Although AWS is making progress, I doubt that they'll be able to take over the entire HPC space[1]. First off, if a research group does not have access to a da

Although AWS is making progress, I doubt that they'll be able to take over the entire HPC space[1]. First off, if a research group does not have access to a data center[2], then AWS and GCP are the way to go. Just off the cuff, I'd say to go GCP unless you have big memory or specialized hardware needs, then go AWS. Other than that, AWS (and GCP) are great if you need to quickly spin up something to investigate the usefulness of new technologies: If you're not sure if a GPGPU or ASIC is going to work for your code, then test it at AWS. But, if you plan on doing _a lot_ of computing, then once you've done the validation, it's typically worth investing in the hardware so you can run it locally. Another reason, related to hardware, is that some places have better treatment for capital costs (like computing hardware) vs. regular expenses (like AWS and GCP time). If AWS could come up with a GAAP method for capitalizing AWS time, then that would help them out alot. MPI is mentioned alot in the article, and it's definitely an issue that affects the more "traditional HPC" crowd, but there's something else: Alot of clusters use Infiniband, for storage access and also for MPI (and other inter-node communication). Infiniband is not just high-bandwidth (single-link EDR is 25 Gbps, 4x is almost 100), it's also _very_ low-latency (less than a microsecond). The filesystems used over IB are similarly highly-performant (like Lustre). The speeds (and, just as important, the latencies) in HPC data centers are amazing, much better than what you get with 10 Gbps Ethernet. People who have experienced that want to keep it. However, IB is a different type of topology than Ethernet, which doesn't work as well with the "your X can land anywhere in an AZ" nature of AWS. I'd also suggest that AWS get involved in SLURM development. Don't take it over or absorb it; work with SLURM developers to get it to work naturally with AWS. So, unless (or until) AWS can address those issues, there's still a big place for on-prem HPC! [1] For anyone who might reply "AWS isn't thinking to take over all of it", I would say, in the backs of some people's minds, I bet they are. [2] "Data Center" could mean a rack or cage at a colo facility, but I am _not_ referring to a small server room or closet; you typically need more power and/or cooling than those rooms can provide.