*[DediGPU](https://dedigpu.com/) — Markdown mirror of [https://dedigpu.com/guides/renting-vs-buying-an-h100](https://dedigpu.com/guides/renting-vs-buying-an-h100) · updated 2026-09-02 · index for LLMs: [llms.txt](https://dedigpu.com/llms.txt) · everything: [llms-full.txt](https://dedigpu.com/llms-full.txt)*

# Renting vs buying an H100: the 36-month math, line by line

> What an H100 costs to own and run over three years, from the same cost sheet that prices our servers, and the break-even against renting at cost.

*Economics · published 26 August 2026 · updated 2 September 2026 · by DediGPU engineering*

In short:

- Over 36 months, renting an H100 at cost totals $40,824 per card; owning is the same six lines plus cash up front, retail power, colocation and idle time.
- Owning wins only at our purchase price, $0.05 power, a full rack and 100 % utilisation for three years; then you keep the residual value.
- The cloud sits between the two: about twice our cost sheet for a month of H100.

People compare renting with buying by looking at the price of the card. That is one of six lines. Because our price is the cost sheet itself, we can show the whole comparison with our own numbers, and it is honest in both directions: the same lines that make renting cost $1,134 a month make owning cost about the same, if you can run the machine as well as a data centre does.

## What one H100 costs per month, owned or rented

| Line | Assumption | Per month | Over 36 months |
| --- | --- | --- | --- |
| **Hardware** | $27,000 card + $8,750 chassis share, over 36 months | $993.06 | $35,750 |
| **Power** | 529 kWh at $0.052 · 900 W, 70% load, PUE 1.15 | $27.50 | $990 |
| **Rack & cooling** | HGX baseboard, liquid-assisted | $45.00 | $1,620 |
| **Network & IP** | transit share, one IPv4, port | $9.00 | $324 |
| **Spares & failures** | 2% of the card price per year | $45.00 | $1,620 |
| **Operations** | monitoring, on-call, provisioning | $14.00 | $504 |
| **Total** | rounded up to the dollar | $1,133.56 → $1,134 | $40,824 |

The hardware line is the card at $27,000 plus its share of an HGX node ($8,750 of CPUs, RAM, NVMe, NICs and baseboard) amortised straight-line over 36 months. That share is not optional: an SXM card only works on that baseboard.

## What owning adds that the sheet does not show

- **Cash up front.** $35,750 per card, $286,000 for a node, paid before the first token. Renting spreads it at zero interest.
- **Power at retail.** Our sheet uses $0.052 per kWh on long Nordic contracts. Office or colocation power in Western Europe or the US is two to four times that; the power line alone becomes $60 to $110 a month per card.
- **Colocation and cooling.** A rack with 40 kW of cooling for an eight-H100 node is a specialised product; expect $1,500 to $3,000 a month per rack, against our $45 per card.
- **Failures.** Our spares line is 2 % of the card price per year, spread across a fleet. With one node, a failed card is a support case with the vendor and weeks of downtime, not a swap from the shelf.
- **Utilisation.** A rented server can be cancelled the month the project ends. An owned one keeps depreciating. At 50 % utilisation the effective cost of ownership doubles.
- **Residual value.** The one thing in owning's favour: after 36 months the card is worth something, perhaps 20 to 30 % of its price for a Hopper card in 2029. Our sheet amortises to zero, which is why our price is a cost, not a bargain.

## The break-even

Over 36 months, renting at cost totals $40,824 per card. Owning totals the same hardware, plus power, rack, spares and operations at whatever you can negotiate. If you can buy at our purchase price, get power at $0.05 per kWh, colocate at $45 per card per month and keep the card busy 100 % of the time for three years, you break even, and you keep the residual value. At any smaller scale, at any lower utilisation, or with any commitment shorter than three years, renting at cost is cheaper.

## Where the cloud sits

Between the two. The median on-demand H100 in 2 September 2026 was about $2,460 a month at 730 hours, twice our cost sheet, because the cloud price carries a margin and the cost of idle capacity. Clouds win for bursts of hours; renting wins for months; owning wins only at scale and full utilisation. The [cloud comparison](https://dedigpu.com/guides/dedicated-gpu-server-vs-gpu-cloud) goes through the hidden lines.

## A note on the PCIe H100

The PCIe card costs $23,000 and needs only a $2,500 chassis share, so its sheet lands at $810 a month. For single-card inference and fine-tuning it is the same silicon with less bandwidth and no NVLink, at 29 % less. The same math applies to buying one.