Short, vendor-neutral guides for people deciding what to buy, what to rent, and what to ignore. Every figure dated, every assumption stated.
Two guides so far. The first is about what the hardware can do. The second is about whether it's worth owning.
Three local deployments, priced against the APIs serving the same models. A $7,000 desktop pays back in three months. An $83,000 server takes nearly three years. A third machine never breaks even at all — and the reason has nothing to do with the silicon.
Read the cost breakdownCapacity only answers whether a model loads. Bandwidth decides whether you can stand to watch it answer. A field guide to memory, model architecture and quantization — then six classes of machine mapped against each other at current prices.
Read the hardware primerMost writing about local AI is either a vendor's spec sheet or someone's enthusiasm. Neither tells you whether the machine will do what you need, or whether buying it makes any sense against the alternative.
These guides try to answer both, with the arithmetic shown so you can substitute your own numbers. Where a figure is modeled rather than measured, it says so. Where the honest answer is "don't buy it," it says that too.