Ditch the Bare Metal - Android Emulation on EC2 with Nested Virtualization

If you run Android CI pipelines at scale, you know the pain: bare metal instances are the only way to get the KVM acceleration your emulators need, and that comes with a real cost premium. Bare metal instances serve one customer at a time. You pay for the full physical host whether you run one emulator or twenty. That changed in February 2026 when AWS announced nested virtualization support on C8i, M8i, and R8i instances . You can now run KVM-accelerated Android emulators on standard virtualized EC2 instances for the first time.

This post covers what KVM is, why emulators need it, why bare metal was previously required, and how nested virtualization changes that. I also walk through getting Android emulation working on a virtualized EC2 instance from scratch.

KVM (Kernel-based Virtual Machine) is a full virtualization solution for Linux on x86 hardware. It exposes processor virtualization extensions (Intel VT-x or AMD-V) directly to guest workloads. KVM lets the Linux kernel act as a hypervisor. A virtual machine runs guest code directly on the host processor instead of relying on software-based instruction translation.

For Android emulators, that distinction determines whether the emulator is usable in CI. Without KVM, the emulator translates every ARM instruction to x86 in software, block by block. This makes boot times and test execution far too slow for CI pipelines. With KVM, the emulator runs guest code at near-native speed.

Standard EC2 instances are virtualized: multiple instances share the same physical host, managed by the AWS Nitro System . The Nitro hypervisor sits between your instance and the hardware, so it does not expose processor virtualization extensions to the guest by default. This blocked KVM on standard instances. Teams that needed KVM used bare metal instances, where the Nitro hypervisor handles security and networking but exposes the full CPU. Bare metal works, but the dedicated tenancy model makes it expensive at pipeline scale.

Nested virtualization, announced in February 2026, works by having the Nitro System pass processor virtualization extensions through to your EC2 instance. This creates three layers:

Your EC2 instance uses KVM the same way a bare metal host does, without the dedicated tenancy cost. AWS supports nested virtualization on C8i, M8i, and R8i instances.

Enable nested virtualization by checking the "Nested virtualization" flag under "Advanced Details"  when launching a new instance. If your instance already runs without the flag, stop it first, then modify the instance attribute to enable it before restarting.

Once the instance runs, verify KVM is available:

You want to see /dev/kvm exist, a vmx count greater than 0, and kvm-ok print "KVM acceleration can be used".

I tested on ami-0d13e2317a7e75c95 (ubuntu-noble-26.04-amd64-server-20260421), the latest Ubuntu LTS image at publication time. Install Java and the system libraries the headless emulator requires:

Each package links to the Ubuntu package index for reference:

kvm-ok passes under sudo, but the emulator runs as your user and needs group access to /dev/kvm. On a fresh instance, the kvm group has no members by default. Add your user explicitly:

If you are connected to the instance when you run this, log out and back in before the group change takes effect.

Persist the environment variables so they survive reconnects:

The per-vCPU rate is identical for virtualized and bare metal instances. The difference: bare metal forces you to buy the entire physical host regardless of how many emulators you run. All prices below are US East (N. Virginia), Linux, on-demand as of June 2026.

A c8i.4xlarge costs $0.75/hr and handles a typical emulator workload. That is 5x cheaper than m5zn.metal ($3.96/hr), a common pre-nested-virt choice, and 12x cheaper than an equivalent 8th-gen bare metal host.

Virtualized instances also unlock purchasing options that bare metal does not support. You can run emulator fleets on Spot Instances for up to 70% savings on interruptible CI jobs. Savings Plans let you commit to steady-state capacity at a lower rate. You can match instance size to the number of parallel emulators your pipeline runs instead of paying for a fixed physical host.

Pair Auto Scaling  with EC2 Warm Pools  to handle scale-out latency. Emulator instances take time to boot and reach a ready state. Pre-initializing a pool of stopped instances lets your pipeline scale out without a full cold start on every run. Save a custom AMI with all the changes above.

For the first time, KVM-accelerated Android emulators run on standard EC2 instances. That ends the bare metal tax. You no longer buy a full physical host that serves one job at a time. You run a virtualized instance sized to the emulators you actually launch, and you back bursts with Spot capacity at up to 70% off. The shift is bigger than a lower hourly rate. Right-sizing, Spot, Savings Plans, and Auto Scaling were all off the table on bare metal. Nested virtualization puts every one of them back in play for Android CI. The steps above boot a single emulator on KVM on a C8I instance. To get from there to a production fleet, do three things: bake a custom AMI from the instance you just configured, put it behind an Auto Scaling group, and front that group with a warm pool so scale-out skips the cold start. Same emulators, a fraction of the bill.