GPU rental vs purchase break-even - Cloud per-GPU-hour vs used hardware asking prices

DERIVED ANALYSIS, not a primary source. This record synthesizes data from multiple records in this repo to compute the rental-vs-purchase break-even point. Using Together AI's on-demand H100 rate ($3.99/GPU-hr) as the rental benchmark and the used H100 SXM5 asking median ($26,880) as the purchase price: break-even = $26,880 / $3.99 = 6,737 hours = 281 days of 24/7 usage. At the cheapest managed rate (Beam H100 $1.83/hr): break-even = $26,880 / $1.83 = 14,689 hours = 612 days. At CoreWeave on-demand ($6.16/GPU-hr): break-even = $26,880 / $6.16 = 4,364 days = 182 days. At AWS p5 on-demand ($6.88/GPU-hr): break-even = 3,907 hours = 163 days. KEY INSIGHT: At current cloud rates, the break-even point for buying vs renting an H100 is 5-21 months of continuous 24/7 usage. For most inference workloads that are not 24/7, renting is cheaper. The break-even only favors buying for steady-state, high-utilization workloads. This analysis excludes electricity, cooling, colocation, and maintenance costs for the purchased hardware, which would extend the break-even period. It also excludes the resale value of the hardware at end of life, which would shorten it. The used H100 SXM5 asking median of $26,880 is from the market-gpu-price-bands-2026-10 record (n=5, small sample).