At a guess the downsides of Arm instances are lower performance, reduced software support, leading to potentially increased setup and support costs. Or maybe it's fast enough for most things, you can of course scale horizontally, and software support is ok.
In AWS small amounts of compute are fairly cheap, it's storage and bandwidth that can really increase the costs. t2 instances are priced well compared with A instances, though you get burst CPU rather than constant, but that usually works fine for web servers.
A t3.micro or even a t3.nano can meet a surprisingly high load, when set up correctly with caching, costs less than A instances, and is Intel x86 based so support is probably easier. AMD t3a instances are 10% cheaper again, and I'm running them fine with the same stack that ran on t2 Intel instances.
I have a LEMP stack running on a t3a.nano instance, with a little bit of swap, performance is not as good as Intel based but is close and is good. The main trick for running a full stack on a 512MB RAM instance is to disable the performance schema.
1A few hundred dollars? – Michael - sqlbot – 2019-07-27T00:30:09.340
3 year reserved instance – Kevin Swindon – 2019-07-30T20:15:08.783