How this is calculated Close The arithmetic, with your inputs Where each input comes from The rate — one published per-GPU on-demand price. Spot and committed contracts are cheaper and are not used: they are not the same product.Median or cheapest — median is the middle rate across providers, one vote each. Cheapest is one provider, on one day, possibly with no capacity.Cards to buy — the most you run at once, not the sum across stages. The same cards get reused when one stage ends.The purchase price — yours, because we cannot collect it. Most data-centre GPUs have no list price; an H100 SXM5 is not sold as a card.Board power — the vendor’s own figure, archived with a hash like every price here. The B200 is the exception: it has no board figure, only ~14.3 kW for an eight-GPU chassis, so it is offered in whole chassis and multiplied rather than divided.What the power figure leaves out It is boards at rated power and nothing else. Too high in one way:
Nothing runs at rated power — it is a ceiling, and some parts are configurable lower. Too low in three:
CPU, memory, NVMe, fans and supply losses all draw too. Cooling is not counted — the facility’s overhead, or your air conditioning. Nor is anything else about owning hardware: host server, rack space, network, spares, failures, resale, and someone to run it. We could fold those in with plausible multipliers. We do not: a coefficient presented as a measurement is the one thing this site will not publish.