Can't upvote this enough! Most people habitually build those overcomplicated things, and THEN base them on clouds where they will be 40x-200x more expensive to run than on dedicated machines, as a result, getting things that are fragile, expensive to build, and extremely expensive to run.
Do a serious calculation of a load you can hope to have, what it will take to process all of it, and build a simple solution and deploy it on a single dedicated machine of appropriate specs in a good datacenter, or if it requires low latency, several of them with simple DNS-based load balancing, and you will get your thing much sooner, and much cheaper, and a good dedicated server or one per region, will be enough for vast majority of startups.
I recently had a laughable story. A client of mine wanted to build an AI setup to analyse videos that come from security cameras. A security camera is 24x7 by definition. There is no justification whatsoever to use clouds for that: you won't suddenly have 100x the load in minutes if your load comes from fixed security cameras that operate 24x7. So 90%+ of what you pay for a cloud is a direct waste. It took days of painful discussions to explain that.
