GPU instances can take minutes to cold start. If your workload has sparse invocations or burst traffic, the time spent recreating instances after each request adds significant latency to subsequent calls.
The delayed release feature keeps GPU-accelerated instances alive for a configurable period after a function invocation completes, so the next request reuses a warm instance instead of waiting for a new cold start, improving service stability. The default delay is 5 minutes. Starting at 00:00 on November 12, 2025, Function Compute is rolling out delayed release for GPU-accelerated instances by region, covering all GPU card types. Availability dates for specific regions will be announced as the rollout progresses.
Billing impact
On-demand instances
After delayed release is available in your region, Function Compute checks the instance status every second after a function invocation completes. Instances are billed based on their status at the rate for either active or idle on-demand instances. The idle window does not exceed the delay duration you configure (default: 5 minutes).
Keeping an instance warm reduces cold starts, but you pay for the duration at the applicable on-demand rate. Set a delay that matches your traffic pattern to balance latency against idle costs.
For rate details, see pay-as-you-go billing.
Provisioned instances
No additional charges apply when using delayed release with the subscription billing model.