First part is billing model. You pay per request instead of per server. Say you have a server that can do 100 request/s. But you’re actually only doing 2. You’re overpaying by 50x if you pay for a server. At a small scale this difference is huge, as you grow it converges though.

But really, most application developers don’t care, and don’t want to care, about the underlying servers. Figuring out how autoscaling works, or keeping the kernel up to date. Individually none of these things are that hard, but there’s a bunch of time that adds up. Time that could be spent adding features that make money.

There is still a lot of marketing hype around it. Things like infinitely scalable fall apart when you have connection limits on a db and lambda doesn’t support pooling. Aws pushes it because it gives them extra flexibility - doubling the density of functions is probably easier than instances.

I don’t think the current form of api gateway + lambda is the final state, but something more like fargate, where you get a whole container and no time limit, is really interesting. Or we can all just go back to heroku.