I have an idea to create an EC2 instance with a CUDA AMI (with NVIDIA GPUs), deploy stable diffusion with Flux Dev and create an API from the EC2 instance only open to the VPC. Then creating a Lambda function with an external API that will:

1. The Lambda function starts the EC2 instance with a CUDA-enabled AMI.
2. It waits for the instance to be in running state using a waiter.
3. Calls an API endpoint on the EC2 instance using the VPC.
4. Stops the EC2 instance after receiving the API response.
5. Returns the API response from the instance back to the public API Gateway endpoint.

I understand that the inference time will be long because of the time it takes to start the instance, but the main purpose is to cut down costs and only have the expensive GPU EC2 instance running when needed.

Has anyone else tried this? Can this actually be feasible?