GPU Buy vs Rent Break-Even Calculator

The calculator compares two ways to get GPU time. Owning means you pay the hardware price once (capex), then only electricity and any hosting fee each month. Renting means no capex, but you pay a flat cloud rate for every GPU-hour you use. Because the cloud bill scales with usage while your electricity cost per hour is far lower, heavy users eventually recover the purchase price — the moment that happens is the break-even.

Every month is treated as 730 hours (8,760 ÷ 12). The math is fully deterministic:

1. Power cost per active hour= TDP(kW) × PUE × electricity rate. PUE (Power Usage Effectiveness) grosses the card's draw up for cooling — about 1.1 for a home rig, 1.4–1.6 in a data centre (Uptime Institute). A 450 W card at PUE 1.1 and $0.13/kWh costs ≈ $0.064/hr.
2. Monthly active hours = utilization% × 730. Utilization is the share of wall-clock the GPU is genuinely busy.
3. Owned running cost / month = power cost per hour × monthly active hours + fixed hosting. Capex is excluded here — it is paid once, up front.
4. Cloud cost / month = cloud rate × monthly active hours, using published on-demand/community rates (Lambda, RunPod, Vast.ai, AWS).
5. Monthly saving from owning = cloud cost − owned running cost. If this is zero or negative, owning never gets cheaper and the tool says rent — no division by zero.
6. Break-even months = capex ÷ monthly saving, where capex = purchase price + extra rig cost. Multiply by monthly active hours for the break-even in GPU-hours.
7. Effective owned $/hr over an horizon of M months = capex ÷ (active hours × M) + power cost per hour + hosting ÷ active hours. At the break-even month this equals the cloud rate exactly — the tool verifies the break-even a second way by intersecting the two cumulative-cost curves, so the number is checked, not just asserted.
8. Verdict: if the monthly saving is zero or less, rent. Otherwise, if break-even lands within your ownership horizon (default 24 months), buy; if it lands beyond, rent.

Version 1 keeps the model honest and conservative: the purchase price is treated as sunk (no resale credit), the purchase is assumed to be cash (no financing interest), and cloud pricing uses steady on-demand rates rather than volatile spot bids. That means a "buy" verdict is a floor — resale value would only improve it.

• NVIDIA — GeForce graphics cards (RTX 4090/5090 MSRP & board power)
• NVIDIA — Data Center GPUs (A100 / H100 specs & TDP)
• Lambda — GPU Cloud on-demand pricing
• RunPod — GPU cloud pricing (community & secure)
• Vast.ai — marketplace GPU rental pricing
• AWS — EC2 P5 (H100) instance reference pricing
• U.S. Energy Information Administration — average electricity price
• Uptime Institute — Power Usage Effectiveness (PUE) guidance

Preset prices, cloud rates and TDPs were last re-checked against these sources on 2026-07-15. They are illustrative references, not a live feed — every field is editable so you can drop in a current quote. This is a cost break-even, not a performance benchmark: it does not claim a rented GPU matches an owned one in speed.