTIVATING, RUNNING and DEACTIVATING, including tasks still serving traffic while they drain during scale-in. That clause is the bug it fixes: the old count under-reported during scale-in, so a service draining 20 tasks looked smaller than it was and the policy could over-correct. If you scale on task count, switch the metric and re-tune the target. RunningTaskCount in Container Insights is now free.

Region switch checks the quotas before the failover

On 14 October ARC Region switch added four capabilities: Systems Manager Run Command and Lambda-payload execution blocks, service quota parity checks, and a wait for target group health on compute scaling actions. The quota check is the one to adopt this week. Failover plans fail in the dullest way imaginable: the standby Region has a 32-vCPU default limit where production has 2,000, discovered at scale-out time during the incident. Quota parity evaluated in advance turns that into a ticket on a quiet Tuesday. The target-group wait closes the other classic gap, where the runbook declares the Region up when the ASG reaches desired capacity while the ALB is still failing health checks.

Also this week

The throughline

Every headline this week replaces something you used to estimate with something you can measure. Transform's assessment says how much of an Azure codebase is convertible instead of letting a steering committee guess. Web Functions bill the CPU you burn, and the price list gives you the break-even before you move. LiveTaskCount counts the tasks really serving. Region switch checks standby quotas before the day they matter. Less exciting than a keynote, more useful on a Tuesday.

Read this next

For the model and agent reading of the same week, the AI field notes live at ercan.ai. The hub, including consulting on AWS, cloud, CI/CD pipelines, Linux, GitHub/GitLab and more, is at ercanermis.com.

References