AWS announced GPU-backed Event Functions on Lambda Managed Instances on November 18, 2026. The capability runs container-image functions on Amazon EC2 G-series instances; it does not apply to every AWS Lambda function. AWS listed G4dn, G5, G6, G6f, Gr6, Gr6f, and G6e at launch. (AWS announcement)

AWS Lambda console overview of the lmi-gpu-inference function, with a function diagram and a container-image configuration panel.
Match the invocation mode to the job: the 90-minute window applies to eligible asynchronous and event-source-mapping invocations on Lambda Managed Instances.

What workloads fit AWS’s GPU-backed Lambda functions?

AWS Lambda settings panel showing general configuration, capacity-provider settings, and accelerator memory for one function.

AWS positions the capability for inference with custom models, rather than training. It also names media processing, rendering, and scientific computing as possible workloads. The functions run on Lambda Managed Instances, so a workload must use that execution model and a container image to use the GPU capability.

AWS’s demonstration used vLLM and Qwen/Qwen2.5-1.5B-Instruct. AWS reported that four concurrent calls finished in about the time of one, and that it took about 10 minutes to go from a model ID to a live endpoint. Those results describe AWS’s example. AWS also says the example downloads the model and warms up its inference engine while publishing a function version, before that version serves traffic. (AWS announcement)

Before trying that setup, AWS advises checking or requesting the account’s G-series On-Demand instance quota; new accounts may start with a quota of zero.

AWS lists function memory from 2,048 to 32,768 MB, accelerator memory from 3 to 48 GB, and up to 32 GB of ephemeral storage. Container images can be up to 32 GB, with a 10 GB limit per file. Each execution environment can handle up to 48 concurrent requests. These are service limits; the configuration in the console image is one example, not a universal workload setting. (AWS announcement)

How long can a Lambda invocation run?

AWS documents a 30-minute maximum for an ordinary invocation. For Lambda Managed Instances, asynchronous and event-source-mapping invocations can run for up to 90 minutes, except Amazon MQ and Amazon DocumentDB event-source mappings. (AWS Lambda quotas)

Invocation modeMaximum durationConditions
Ordinary invocation30 minutesOrdinary invocation limit
Asynchronous or event-source-mapping invocation on Lambda Managed InstancesUp to 90 minutesAmazon MQ and Amazon DocumentDB event-source mappings are excluded

That distinction matters when a job may take longer than 30 minutes: the 90-minute limit applies to the specified Lambda Managed Instances invocation modes, not to ordinary invocations.

What does AWS’s GPU pricing include?

AWS says GPU-backed Lambda usage incurs the underlying EC2 instance charge plus a 9.75% GPU-instance management fee. Lambda requests cost $0.20 per million requests. AWS says existing Savings Plans and Reserved Instances apply to the EC2 instance charge. (AWS announcement)

For a workload estimate, use the EC2 rate for the instance and Region you plan to run, add the GPU management fee, and account for request charges and runtime. The request rate alone is not the GPU compute charge.

What GPU metrics can you monitor?

Four AWS monitoring charts for GPU utilization, GPU memory utilization, GPU memory used, and GPU power utilization.

AWS says the Lambda console and CloudWatch provide GPU utilization, GPU memory utilization, GPU memory used, and GPU power utilization metrics. These let an operator track accelerator use and memory alongside the running function. (AWS announcement)