Buy it or rent it?
Used datacenter GPUs, especially A100s coming off hyperscaler and lab fleets, now sell for a fraction of launch price. This calculator turns a purchase price, your power cost, and your utilization into an effective $/GPU-hour, and shows the break-even point against renting the same card.
Defaults: indicative secondary-market unit price (2026-09); power modeled at 70% of TDP while busy plus 40% datacenter overhead (cooling, PUE); hosting/colo and ops labor excluded, add your own numbers mentally or in the price field.
Your effective cost
| $/GPU-hr | 3-yr cost @ util | Notes |
|---|
Reading the result. Ownership wins on sustained utilization and loses on idleness, the crossover is usually between 25% and 50% utilization at 2026 prices. Rental prices are our tracked on-demand list rates (2026-09); reserved contracts cut them 30-70%, which moves the crossover up. If the answer is "own," the next question is what the card can actually serve, check the VRAM calculator and the A100 in 2026.
If the math says own but you don't want to operate: Charg, from the team behind Virtualized, runs dedicated single-tenant A100 clusters at a recovered cost basis under contract: validated hyperscaler-grade fleets, U.S. datacenters, InfiniBand, deployable now. Reserve capacity →