An 8×H100 SXM node lists around $280k. Add InfiniBand, NVMe, power (10kW sustained ≈ $12k/yr), and a part-time cluster engineer, and year-one true cost lands near $330k — before downtime.
If sustained utilization is under 45%, renting wins outright. Between 45–70%, reserved rentals ($1,399/mo/GPU) still beat owning once you price power, spares, and upgrades. Only north of 70% 24/7 for 3 years does buying pull ahead — and then you're stuck on H100s while H200s ship.
See live rental pricing → or get a TCO sheet for your workload →