commodity hardware
There are two ways to build a fleet of delivery vehicles. One is to buy a handful of custom-engineered armored trucks, each expensive, individually maintained, and very reliable. The other is to buy a large number of ordinary mass-produced vans: each one is cheap, none is special, any of them might break down — but they are interchangeable, and you keep spares. Warehouse-scale computers take the second approach, and the ordinary vans are called commodity hardware: standard, mass-produced, off-the-shelf servers and parts, chosen because they are cheap and plentiful rather than because any one of them is exceptional.
Concretely, commodity hardware means building the fleet from the same widely-available components the whole market uses — mainstream server CPUs, standard DRAM modules, ordinary disks and solid-state drives, common ethernet — rather than custom, exotic, or specially-hardened parts. The payoff is cost: mass production makes these parts cheap, competition keeps prices down, and interchangeability means a failed unit is replaced by an identical one from a deep pool of spares. The trade is that any individual commodity part is only ordinarily reliable; it is not built to never fail. WSC designers accept that on purpose and lean on redundancy and software to turn a pile of so-so machines into a dependable whole — reliability bought at the system level, not the component level.
Commodity hardware is one of the defining choices that separates a warehouse-scale computer from a classic supercomputer. The supercomputer tradition buys the fastest, most reliable, most specialized hardware money can build; the WSC tradition buys boring, cheap, replaceable parts in vast quantity and engineers reliability and performance across the whole fleet in software. This choice flows straight from the economics: when you are buying tens of thousands of machines, a small per-unit saving multiplied across the fleet is enormous, and paying a premium for individually gold-plated reliability makes no sense when redundancy already covers failures.
Two honest caveats. First, 'commodity' does not mean 'whatever is cheapest off the shelf with no thought' — the parts are standard, but at WSC scale even tiny differences in efficiency, failure rate, or power draw multiply across the fleet, so the selection is careful and the hardware is often customized at the edges (custom racks, boards, or power) while staying built from commodity silicon. Second, the trend is not purely toward commodity: for some workloads, specialized accelerators earn their keep, so modern WSCs mix mostly-commodity servers with a layer of purpose-built hardware where the volume justifies the design cost.
A WSC operator buys 50,000 identical servers built from mainstream CPUs, standard DRAM, and ordinary SSDs — each a few thousand dollars, none specially hardened. When one fails (and several do every day), it is swapped for an identical spare and the software has already routed work to the replicas. Cheap parts plus redundancy beat expensive parts alone.
Cheap, interchangeable, ordinarily-reliable parts in vast quantity — with redundancy supplying the reliability.
'Commodity' is not 'whatever is cheapest, chosen carelessly'. At fleet scale tiny per-unit differences in power and failure rate multiply hugely, so parts are picked carefully and often customized at the edges.