Regarding AI and distributed systems, this might not be what you have in mind, but take a look at federated learning. I'm currently a computer science PhD candidate at a small school, and a couple of our graduates in the past year worked on the fundamentals and applications of federated learning.

I came into this PhD program thinking that I wanted to work on stuff like the distributed databases that you listed, or the stuff they're built on like clock synchronization. I did my master's degree in 2017-2018 and I was fascinated by an "advanced databases" class that covered these things. Unfortunately, nobody in my department works on such things, and I agree with you that I don't hear much about that area anymore.

I'm sure the next generation of supercomputers (eg anything upcoming with Nvidia's GB200 architecture, Tesla's Dojo, etc) will have their own share of challenges that need to be solved, at their specific scale.

No, AI has nothing to do with this. The late 1990s to early 2010s were a massive period for distributed systems research, but that surge was driven by advancements in computing infrastructure and the growing needs of industry giants—not AI. From theoretical frameworks like Chord to practical implementations like BigTable, Spanner, Cassandra, and TAO, the focus was on solving complex problems in scalability, consistency, and fault tolerance. The efforts to properly implement Paxos were part of that journey. AI had little to do with the sheer volume of breakthroughs that came from this time; it was all about improving distributed systems, not artificial intelligence. Not everything needs to loop back to AI.

Everything is a distributed system now, and system designs are changing in important ways. As you indicate, we you are currently seeing a redesign of basically all computer systems from mobile to datacenter/cloud/hpc to better support AI workloads. I expect there are still many opportunities for distributed systems research.

If anybody needs a practical launching point to study distributed systems maybe I can help.

I have a browser/Node.js application that I worked on for several years that is a browser based peer-to-peer file system sharing tool. The application is massive in size, but its easy to tinker with, the build takes about 8 seconds, and the test automation is decent enough for anybody interested to get started. The idea is to invert the transmission model of the web from anonymous connections towards a server instead to trusted connections between owned computers and even restricted connections to different users (and their various devices) all in real time using sockets and streams and a tiered distribution model. https://github.com/prettydiff/share-file-systems

I am currently working on a web server application that allows spinning up many different web servers quickly and a monitoring dashboard that shows the various servers, their connected sockets, port management, and various other things. This all started with the idea that you can proxy and redirect anything super efficiently where a proxy in Node.js is as simple as:

    proxy.pipe(socket);socket.pipe(proxy);