diff --git a/docs/ci/ec2-runners.md b/docs/ci/ec2-runners.md
new file mode 100644
index 0000000..e688ccb
--- /dev/null
+++ b/docs/ci/ec2-runners.md
@@ -0,0 +1,28 @@
+# CI/CD with EC2 Runners
+
+## Problem
+
+Projects run E2E tests on EC2 runners for small, medium, and large jobs. (Some
+projects may have other names for such jobs, like `smoke` in `training`).
+
+These runners are used to get access to accelerated hardware (e.g., GPUs) to
+run compute intensive processes.
+
+Access to instances with such hardware is sometimes limited and depends on the
+current demand among all EC2 users in a particular zone. This means that
+sometimes requested instance types are not available, which makes jobs that
+rely on these instances fail.
+
+## Solution
+
+Availability depends on a particular zone. If a zone is busy, we can try
+another zone.
+
+For this, a new
+[`launch-ec2-runner-with-fallback`](https://github.com/instructlab/ci-actions/tree/main/actions/launch-ec2-runner-with-fallback)
+action was implemented in `ci-actions` repository. If adopted, this action will
+walk through AZs and try to request an instance in each AZ until it finds one.
+
+All projects that rely on AWS EC2 runners should adopt the
+`launch-ec2-runner-with-fallback` action in all of the jobs to avoid fluke test
+failures.