With Moore's law now still continuing to double transistor counts every 2.5 years or so, if AWS was serious about passing on savings, it'd reduce the cost of EC2 instances by about 5% _every couple of months_. That's clearly not happening, and because compute isn't exactly fungible (despite being called a "commodity") thanks to things like subsidized data ingress, AWS is making an absolute killing. The margins, even after building huge datacenters, are ridiculous.
CPU cost is only part of the total cost of an EC2 instance.
According to James Hamilton who speaks on behalf of Amazon, Servers make up ~57% of the cost of running a DC. CPUs are only a fraction of the cost of a server, maybe 25%.
So moores law only really applies to ~15% of the total cost of EC2. That's before factoring in all of Amazons capex to build the software that is EC2.
Yes, the CPU cost is only a fraction, however, the costs of the rest of server (ie. storage and memory) also is dropping precipitously, and has for decades. The cost of networking equipment and any network transit should be factored into the cost of network egress (which starts at healthy $0.09/GB).
The capex on software is negligible, because the cost of duplication is effectively zero. AWS can scale to whatever size they want (as long as their software is architected correctly, of course), and not have to spend more cash.
That said, I'm not faulting AWS for charging what they charge. They have a remarkable service. It's just hard to stomach a post bragging about dropping the price by 5% when they're effectively printing money.
Do remember Intel builds the CPUs, prices them, and basically has no competition.
I have no education/insight at all into this matter, but I would keep that in mind and reconsider your assessment of Moore's law's direct influence on AWS' bottom line.
It's not like storage or network bandwidth are sitting still pricewise. Especially on the SSD front. What was a reasonable price 2 or 3 years ago for some amount of SSD would be considerably overpriced today.
Wouldn't that only be true if they replaced/upgraded everything every few months? You'd have to factor in the cost of that. And of course, it wouldn't work that way anyway because Moore's Law isn't actually continuous. It's is a line drawn through discrete points in the timeline (new product launches from chip makers).
Also, correct me if I'm wrong, but Moore's law doesn't say anything about ancillary costs like power/air conditioning. I have no idea what the trend is for those.
It doesn't matter what the trend is, it just matters that Moores law only applies to a fraction of the cost of a server, which is only a fraction of the cost of running DCs.
With AWS, you're paying for so much more than just the bare cost of the resources. Like it or not, things like CodeDeploy, Auto-scaling triggers, OpsWorks, Elastic Beanstalk and all the other services that are "included" ("pay only for the resources you use") are likely factored into the EC2 costs.
