Elasticity is very expensive, in practice people only use it for one-off jobs, preferably using the cheaper unreliable "spot" instances (meaning the job must support being partially re-run to recover, which implies a complex job splitting and batching platform).

For traditional, always-on servers, you should reserve them for 3 years. You still have the ability to scale up, just not down. You can always go hybrid if you don't know what your baseline usage is.