@shitpostalotl @parismarx AI is the solution to the cloud provider problem. The problem is that the economies of scale for the cloud depend on compute being expensive. A cloud provider may be able to get 60% average utilisation, whereas an on-premises server may get only 5-10% (but needs to be able to handle the peaks, so is massively overprovisioned for the common case). A 128-core system that consolidates a load of workloads is cheaper than 16 eight-core ones. This is great, except that computers keep getting faster. In the ‘90s, you needed a big system to handle corporate accounting for a medium-sized enterprise, now a RPi can do it with a 5W power budget. There was a story a few days ago about a payment processor that handles 20k transactions per second. They have 60 servers, worldwide, and a lot of those are for redundancy. Every generation of improved compute pushes the efficiency arguments towards simpler hosting.

For clouds to be successful, they need compute demand to keep increasing, ideally at a faster rate than efficiency increases. Demand for compute for most tasks grows at a rate aligned with number of employees or number of customers, which means that the asymptote is the rate of population growth in regions with disposable income. AI, in contrast, demands huge amounts of compute power. If you can convince businesses that they need to put an AI in their products or loose out when their competitor does it first, you’ve guaranteed growth for the next few years. During that time, you can hunt for the next thing that demands vast compute resources.