Running a CI/CD pipeline that fires a hundred short jobs a day? Or an AI coding agent that needs a clean, isolated environment for each task? You’ve probably been paying for servers that sit idle between those bursts — because traditional cloud VMs don’t stop charging just because they’re not doing anything.
DigitalOcean just moved MicroVMs into public preview, and the core idea is simple. A MicroVM is a tiny, fully isolated virtual machine that starts in milliseconds, runs your workload, then pauses automatically when idle. Paused instances pay only for storage. You get charged by the second for compute only when the VM is actually running.
MicroVMs are built on Firecracker, an open-source virtualization technology originally developed by AWS for their Lambda and Fargate services. Firecracker is a virtual machine monitor, which means software that creates an extremely lightweight VM with its own operating system kernel and strong hardware-level isolation, but without the overhead of a full server setup. That isolation is what makes MicroVMs safe for running untrusted code — like user-submitted scripts or AI agent actions that shouldn’t touch your core infrastructure. Pricing is $0.022 per vCPU-hour and $0.0085 per GiB of memory per hour, billed to the second.
The three natural fits are AI coding agents, sandboxed code execution, and CI/CD runners. Each of those workloads is bursty by nature. An AI agent writes some code, needs an environment to test it, runs the test, then waits. A CI pipeline triggers on a pull request, runs for three minutes, then goes quiet for hours. With traditional compute, you’re either pre-warming expensive VMs or waiting for cold starts that ruin your developer experience. MicroVMs eliminate that trade-off.
For teams managing AI agent infrastructure, this is a practical cost-control tool. Agents that spin up dozens of isolated environments per day could see costs drop significantly compared to running persistent containers or standard cloud instances. The per-second billing removes the incentive to keep VMs running continuously just to avoid startup delays. You pay for work done, not time reserved.
The infrastructure world is quietly reshaping itself around the assumption that AI workloads are short, bursty, and parallel. Long-running servers are the wrong abstraction for a coding agent that needs 15 seconds of compute, a clean filesystem, and then nothing. MicroVMs fit that pattern precisely. This one is worth paying attention to as AI workloads become a bigger part of your infrastructure bill.
Want to explore how AI agent infrastructure could benefit your business? Let’s talk.