AWS announced GPU support for Event Functions on Lambda Managed Instances on November 18. Functions use G-series EC2 instances and can process multiple invocations concurrently within an environment. Announced pricing includes EC2 resources, a 9.75% management fee and $0.20 per million requests. This is not a price based solely on tokens actually generated by an AI model.

Managed execution has concrete constraints

Documentation specifies one GPU or GPU fraction per execution environment, container packaging and concurrency controls. It also notes an initial EC2 quota of zero vCPUs for the relevant G/VT On-Demand instances. Product availability therefore does not establish that a particular account can already obtain the required resources.

For a team processing images or documents as they arrive, this adds an architectural option. An event can initiate accelerated work without a separately written dispatch system for a self-managed GPU worker. Suitability depends on task size, arrival frequency and acceptable waiting time.

Measure concurrency alongside memory

Our engineering interpretation starts with the working set of one request. Several invocations may share an accelerator, but also compete for memory and execution time. More concurrency is not an automatic guarantee of lower latency for every task.

An acceptance example would compare a short stream, a sudden burst and sustained activity. Record startup time, total time to output, errors and peak memory. That would help distinguish startup limitations from processing or capacity limits.

The bill belongs to the complete configuration

Selection should include idle periods, scaling policy and time needed to load the model. We suggest calculating the actual daily workload profile before the cost of an idealized request. This is a comparison method, not a savings measurement performed by SIExpo.

A large container and temporary-storage allowance do not establish fast startup. Applications with a deadline need to measure from the triggering event to confirmed output, including the first execution after inactivity.

The announcement expands options for event-driven GPU computing. It does not remove resource budgeting or quota checks. SIExpo reviewed the announcement and documentation without deploying a function or comparing actual cloud invoices.