# B200 reference price

> The GPUQuant B200 reference price: the on-demand list price of one B200 for one hour on the eight-GPU node, averaged across the hyperscalers that publish one, monthly since the first listing.

Updated 2026-09-06. Canonical: https://gpuquant.com/docs/market-data/b200-reference-price

## What it measures

One NVIDIA B200 180GB, rented for one hour on the eight-GPU HGX node, at the on-demand list price for Linux, in US dollars, from the hyperscalers that publish one: AWS p6-b200 and Google Cloud a4-highgpu-8g, with Oracle’s uniform per-GPU price beside them from August 2026.

Azure’s GB200 sizes are excluded, not because they are not Blackwell but because Microsoft states no accelerator count GPUQuant could verify for them, and a figure that rests on a guessed count would not be the same kind of number as the others. The exclusion is logged and shown on the coverage page.

The B200 is the newest part on the site, and the series starts where the first hyperscaler listed it. Google Cloud’s A4 family publishes a bundled per-GPU slice rather than separate machine SKUs, so it is already a complete per-GPU figure and is not rebuilt the way the A2 and A3 families are.

Oracle’s arrival in August 2026 moved the figure from $16.39 to $15.59; the maths did not change, the inputs did.

## How it is built

In two steps, every month. Per region, GPUQuant takes the lowest eligible published price for the eight-GPU node and divides it by eight. Per provider, it averages that provider's regional prices, so a provider that publishes the B200 in forty regions counts no more than one that publishes it in six. The reference is the average of the provider figures, and the median is computed and stored beside it on every row.

A provider with no eligible offering is absent, never zero. A provider that has not republished holds its last price forward and is marked held. Google Cloud prices the accelerator separately from the machine, so its complete machine price is rebuilt from the GPU, core and memory SKUs before dividing; Oracle prices per GPU-hour for a whole shape and is used as published. The arithmetic is on [The reference price](https://gpuquant.com/docs/reference-price) and [Normalization](https://gpuquant.com/docs/normalization), and the calculation version is **v4**.

## Publication

**Source**

GPUQuant

**Unit**

USD / GPU-hour

**Frequency**

Monthly: the price lists are read daily, and a month’s figure is the last reading taken in it

**Ticker**

None; a GPUQuant figure

## What GPUQuant shows

The whole history, from the first month GPUQuant priced the part, with every provider's own regional average as a row under the reference, the provider comparison and every regional offering behind it under the chart, and the change log of every list-price move dated to the provider's own effective date. Nothing is licensed and nothing is windowed: every figure has a URL behind it.

On [Market Data](https://gpuquant.com/market-data) as GPUQuant Reference under B200.

## Reading it beside the others

-   [B200 neocloud index](https://gpuquant.com/docs/market-data/b200-neocloud-index): The list price of the same hour on the specialist clouds, roughly a third of the hyperscaler figure.
-   [Silicon Data B200 index](https://gpuquant.com/docs/market-data/silicon-data-b200-index): Silicon Data's modelled figure, which counts negotiated and committed-use deals and sits below any list price.
-   [Ornn B200 index](https://gpuquant.com/docs/market-data/ornn-b200-index): Ornn's transacted figure, the prices B200 rentals actually cleared at.
-   [B200 forward curve](https://gpuquant.com/docs/market-data/b200-forward-curve): What the market expects the B200 rental index to be out to three years.
-   [B200 residual value](https://gpuquant.com/docs/market-data/b200-residual-value): What one B200 is still worth, from that curve.

## Sources

-   [Data sources](https://gpuquant.com/docs/data-sources), the provider price lists every figure is read from.
-   [Coverage](https://gpuquant.com/docs/coverage), which provider publishes the part where, and every exclusion with its reason.
-   [Time and coverage](https://gpuquant.com/docs/history), how a held price and a missing provider are shown.