AWS announced on November 18, 2026 that Lambda Event Functions can now run on GPU hardware through Lambda Managed Instances, backed by Amazon EC2 G-series machines. Lambda handles the GPU drivers, frameworks and scaling, so developers can run inference, media processing or rendering as a normal step in an event-driven application.
Running GPU code has usually meant provisioning servers, matching CUDA and driver versions, and writing custom autoscaling. AWS says this launch cuts that work to minutes, and the functions plug into the same 200-plus AWS event sources as any other Lambda function. It is available in all commercial Regions where both Lambda Managed Instances and G-series instances exist.
What Is New
Lambda Managed Instances already let functions run on EC2 capacity that Lambda operates in your account. The new option adds GPU instance types to that model. Lambda launches the instances, patches the operating system and GPU drivers, runs health checks and scales the fleet.
- Supported families include G4dn with NVIDIA T4 (16 GB per GPU), G5 with NVIDIA A10G (24 GB) and G6 with NVIDIA L4, among others listed in the docs.
- Ready-made base images bundle CUDA, PyTorch and CuPy along with vLLM and SGLang for serving models, tested against the driver Lambda installs.
- Both the container image and local NVMe scratch space can go up to 32 GB for large model artifacts.
- Startup is sped up with Seekable OCI (SOCI) lazy loading and by streaming weights straight from Amazon S3.
- GPU, GPU memory and GPU power utilization metrics appear in CloudWatch and the Lambda console.
How It Works
You first create a capacity provider whose allowed instance types list only supported GPU instances.
You can include several G-series generations, and Lambda picks the newest one that fits, dropping to an earlier generation if the newer one is unavailable. Next you build a container image from a Lambda GPU base image or your own, then create the function and use AcceleratorConfig to state how much GPU memory it requires.
Publishing a function version tells Lambda to launch GPU instances, pull the image and start execution environments, after which the version becomes active.
Each environment serves several invocations at once, controlled by the PerExecutionEnvironmentMaxConcurrency setting. Lambda adds or removes instances based on how busy the GPUs and CPUs are and on concurrency, and scale-to-zero can be turned on to avoid paying for idle GPUs.
One practical catch is quotas. The EC2 On-Demand quota for accelerated instances starts at 0 vCPUs, so you must request an increase before your first deployment. Configuration works through the console, AWS CLI, SDKs, CloudFormation, CDK and SAM, and AWS also offers a Lambda Managed Instances skill for coding agents in its Agent Toolkit.
Pricing
GPU functions use the same pricing model as other Lambda Managed Instances. The bill covers On-Demand EC2 rates for each instance, with Savings Plans and Reserved Instances applying, plus a 9.75% Lambda management fee. Requests cost $0.20 per million.
That makes cost depend mostly on instance hours rather than per-millisecond duration. Keeping GPUs busy with concurrent invocations, and letting idle capacity scale to zero, is the main way to control the bill.
What It Means for Developers
For backend and data teams, this brings GPU compute into the serverless toolbox without a separate cluster.
A file landing in S3, a message on SQS or an API Gateway call can trigger a transcoding job, an image model or a speech-to-text pass with no glue code. That suits workloads that are bursty and tied to events more than long-running training jobs.
AWS itself points users to Amazon Bedrock for foundation models and to SageMaker AI for training and full model lifecycle control. Lambda GPU functions fit custom small and mid-size models that live inside an application, along with rendering and numerical computing. Pick the tool by who should own the model serving stack.
If you already tested the earlier ability of Lambda Managed Instances to scale to zero, that setting is especially useful here because idle GPU hours are expensive. Start with a small instance family such as G4dn or G6, measure throughput with realistic concurrency, and only then widen the allowed instance list.
Getting Started
Request the EC2 accelerated instance quota in your target Region, then follow the getting started guide in the Lambda Developer Guide. Use a Lambda GPU base image to skip driver matching, and enable SOCI and S3 streaming for large weights. Watch the GPU metrics in CloudWatch to tune per-environment concurrency.
Conclusion
GPU support on Lambda Managed Instances lets developers add accelerated steps to event-driven applications while Lambda runs the drivers, images and scaling. Pricing follows EC2 rates plus a management fee, so utilization planning matters. For teams that want inference or media processing without running a GPU cluster, it is a notable new option.
Sources
- AWS What's New: AWS Lambda adds support for GPU instances on Lambda Managed Instances
- AWS Lambda Developer Guide: Run GPU-accelerated Event Functions on Lambda Managed Instances
See my articles more often when you search Google.
