Watching · AWS · Untrusted code, in Australia or next to private data

Lambda MicroVMs

A Firecracker VM per session for code nobody reviewed, inside your AWS account, in Sydney, with VPC access. It suspends with its memory intact and resumes on the next request.

Launched in June 2026 and in Sydney since August: a new Lambda primitive, separate from functions, that runs a Linux environment built from a Dockerfile in its own Firecracker VM. It's what you reach for when your product runs code nobody reviewed (an agent's code, a customer's script) and that code must stay in Australia or reach private data in a VPC.

I'm watching it rather than using it: I haven't run it in production, and for short bursts of code in a product on Workers, Cloudflare Sandbox is cheaper and ships a ready-made toolkit. MicroVMs win as soon as residency, the VPC, bigger machines or long sessions that keep their running processes matter. Both, priced on two real cases, are in Cloudflare Sandbox or Lambda MicroVMs?.

The part I like most is suspend and resume: a session idle for a while is snapshotted, memory included, and picks up exactly where it was on the next request, for up to 8 hours. The part I like least is that you build the server inside it yourself.

Use it when

  • The code must run in Australia: Sydney since 19 August 2026.
  • The code must reach private data in a VPC, such as the client's Postgres, through a network connector.
  • Long, stateful sessions: a coding agent's dev server, node_modules and memory survive the user leaving, and come back on the next request.
  • More than 4 vCPU or 12 GiB: up to 16 vCPU, 32 GB of memory and 32 GB of disk per MicroVM.
  • Running untrusted files or binaries with a full VM's isolation, inside the AWS account and its security tooling.

Don't use it when

  • Short bursts of code in a product on Workers: Cloudflare Sandbox bills active CPU only and costs about a quarter as much for that pattern.
  • The code is yours and reviewed: a Lambda or a Fargate task is simpler.
  • Sessions longer than 8 hours in total.

Pricing

Baseline compute billed per second while the MicroVM runs, from 1 vCPU and 2 GB (about $0.126 an hour at US East Arm rates: $0.0000276944 per vCPU-second and $0.0000036667 per GB-second). Resources above the baseline, up to 4×, billed only while used. Suspended: no compute, only snapshots ($0.0038 per GB written, $0.00155 per GB read, $0.08 per GB-month stored). Sydney rates not checked.

Gotchas, with fixes

Gotcha: There's no exec or code-interpreter SDK: Lambda returns an HTTPS endpoint to your own app inside the VM.

Fix: Bake a small server into the image (run a command, read and write files) and drive it over HTTPS with the short-lived token Lambda issues in the X-aws-proxy-auth header.

Gotcha: The baseline is billed every second the MicroVM runs, busy or idle.

Fix: Set an idle policy so it suspends quickly, and stop it when the session ends. For sessions of a few minutes, don't suspend at all: one suspend and resume of a 3 GB snapshot costs more than the session's compute.

Gotcha: Every MicroVM starts from the same snapshot, so anything generated at initialisation (random IDs, open connections, temporary credentials) is cloned into every copy.

Fix: Generate unique values and open connections after resume, using the hooks the service provides.

Gotcha: Arm64 only.

Fix: Build the image for arm64, and check that every native dependency has an arm64 build.

Provisioning

Official Pulumi package · Official Terraform provider

In @pulumi/aws (lambdamicrovms.Image and lambdamicrovms.Microvm, with egressNetworkConnectors for the VPC) and aws-native (lambda.MicrovmImage). The image is infrastructure; the MicroVMs themselves are usually started by the app at runtime (run-microvm), one per session.

Instead, consider

Decision guides

The concept behind it

Where it fits

  • AI that runs the code it writes: Pick 02.

    When the code must run in Australia, reach a private VPC, use more than 4 vCPU, or keep its processes through long sessions.

Sources

Last reviewed by Nicolas Dao

Want this stack in your AWS account?

Cloudless Labs is a Sydney consultancy. We migrate on-premises systems to AWS, build serverless apps on this stack, and coach teams to ship with Claude Code and Codex.

Talk to us

Contact Us

For general inquiries, please contact us at:

For e-learning capabilities, please contact: