Warehouse-Scale & Datacenter Computing

total cost of ownership

Buying a car is not the real cost of owning one. The sticker price is just the start; over the years you also pay for fuel, insurance, repairs, tires, and parking — and those running costs often add up to more than the purchase itself. Judging the car by its price tag alone would badly mislead you. Total cost of ownership is the discipline of adding up everything a thing costs you over its whole life, not just what it costs to buy. For a warehouse-scale computer, TCO is the master number that every design decision is ultimately judged against.

TCO splits into two broad buckets. Capital expense (capex) is the up-front money: the building, the servers, the networking gear, the power and cooling infrastructure — paid once and amortized (spread out) over the years of useful life. Operating expense (opex) is the ongoing money: the electricity to run the servers, the electricity to cool them, the people who maintain the fleet, the bandwidth, and the steady replacement of hardware that wears out. To compare designs fairly you convert everything to a common basis — often cost per server per month, or cost per request served — by spreading the capex over the hardware's lifetime and adding the monthly opex. A striking result of doing this honestly is that energy (power plus cooling) is a large and growing slice, sometimes rivaling the amortized cost of the servers themselves.

TCO matters because it reframes what 'a good design' means at scale. A cheaper server that burns more power, or a denser rack that needs costlier cooling, can easily lose on TCO even if it looks better on the spec sheet. This is why WSC architects obsess over power usage effectiveness, energy proportionality, and reliability: a watt saved is saved every hour for years, and a machine that fails often costs money in replacement and lost work. The right question is never 'what is the fastest or cheapest box' but 'what minimizes total cost per unit of useful work delivered over the fleet's lifetime.'

An honest caveat: TCO is a model, and models depend on assumptions — the price of electricity, the lifetime you assume for hardware, the cost of money over time, and how heavily the equipment is utilized. Change those assumptions and the ranking of two designs can flip. TCO is indispensable for making rational tradeoffs, but it is an estimate built on forecasts, not a precise truth, and the honest practitioner states the assumptions out loud.

A $3,000 server amortized over 4 years costs about $63/month in capex. If it draws 200 W and power-plus-cooling costs $0.15/kWh at a PUE of 1.5, its energy bill is about 200 W x 1.5 x 720 h x $0.15/kWh = about $32/month. Energy alone is roughly half the server's amortized hardware cost — which is why a watt saved matters so much.

Capex (amortized hardware) plus opex (energy and operations): at WSC scale, energy is a huge share of the total.

TCO is a model built on assumptions — electricity price, hardware lifetime, utilization, cost of money. Change them and which design 'wins' can flip. State your assumptions, because the ranking is only as honest as they are.

Also called
TCOtotal cost擁有總成本TCO