How much does it cost to rent an H100 per hour?

Between roughly $1.38 and $12.29 per GPU-hour on published on-demand rates, surveyed 2026-09-09 — a spread of about 9x for a physically identical accelerator. Marketplaces sit at the floor, hyperscaler list prices at the ceiling. Committed and negotiated rates are lower than every one of those figures and are not published by anyone.

Why is the same GPU so much cheaper at one provider than another?

Because you are not only buying the chip. The floor is brokered capacity on hardware somebody else owns, with variable interconnect quality and no guarantee about which machine you get. The ceiling is a hyperscaler list price that exists mainly to make its own committed rate look like a discount. Between them sit differences in support, networking, storage, region and contract terms — and in who the counterparty is.

Does Macrostack measure these prices itself?

No, and the page says so at the top. We do not rent clusters and we do not benchmark them. Every figure is a published list rate read from the cited source on the observation date. Inventing measurements would make this index worse than useless.

Layer 2.5 · price index · surveyed 2026-09-09

GPU rental price index

Name: Macrostack GPU Rental Price Index
Creator: Macrostack
License: https://creativecommons.org/licenses/by/4.0/

The same NVIDIA H100 SXM 80GB rents for $1.38 to $12.29 per GPU-hour — a 8.9x spread for a physically identical part. That gap, not any single rate, is the finding worth carrying away.

What this is, and what it is not

We did not measure these prices. Macrostack does not rent clusters and does not benchmark them. Every figure below is a published list rate, read from the source cited beside it on 2026-09-09. Where a range appears, the range is what the source reported.

Nobody serious pays these numbers. Committed, reserved and negotiated rates are materially lower than every figure on this page, and no provider publishes them. Treat this as the ceiling of what you should pay, not an estimate of what you will.

Any single rate here will be stale within a month. Rates at this layer move weekly. The spread is what survives, which is why the spread is the headline.

NVIDIA H100 SXM 80GB — $1.38 to $12.29 per GPU-hour

8.9x spread. Roughly a 9x spread for a physically identical accelerator. This is the single most important number on this page: the chip is not the variable, the counterparty is.

Where	USD / GPU-hour	Tier	Source
Marketplace and boutique floor — Brokered capacity. The machine you get is not a machine you chose, and interconnect quality is the thing that varies.	$1.38 - 1.49	on-demand, marketplace	intuitionlabs.ai
Specialist neocloud and marketplace on-demand — Where RunPod and Vast.ai sit for a single H100 SXM. Spot listings dip below this.	$2.00 - 3.00	on-demand	intuitionlabs.ai
Broad market median — Roughly half what the same rate was a year earlier — the clearest evidence that H100 capacity stopped being scarce.	$3.38	on-demand	shattered.io
Hyperscaler on-demand ceiling — Nobody signing a real contract pays this. It is a list price that exists to make the committed rate look like a discount.	$11.68 - 12.29	on-demand, hyperscaler list	intuitionlabs.ai

Specifications for this part: NVIDIA H100 SXM 80GB on Layer 2.

NVIDIA H200 SXM 141GB — $3.59 to $4.59 per GPU-hour

1.3x spread. The tightest spread of any part here, because H200 supply arrived into a market that had already learned what H100 was worth.

Where	USD / GPU-hour	Tier	Source
RunPod community cloud — Held at roughly this rate through most of 2026. Community capacity is other people's hardware.	$3.59	community	gpusmith.com
RunPod secure cloud — The same chip in a datacentre the provider controls. The delta is the price of knowing where your data physically is.	$4.39 - 4.59	secure	gpusmith.com

Specifications for this part: NVIDIA H200 SXM 141GB on Layer 2.

NVIDIA B200 — $3.44 to $16.11 per GPU-hour

4.7x spread. A 4.7x spread, and the widest in absolute dollars. Blackwell capacity is still being allocated rather than sold, which is what a wide spread looks like from the outside.

Where	USD / GPU-hour	Tier	Source
Vast.ai — The floor of the surveyed market.	$3.44	marketplace	intuitionlabs.ai
Lambda	$4.99 - 5.29	on-demand	intuitionlabs.ai
RunPod	$5.89	on-demand	intuitionlabs.ai
Median across surveyed providers	$6.11	on-demand	intuitionlabs.ai
Google Cloud — The ceiling of the surveyed market, and 4.7x the floor for the same part.	$16.11	on-demand, hyperscaler list	intuitionlabs.ai

Specifications for this part: NVIDIA B200 on Layer 2.

What the numbers say

Neoclouds and marketplaces price 40% to 400% below the three major hyperscalers for the same accelerator. (www.spheron.network)
H100 on-demand rates roughly halved during 2026, while Blackwell-class parts held a wide spread — the signature of a market where the previous generation has become a commodity and the current one is still allocated. (shattered.io)

Use this data

Published under CC-BY-4.0. Commercial use, redistribution and modification are all permitted. The single condition is attribution: credit Macrostack with a link to macrostack.net.

Machine-readable: /compute/gpu-prices/index.json — the same data, with the source and observation date on every row.

If you maintain a pricing page and think a figure here is wrong, it probably is: tell us and we will correct it with your source. Being correctable is the only thing that makes an index worth citing.

Sources

For hands-on performance ratings of the providers behind these prices — security, orchestration, storage, networking, reliability — SemiAnalysis runs ClusterMAX, which tests the clusters directly. This index covers price; our compute layer covers counterparty risk and exit. Neither of those is a benchmark and we do not present them as one.

Where this sits

Layer 2.5 — Compute · Layer 2 — Silicon · Layer 1 — Energy · The five-layer map