7. Futures, commodities and energy

GPU rental prices (GPU)

GPU collects published hourly rental prices for AI chips by model and price basis, with medians and history, so compute can be tracked like a commodity.

  • 4 min
  • 3 questions
  • Lesson 6 of 7

Why you would care

Is AI compute getting cheaper or scarcer? The cost of an hour on a high-end AI chip drives the economics of every AI company, cloud provider and chip maker. When rental prices fall, chips are plentiful; when they rise or offers disappear, compute is scarce. Investors now watch GPU-hour prices the way they watch oil or freight rates.

The idea from scratch

A GPU (graphics processing unit) is the chip that trains and runs AI models. Companies rent them by the hour from:

  • Hyperscalers

    the largest cloud providers, with published list prices.

  • Neoclouds

    newer GPU-focused cloud companies.

  • Marketplaces

    platforms where owners post offers (asks) for spare capacity.

The same chip has several prices, depending on the basis:

List price
The published on-demand rate per GPU-hour
Provider-declared spot
A cheaper rate for capacity that can be taken back at short notice
Marketplace ask
Offers posted on marketplaces, which move with supply

Prices are compared per GPU-hour (one chip for one hour), and the board shows a median across providers so one outlier does not dominate.

These are posted prices, not transaction prices: big customers negotiate discounts that are not public.

Diagram: Hyperscaler rate cards leads to Board by GPU model and basis; Neocloud prices leads to Board by GPU model and basis; Marketplace asks leads to Board by GPU model and basis; Board by GPU model and basis leads to Medians, 1D/7D/30D changes; Board by GPU model and basis leads to History and changes; Board by GPU model and basis leads to Related stocks: chip makers, clouds, builders.
Wide diagram: scroll sideways to see all of it.

See it in Gloom

Gloom screenshot: GPU, Board tab, free preview
GPU, Board tab, free preview: A10 list prices with the hyperscaler median and a provider-declared spot price, per GPU-hour. Provider names blurred here. Captured 2026-10-06.

open ittype GPU, or GPU H100 for one model. Tabs: Board, History, Changes, Equities.

  1. A10 List price (2): the A10 chip's list prices, from two quotes.
  2. Hyperscaler median PCIe 24GB 2.00 0.0%: the median list price among the largest clouds, 2.00 dollars per GPU-hour, unchanged on the day.
  3. Chips PCIe and 24GB: the form factor and memory of the variant.
  4. A10 Provider-declared spot (1) 2.19 +0.1%: a spot rate, here above the list median (spot is not always cheaper).
  5. Footer: as of Oct 6, 2026 11:54 UTC Free preview. Pro shows every model (H100, B200...) and provider, with full history.

Practice and recap

Try it3 tasks
  1. 500 GPUs for 10 days at 2.00 dollars an hour: cost? (500 x 240 h x 2.00 = 240,000 dollars.)
  2. Why use a median rather than an average across providers? (One outlier price would move an average a lot.)
  3. Name one reason posted prices can differ from what big customers pay. (Negotiated discounts and long-term contracts.)
Common mistakes4 mistakes
  • Comparing different GPU models or memory sizes as if they were the same.
  • Mixing list prices with marketplace asks.
  • Reading posted prices as transaction prices.
  • Ignoring availability: a cheap price with no capacity is not a real offer.
Check yourself3 questions
  1. What is a GPU-hour?
  2. Name the three price bases.
  3. What does a falling marketplace median suggest?
Answers
  1. One GPU used for one hour: the unit for comparing rental prices.
  2. List price, provider-declared spot, marketplace ask.
  3. More supply of that chip, or weaker demand.
Words in this lesson6 words
GPU
The chip used to train and run AI models.
GPU-hour
One GPU rented for one hour.
hyperscaler
One of the largest cloud providers.
neocloud
A newer, GPU-focused cloud company.
list price / spot / marketplace ask
Published rate / interruptible discounted rate / posted offer.
median
The middle value of a sorted list.

Educational material about reading market data, not investment advice.