Evaluate financial capital recovery horizon, monthly cloud rental expenditure vs on-premise hardware purchase and colocation upkeep for NVIDIA H100 80GB SXM5 Node.
With the defaults ($320,000 purchase, $24/hour rent, 24 hours a day, $1,800 a month running cost), payback is 20.4 months. At 12 hours a day it stretches to 46 months.
It is an illustrative input equal to $3 per GPU-hour for eight GPUs. Prices vary widely between providers and over time, so enter a current quote.
Staff time, financing, resale value, networking and storage, and reserved-instance discounts. Use the result as a first filter before a full cost analysis.
This calculator asks a single question: if you need an 8-GPU H100 server, how long does it take for buying one to cost less than renting the equivalent from a cloud provider? The defaults describe an 8 × H100 80 GB SXM5 node (the HGX H100 class of server) bought for $320,000, rented for $24 per hour (equivalent to $3 per GPU-hour), used 24 hours a day, with $1,800 a month for colocation space, power and maintenance. These are illustrative inputs, not quotes. H100 server prices and cloud rates have changed a lot since the GPU launched and differ widely between providers, so replace them with prices you have actually been offered.
If payback is longer than 36 months the calculator recommends renting, on the reasoning that accelerator hardware is usually written off over three to five years and a newer generation will have arrived by then.
| Line | Working | Result |
|---|---|---|
| Cloud rent per month | $24 × 24 h × 30.4 days | $17,510 |
| Owning: running cost per month | input | $1,800 |
| Monthly saving from owning | 17,510.4 − 1,800 | $15,710 |
| Payback | 320,000 ÷ 15,710.4 | 20.4 months |
| Net after 24 months | 24 × 15,710.4 − 320,000 | $57,050 in favour of owning |
| Net after 36 months (same formula) | 36 × 15,710.4 − 320,000 | about $245,600 |
The figure that most changes the result is hours of real use per day. The cloud bill falls when you stop renting, but the purchase price does not:
| Scenario | Payback | 24-month net |
|---|---|---|
| 24 h/day at $24/h (default) | 20.4 months | +$57,050 |
| 12 h/day at $24/h | 46.0 months (calculator recommends cloud) | −$153,075 |
| 8 h/day at $24/h | 79.3 months | −$223,117 |
| 24 h/day at $16/h | 32.4 months | −$83,034 |
| 24 h/day, $4,000/month running cost | 23.7 months | +$4,250 |
Buying pays off for steady, round-the-clock work such as a production inference service or a long training programme. For bursty research or occasional fine-tuning, renting is usually cheaper even at on-demand rates, and reserved or committed-use cloud pricing pushes the break-even point further out.
Check the running-cost input carefully. A node that can draw about 10.2 kW uses roughly 7,450 kWh a month at full load. At $0.12/kWh that is about $894 of electricity, or about $1,117 once a facility PUE of 1.25 is included, before rack space, cross-connects, bandwidth, support contracts and spare parts. For a fully loaded node, $1,800 is optimistic in many markets; try $3,000–$4,000 to see how sensitive your decision is. The site's data-centre PUE calculator works out the electricity part in detail.
Use the payback figure as a first filter. If it is under about 18 months at honest utilisation, a full total-cost-of-ownership analysis is worth doing; if it is over 36 months, renting is almost certainly the better choice.