At the AWS re:Invent conference in Las Vegas this week, the cloud giant has introduced Lambda managed instances, which blend the serverless model of Lambda functions with the ability to rent specific VM (virtual machine) instances.
Lambda is popular with developers for its ability to run functions on demand, with no need to manage servers and scaling to zero when not in use; but if used intensively it is expensive in comparison to running the same functions on EC2. Other issues with Lambda are cold start latency (unless paying extra for provisioned concurrency, which keeps functions permanently initialized) and lack of control over the hardware used other than the architecture (x86 or ARM) and the amount of memory.
Lambda managed instances are a new way to provision serverless functions, where the user can specify the exact EC2 instance type used and get discounted prices in return for commitments using compute savings plans or reserved instances. Each instance can handle multiple concurrent requests. These instances are only used for Lambda, and are managed automatically with operating system updates, load balancing, and a maximum lifetime of 14 days. Users define what AWS calls a capacity provider to specify the managed instances and how they should scale, with the initial default being three execution environments across availability zones for resiliency, scaled up automatically based on CPU resource consumption. The concurrency – meaning the number of functions invocations that run simultaneously – can be set to a maximum of 64 per vCPU (virtual CPU).
Pricing for Lambda managed instances is based on standard EC2 pricing plus a 15 percent “compute management fee” and a per-request charge of $0.20 per million. Managed instances eliminate the duration charge which applies to normal Lambda functions. This, combined with the ability to apply discounts, means that high volume users may get lower costs as well as better performance. For a light user of Lambda functions, the reverse is true, since the managed instances will be a waste of resources when no functions are running.
The initially supported runtimes are Java, Node.js, Python and .NET.
In a news post, AWS said the value of managed instances lies in “accessing specialized compute options and optimizing costs for steady-state workloads without sacrificing the serverless development experience.” There will be some large customers for whom managed instances will be a substantial benefit; the new enhancement reduces but does not eliminate the price premium for using the serverless model.
Alongside the Lamba news, AWS also previewed a set of solutions called EKS (Elastic Kubernetes Service) Capabilities, which allow users to add managed features including Argo CD (a continuous deployment tool), AWS controllers for Kubernetes which enable management of cloud resources, and Kube Resource Orchestrator for creating and managing custom resources.
Another key announcement is that of the AWS MCP Server, which combines the capabilities of the existing API MCP server and AWS knowledge server. Using the new MCP server, AI assistants will be able to execute tasks across multiple AWS services, improving the experience for AI prompters but also raising security concerns regarding risks from prompt injection, hallucination or other AI pitfalls. AWS promise that identity and access management as well as audit logging through AWS CloudTrail will give users “full control over resources and permissions.”
Attendees at re:Invent can expect to hear plenty more on this subject, with Swami Sivasubramanian, VP agentic AI, slated to deliver a keynote on Wednesday on the future of agentic AI.