NVIDIA H100 rental price index

What it costs to rent one NVIDIA H100 for an hour, in US dollars, as the Racks H100 index.

Updated Nov 10, 2026, 10:45 UTC

Price

$2.94per GPU-hour

Range

2.852.902.953.003.05Oct 15Oct 22Oct 29Nov 5

Figure 1. Index level in US dollars per GPU-hour, past month, Oct 12 to Nov 10, 2026. The ring marks the latest value.

H100 rental price in dollars per GPU-hour, past month, Oct 12, 2026 to Nov 10, 2026: low $2.83, high $3.07, latest $2.94.

**Table 1.** Key figures of the Racks H100 index.
Parameter	Value	Unit
Price	$2.94	per GPU-hour
Change since inception^[3]	+11.29%	+$0.30
Past 24 hours	−0.43%	change
Past 7 days	−1.83%	change
Low since inception^[3]	$2.64	per GPU-hour
High since inception^[3]	$3.07	per GPU-hour
Realized volatility	44%	a year, past 30 days
B200 for each H100	2.31	times the price
Sources counted	4 of 4	now
Racks H100 Index^[4]	105.18	base 100

1Cost to run

At $2.94 an hour for each GPU, a month being 730 hours:

**Table 2.** What running GPUs costs at the H100 index price, all day.
Size	A day	A month
8 GPUs, one node	$564	$17,170
1,000 GPUs	$70,560	$2.15 million

H100$2.97

B200$3.02

Figure 2. Price of one petaFLOP-hour of dense BF16 compute at each chip’s peak, at today’s prices.^{[1, 2]}

An H100 costs 2% less than a B200 per unit of peak compute. Real work reaches a fraction of peak, and the fraction differs by job.

2Method

The Racks H100 index follows what an NVIDIA H100 rents for by the hour. It counts four published rates: Compute Desk’s US 1-year desk rate, Liquid Compute’s US reference, Silicon Data’s rental index for neo-clouds and Ornn’s OCPI of on-demand rentals. Each is compared with where it stood when the period opened.

The index moves by the geometric mean of the four changes: multiply them and take the fourth root. Each change counts in full, and the index needs two live sources, or it holds its level.

Index = 100 \times \prod_{i = 1}^{4} {(\frac{p_{i}}{p_{i, 0}})}^{w_{i}}

(1)

Equation 1. The index. Here p_i is a source’s counted price now, p_i,0 its counted price when the period opened, and w_i its weight. The weights add up to 1, and each is one quarter while all four count. The Racks H100 index follows the geometric mean of four published H100 rental rates: Compute Desk, Liquid Compute, Silicon Data and Ornn’s OCPI.

**Table 3.** The four changes this week and the index they make.
Source	Ratio to the open	Change
Compute Desk	1.0181	+1.81%
Liquid Compute	1.0140	+1.40%
Silicon Data	1.0033	+0.33%
Ornn	0.9529	−4.71%
$100 \times {(1.0181 \times 1.0140 \times 1.0033 \times 0.9529)}^{\frac{1}{4}}$
Index		99.67

The price in $/hour is the geometric mean of the four prices when the period opened, moved by the index: while all four count, the geometric mean of the four prices now.

The rules are in the docs.

3Questions

What does an H100 cost to rent per hour?

The H100 rental price index stands at $2.94 per GPU-hour as of Nov 10, 2026, 10:45 UTC. It is an index of four published rental rates, so what one provider charges can differ.

What is the H100 rental price index?

Racks’ index of what an NVIDIA H100 rents for by the hour. It counts four published rates: Compute Desk’s US 1-year desk rate, Liquid Compute’s US reference, Silicon Data’s rental index for neo-clouds and Ornn’s OCPI of on-demand rentals. Every period starts at 100, and the index shows how far prices have moved since, in percent. The price in $/hour is the geometric mean of the four prices when the period opened, moved by the index: while all four count, the geometric mean of the four prices now.

Can I check the number?

Yes. The Method section of this page shows how the index is built from the four published rates, and the docs give the method in full.

How can I trade it?

Trading on Racks is not open yet. Join the waitlist.

4Notes

NVIDIA H100 datasheet: BF16 Tensor Core 1,979 teraFLOPS for the H100 SXM, with sparsity. Dense is half, 989.5 teraFLOPS. nvidia.com/en-us/data-center/h100
NVIDIA HGX platform: HGX B200 FP16/BF16 Tensor Core 36 petaFLOPS with sparsity across 8 GPUs, and “dense is half”. That is 18 petaFLOPS dense, 2.25 for each GPU. nvidia.com/en-us/data-center/hgx
Before Oct 12 the history is back-calculated: the index’s rule applied afterwards to stored prices, not the levels posted at the time. The figures marked with this note include it.
The Racks H100 Index chains each week’s index into one long series, from 100 at the open of the week of Oct 5, 2026. The docs give the rule.