I don’t particularly like the either or approach. You can certainly spin up some minimum on local hardware. You have some up front capex but something that doesn’t have a fluctuating, expensive monthly opex bill.
You can then use cloud architecture to add capacity resources on demand and in different geographic locations. You can also utilize multiple cloud architectures to further add redundancy and cost optimization.
If you build out the scripts used to dynamically scale to also pull current pricing, you can have something that is both heavily redundant and somewhat cost effective. Sure it’s not like azure, AWS, Google cloud, or any other public cloud option changes their pricing that frequently, but it would give a good way to compare specifically in different regions.
For a game like this, building capacity and the ability to scale early was clearly more important than optimizations in the server code base. 500k/mo isn’t actually a lot to companies and it’s likely to go down as optimizations are implemented and popularity stabilizes.