[electrograv]

Another case where self hosting can be much less expensive (by orders of magnitude, in fact) than cloud providers is compute-bottlenecked tasks, e.g. machine learning experiments with huge training sets.

A single self-hosted compute server that costs ~$20/month in electricity to run 24/7 would cost over $1500/month from AWS.

[Xylakant]

This may be swayed by the fact that you might have a task that can be run in parallel on multiple machines and then finish in an hour, so that you can then move on. Buying or renting enough capacity and then leaving idle for the remaining time is not an option, but with AWS you can easily spin up 100 instances for an hour and do the computational worth for 4 days of a single machine, then shut them down at no extra cost.

[vonmoltke]

That is the use case that AWS is best for. That is not the use case covered in the article, though. The use case being discussed is that for a cluster of always-on machines with a fairly consistent load. If you have a minimum baseline load on your servers, fulfilling that from AWS, or Azure, or similar services doesn't make financial sense. Colocated (if you don't mind dealing with the hardware) or leased (if you do mind) servers for your baseline load combined with cloud machines for load spikes makes the most sense.

[svmegatron]

OTOH, you may have to buy the self-hosted server, which could sway the cost calculation back in favor of cloud hosting.

[kayoone]

I am totally not convinced by AWS, but to be taken seriously you will have to factor in the initial price of that machine if you make a comparison like that.