There's a few things you can look into:

• Running FME Cloud on a schedule (and workspaces on a schedule too). This way it will start at a specified time, run one or more workspaces and then shut down at a specified time. This is useful for batch processing, when the load is more or less the same every time. You can set this up in the FME Cloud management console.
• Using the FMECloudInstanceController custom transformer, that way you can trigger an FME Cloud instance to start or stop. Using two with a FMEServerJobSubmitter in the middle you can trigger it on demand.

Do keep in mind that when your FME Cloud is busy with a job and it gets a signal to shut down (either from the schedule or through an FMECloudInstanceController) it will shut down, no questions asked. A running job will thus be terminated (but should restart when the instance next starts). So you'll need to make sure that if you run it on a schedule you give it enough time to finish the job. For the on-demand option you'll probably need to introduce a bit of delay to allow the instance to properly start before submitting the job.

Hope this helps.

You can also use the FME Cloud API. This blog walks you through using the FME Cloud API in conjunction with AWS SQS to bring up an instance when so many jobs enter the queue.

https://blog.safe.com/2015/02/event-driven-architecture-aws-lambda-s3/

Do not forget that each time you're using FME Cloud, even if it's a 5 seconds process, one hour is compatibilized in the record.