Power Draw in Practice: What a Server Costs to Run

Peak wattage sells PSUs; idle wattage writes the electricity bill. Servers spend most hours far from peak, so the number that compounds across a year is the floor the machine settles to at 2 a.m. That floor is mostly a configuration outcome — which means it is mostly under your control at purchase time.

Real numbers, honestly bracketed

Measured at the wall, typical refurbished builds land near: a single-CPU 1U (one mid-TDP Xeon, 4–8 DIMMs, two SSDs) idles ∼60–90W; a dual-CPU 2U workhorse (2× mid-range, 8–16 DIMMs, a few drives) ∼100–150W; a storage-heavy 12LFF adds ∼5–8W per spinning drive to sit near 160–220W; four-socket and GPU builds start north of 200W before doing anything. Load adds roughly the CPUs' TDP-and-change on top. Configuration swings these brackets more than the badge on the bezel: each populated socket adds ∼20–40W of idle, every DIMM ∼2–4W, every spinner its handful, every PSU a few watts of overhead.

Watts to money, one line

Yearly cost ≈ watts × 8.76 × price-per-kWh. A 120W machine at $0.15/kWh: ∼$158/year; at European $0.30: ∼$315. Useful corollary: every 10W trimmed is $13–26/year forever — which prices the upgrade decisions below rather precisely.

Trimming that actually works

Population choices first: one socket instead of two where cores allow (the second CPU is a live 30W decision), fewer larger DIMMs over many small ones, SSDs where spindles are not earning their watts, right-sized PSUs (a 495W unit runs closer to its efficiency sweet spot than an idling 1100W pair — though keep redundancy where uptime demands it). Firmware settings second: BIOS power profiles set to OS-controlled/balanced rather than maximum-performance let C-states and P-states do their considerable work — the single biggest free lever, worth 15–40W on many duals. Fleet shape third: consolidation — three half-idle boxes onto one well-loaded hypervisor — remains the largest saving available, measured in whole machines.

What not to over-optimize

Performance-critical hosts deserve their performance profile; databases dislike deep C-state latency; and chasing single-digit watts with risky settings on a machine earning its keep is hobby arithmetic. The honest framing for refurbished buyers: a 13G box drawing 30–50W more than its 14G successor costs $40–130/year extra — against a purchase-price gap many times that. Buy the efficient configuration; let the generation follow the workload; and put a metered PDU or smart plug on the rack, because every number in this article improves the moment it is measured instead of guessed.