The week of 20 July was about placement: where data is allowed to sit, where a long-running function keeps its state, and where a GPU node lands relative to its neighbours. AWS opened a Local Zone in Athens, the second in EMEA to carry S3 and EBS Local Snapshots, which makes it a residency answer rather than only a latency one. The Lambda Durable Execution SDK for .NET went GA, so C# teams get checkpointed workflows that can pause for up to a year without an external orchestrator. EKS Auto Mode and Karpenter gained EFA interfaces and EC2 placement groups in the node pool. And SES moved new accounts onto pricing plans and closed its own free tier to newcomers, a change that will surprise whoever creates the next sandbox account.

No AWS Transform, DMS, or other migration launch landed in this window. The .NET SDK is the closest thing to modernization news, and it is a real one for teams that rewrote Windows services into functions and then discovered they needed Step Functions to hold the state.

Athens: a Local Zone that can hold data

Most Local Zones are compute at the edge with the data back in the parent Region. Athens, identifier eu-central-1-ath-1a, ships with S3 One Zone-IA and EBS with Local Snapshots on top of EC2 C7i, M7i, and R7i, plus ECS, EKS, VPC, Direct Connect, and ALB. That combination is what lets you say to a Greek regulator that the data was stored and processed in Greece, which is the sentence the Local Zone is for. Pricing is pay-as-you-go with On-Demand, Savings Plans, and Spot, no commitment. You opt in through AWS Global View or the ModifyAvailabilityZoneGroup API:

aws ec2 modify-availability-zone-group \
  --region eu-central-1 \
  --group-name eu-central-1-ath-1 \
  --opt-in-status opted-in

The trade-off is the "Local" in Local Snapshots and the "One Zone" in One Zone-IA. Both mean the data lives in a single facility. A Local Zone has no sibling AZ to fail over to, and if the residency rule forbids replicating to Frankfurt, your durability story is one building. That can be acceptable for data that is also held by the system of record elsewhere, and unacceptable for data that is not. Decide which bucket each dataset belongs in before the architecture diagram gets drawn around the zone. Also note the service list: no RDS, no Lambda. Anything stateful that is not S3 or EBS runs in the parent Region, and the single-digit-millisecond latency figure is to Athens end users, not to eu-central-1.

Durable execution for .NET: the orchestrator moves into the function

The Amazon.Lambda.DurableExecution NuGet package gives C# functions steps that checkpoint automatically, callbacks for human or agent approval, durable invocation for chaining functions, and waits that suspend execution for up to a year while the function waits on an external event. A companion Amazon.Lambda.DurableExecution.Testing package emulates the runtime locally. The pitch is payment pipelines, agent orchestration, and approval flows without writing progress tracking or standing up a state machine.

For a team that ported a .NET Windows service to Lambda and then bolted on Step Functions to remember where a multi-step job was, this collapses two services into one and moves the control flow back into code where the compiler can see it. The failure mode is the same as every durable-function framework: your code must be deterministic between checkpoints, because on replay the SDK re-executes the function and expects the same sequence of steps. A DateTime.Now or a random ID outside a step breaks the replay in ways that are miserable to debug in production and trivial to catch with the local emulator. Use the emulator in CI from day one. And "pause for up to a year" is the wait limit; the announcement states no overall execution limit or pricing, so read the Lambda pricing page before modelling a workflow that lives for months.

EKS Auto Mode learns about placement groups and EFA

Node pools on EKS Auto Mode and open-source Karpenter can now request EC2 placement groups with cluster, spread, or partition strategies, and can configure EFA network interfaces as EFA-only or standard ENI on EFA-capable instances, for both dynamic and static capacity. EFA-only interfaces do not consume IP addresses. Available wherever EKS is.

This closes the gap that kept distributed training and tightly coupled HPC off Auto Mode. Those workloads need the nodes in a cluster placement group on the same spine and EFA for the collective operations, and until now that meant a self-managed node group with a launch template, which also meant you lost the node repair and consolidation that Auto Mode exists for. The thing to watch is capacity. A cluster placement group narrows where EC2 can place an instance, and Karpenter's strength is picking any instance type that fits; the two pull in opposite directions. Expect more InsufficientInstanceCapacity on provisioning, and give those node pools a static capacity floor rather than relying on scale-from-zero at 09:00 when the training job starts.

SES pricing plans: read this before your next new account

From 21 July SES offers Essentials, Pro, and Enterprise plans, bundled at up to 22% below buying the same capabilities individually. New SES accounts start on Essentials, as do returning customers who have not sent through SES since 1 June 2025. Anyone who has sent since then stays on à-la-carte pricing and can switch at will. The part that matters for platform teams: the SES-specific free tier of 3,000 messages a month for twelve months is gone for new customers, replaced by the general AWS Free Tier credits of up to $200 across six months.

Nothing here changes an established production account. It changes every new account: the per-environment sandbox, the new business unit's org member, the acquisition you are onboarding. If your account vending pipeline assumes SES is free at low volume, it is no longer, and the first signal will be a bill nobody expected. Update the baseline and the budget alerts before the next account is created.

Also this week

The throughline

Each headline this week is AWS letting you pin something that used to float. Athens pins data to a country. Durable execution pins a workflow's state to the function instead of a separate orchestrator. Placement groups pin nodes to a spine. Each pin buys you something concrete, residency, simplicity, bandwidth, and each costs you the flexibility that made the floating version resilient: a single facility, deterministic code, narrower capacity. None of these are bad trades. They are trades, and the announcements describe only the side you receive. The SES change is the same lesson from the billing direction: a default moved, and defaults are where the surprises live.

Read this next

For the AI and agent reading of the same weeks, the field notes live at ercan.ai. Consulting on AWS, cloud, CI/CD pipelines, Linux, GitHub/GitLab, and more runs through ercanermis.com.

References