Falling rental prices have changed the rent-vs-buy math. In 2026, renting wins for almost everyone except those running GPUs flat-out for years. Here’s the break-even.
The simple break-even
An H100 board costs roughly $25,000-30,000+. At the cheapest rental of ~$2.30/hr (see the ranking):
- Continuous rental = $2.30 x 8,760 hr/yr ≈ $20,000/year.
- Buying = ~$30,000 up front, plus power, hosting, maintenance and depreciation.
So even at 100% utilization, year one slightly favors renting; buying only pulls ahead over 2-3 years of high, sustained use.
| Sustained utilization | Rental cost/yr (@ $2.30/hr) | Verdict |
|---|---|---|
| 20% (bursty) | ~$4,000 | Rent — easily |
| 50% | ~$10,000 | Rent |
| 70% | ~$14,000 | Borderline; buy if multi-year |
| 100% (always on) | ~$20,000 | Buy if 2-3+ yr horizon |
Snapshot June 2026 — rental prices change weekly; verify before modeling. Use the cost calculator to plug in your real hours.
The hidden costs of owning
The sticker price is the easy part. Owning adds:
- Power & cooling — an 8-GPU H100 server draws ~10kW; at $0.15/kWh that’s ~$13,000/year just for electricity.
- Hosting — rack space, colocation or a noisy, hot room.
- Networking & maintenance — switches, cabling, failed-part replacement.
- Depreciation — the killer. As rental prices fall (they dropped 60-75% year over year), your owned GPU’s resale value and rental-equivalent value drop with them.
When buying does pay off
- You run GPUs at high, steady utilization (60-70%+) for 2-3+ years.
- You have cheap power and somewhere to host them.
- Data residency or air-gap requirements rule out cloud.
For everyone else — bursty training, experimentation, inference that scales up and down — renting avoids the capital, the idle time and the depreciation risk. And the neoclouds make it cheap: 2-4x less than the hyperscalers.
Bottom line
Unless you can keep a GPU genuinely busy for years, rent. The falling-price trend means the break-even for owning keeps moving further out of reach.