Choosing CPUs by License Cost, Not Benchmarks

Benchmarks assume compute is the scarce resource. Licensing disagrees: when the database bills per core and the hypervisor per socket-or-core, the CPU decision becomes a cost function where silicon is the cheap term. Refurbished buyers hold an unusual advantage here — the exact core count that optimizes the spreadsheet is always in stock, at prices that make right-sizing free.

The models that do the reshaping

Per-core perpetual (SQL Server, Oracle with its multipliers): every physical core in the machine is billed, minimums apply (SQL: 4 per CPU), and the license routinely outcosts the server within a year. Per-socket/per-core subscription tiers (VMware's evolutions, various middleware): socket counts and per-socket core caps steer topology. Per-VM/per-user tiers elsewhere shift pressure off hardware entirely. Step one is always the same: name the licensing model before naming the CPU.

The SKUs built for this exact problem

Intel's frequency-optimized parts exist because enterprises did this math at scale: E5-2637/2643/2667 v4 (4–8 cores, 3GHz+) on 13G/Gen9 platforms; Gold 5122, 6134, 6144 and the 6250-class (8 cores, aggressive clocks, full cache and memory speed) on Scalable. Their trick is delivering top-bin per-core throughput at minimal core counts — precisely inverted from the many-core parts benchmarks celebrate. On the used market these SKUs often price below the core-heavy flagships, because virtualization buyers chase cores; license-aware buyers should quietly enjoy the mispricing.

Two worked examples

SQL Server Standard on a 2U: the reflex build (2× E5-2680 v4, 28 cores) licenses 28 cores; the license-aware build (2× E5-2667 v4, 16 faster cores) licenses 16 — a ∼43% license reduction that dwarfs every hardware number on the quote, while per-query latency improves. Consolidation under per-socket terms: one DL560-class 4-socket node versus two 2-socket nodes carries the same sockets — but density, failure domains and per-socket core caps decide which topology the contract actually favors; the hardware preference and the licensing preference diverge often enough that both must be computed.

The buying pattern

Write the license cost per core per year next to the used price of the CPU that carries it — the comparison is usually 100:1 — then spec cores to the workload's genuine parallelism plus margin, spend the freed budget on RAM and NVMe (unlicensed performance), and keep the swapped-out core-heavy CPUs for the Linux hosts where cores are free. On refurbished platforms every one of these moves costs lunch money to execute — which is why the license spreadsheet, not the benchmark chart, is where experienced buyers start.