cloud datacenters

A Hacker News discussion argues that the world's limited DRAM supply is badly allocated: consumer PCs hold large amounts of memory that sits idle most of the time, while datacenters could use it at far higher utilization. Commenters weigh whether centralizing memory in the cloud makes sense given current chip supply constraints, trading off efficiency against user control and privacy.

FreeToken, a Python project from FlashML-org, is drawing attention for promising datacenter-scale model serving on desktop machines. The tool claims to let users run very large AI models locally, quickly and efficiently, without cloud infrastructure. The project sits at the top ranks of trending repositories, and interest centres on whether it can really deliver heavy-duty inference on ordinary hardware.

Microsoft's Azure revenue jumped 48% to $33.2 billion in a quarter, yet the company is pushing forward with its in-house AI chip, Maia. Microsoft claims Maia 200 delivers 40% more AI work per megawatt than top GPU systems, and the chip already runs Copilot. With datacenters constrained by power rather than demand, efficiency is the key motive: designing custom silicon lets Microsoft squeeze more AI capacity out of every rack it builds.