AWS Lambda can now run GPU-accelerated Event Functions on Lambda Managed Instances, using Amazon EC2 G-series instances. The generally available capability brings GPU workloads into Lambda’s event-driven programming model, with AWS managing instance provisioning, scaling, routing and GPU software compatibility.
Functions use container images built to the Open Container Initiative (OCI) standard. Customers can select an image, configure the GPU memory requirement and connect a function to AWS event sources without operating a GPU cluster or installing drivers and CUDA libraries themselves. The underlying compute remains Amazon EC2 capacity, with Lambda providing the managed execution layer.
Workloads and limits
AWS positions the feature primarily for inference and other GPU-accelerated tasks, rather than model training. Examples include custom machine-learning inference, computer vision, speech processing, document analysis, media transcoding, rendering, scientific simulations and financial modeling. It can also host small-to-medium language models embedded in applications. AWS distinguishes these custom deployments from large foundation models available through services such as Amazon Bedrock.
GPU functions can be composed with other parts of Lambda. They can respond to AWS events, including those from Amazon S3, Amazon SQS, Amazon API Gateway and Amazon EventBridge. Durable functions can coordinate longer workflows around a GPU step, while Lambda MicroVMs can run untrusted or model-generated code in isolated sandboxes before calling GPU-backed inference.
On Lambda Managed Instances, asynchronous and event source mapping invocations can use extended timeouts of up to 90 minutes. The published configuration limits include container images up to 32 GB, ephemeral storage up to 32 GB, function memory from 2,048 MB to 32,768 MB, and accelerator memory from 3 GB to 48 GB. Each execution environment can handle up to 48 concurrent requests, and a capacity provider can scale to as many as 15,000 vCPUs.
Container-based deployment
AWS describes a deployment path for serving an open-source model through a curated vLLM container image. In its example, the model identifier is set in the MODEL_ID environment variable; the function does not require custom inference-server code. The demonstrated model is Qwen/Qwen2.5-1.5B-Instruct, and the example uses Amazon Elastic Container Registry to store the image.
Customers first create a capacity provider and specify eligible GPU instance types, networking and scaling settings. AWS says listing multiple instance types lets Lambda select an available option from the configured list. The function then references that provider and requests accelerator memory through AcceleratorConfig. Lambda places the execution environment on a GPU that meets the requested minimum, which may have more memory than requested.
Managed Instances functions run from published versions. In the example, publishing provisions execution environments, downloads model weights and initializes the inference engine before the version becomes active, moving that startup work ahead of the first invocation. Model versions can be changed by updating MODEL_ID and publishing another version, without rebuilding the image, according to AWS.
The example uses vLLM to batch concurrent requests and share a model copy in GPU memory within an execution environment. AWS reports that four concurrent invocations in its test completed in roughly the time of one, and that the demonstration took about 10 minutes from model selection to a working endpoint. Those results describe the specific example, not a general performance guarantee.
Capacity, monitoring and pricing
At launch, supported EC2 G-series families are G4dn, G5, G6, G6f, Gr6, Gr6f and G6e. GPU instances are billed at the underlying EC2 instance rate, with a 9.75% management fee for GPU instances and the standard Lambda request charge of $0.20 per million requests. AWS says existing Savings Plans and Reserved Instances apply; customers seeking predictable GPU capacity can configure an EC2 On-Demand Capacity Reservation to match the capacity provider.
GPU-specific metrics—including utilization, memory utilization, memory used and power utilization—are available in Amazon CloudWatch and the Lambda console. AWS also lists curated CUDA images for tools including PyTorch, CuPy, vLLM and SGLang, alongside support for custom images. Seekable OCI lazy loading is supported, and AWS says it can reduce execution-environment startup time by up to 70%.
Customers should account for GPU instance quotas before deploying; AWS notes that new accounts may have an On-Demand G and VT instance quota of zero. For production inference, AWS recommends placing model weights in an S3 bucket in the same Region and configuring the serving image to stream them into memory. GPU capacity continues to incur charges while provisioned, so AWS advises deleting the function and its capacity provider when testing is complete.