The price of one GPU-hourFile copy · not for publication

Identical hardware.Quoted the same day.Thirteen times apart.

$1.15Google Cloud · spot
13.0×
$14.99Azure · reserved 1y

Basis is a public study of the part of that price nothing observable explains. Every figure here walks back to the raw response a provider returned.

p5 and p95 of 373 quotes · h100_sxm_80gb · Aug 5 · 08:04 UTC

01 · The brief

A commodity is a thingthat trades at one price.

A GPU-hour should be one of them. Wheat is wheat. Oil is oil. An H100 is an H100, and the silicon is identical.1 So identical things should cost roughly the same. They don’t. Not even close.

Nobody had published, on public data, with a method you can rerun, how much of that price is actually explainable. So we started asking twice a day, and writing down every answer.

Paragraph 4 · struck on receipt · lift it

Every figure in this file walks back to a provider returned, and none of it came from a paid feed.

  • A study, not a product
  • No paywalled feeds
  • Every figure has a receipt
02 · Exhibit A

Two quotes for the sameeighty gigabytes of silicon.

h100_sxm_80gbGoogle Cloud · AU · spot
$1.16/hr
h100_sxm_80gbAzure · ZA · reserved 1y
$14.93/hr

Two sellers, one canonical chip, the same collection day. Nothing about the hardware accounts for the distance between them.

03 · The name

The gap has a name.

When a trader hedges oil with a benchmark, the gap between the benchmark and the price they actually pay is called basis.2 It’s the risk standardization can’t remove. GPUs, the commodity compute is supposed to have become, carry that gap too. This project is named after it, and it measures it.

BENCHMARK PRICEWHAT YOU ACTUALLY PAY
Recorded, never modeled.
04 · The method

Twice a day, we ask five clouds the same question.

Four frames. Keep scrolling.Swipe the strip.

01Collect

At 08:00 and 20:00 UTC we ask five public catalogs the same question: what does an hour of this GPU cost right now?

  • AWS Spot
  • Azure
  • Google Cloud
  • RunPod
  • Vast.ai

No private data and no paywalled feeds: only prices anyone could see.

02File

Every answer is written down exactly as received, in full, and never edited again.

Raw is sacred. Everything derived from it is disposable, and can be rebuilt from the file.

610,652answers on file
03Canonicalise

Every cloud describes the same chip differently.

"H100 SXM""H100 SXM5 80GB"h100_sxm_80gb

A name that matches no rule is set aside, never guessed. Missing information stays UNKNOWN and stays visible.

04Subtract

Then we subtract, one factor at a time, every reason the prices ought to differ.

Where it is. How it’s rented. Who sells it. What comes with it. What survives the subtraction is the finding.

05 · Exhibit B

One day of quotes,one canonical chip.

2026-08-05 · 5 catalogsHover, tap or arrow through a lane to read a quote.

Quoted price · log scale · USD per GPU-hourn = 373
Vast.ai60
AWS Spot15
RunPod6
Google Cloud196
Azure96

Percentiles, never a mean. One mispriced listing would drag an average somewhere no buyer could actually transact.

Quote slipday summary
$5.59/hr
Provider
median of 373
Region
all countries
Commitment
all types
Recorded
2026-08-05

Pick a quote to read the one behind it.

Pull the raw observation
06 · Exhibit C

The settlement sheet.

One hundred units of disagreement, filed against everything sellers disclose.

h100_sxm_80gb · 2026-08-05 · sequential ANOVA on log pricemarket-priced · Azure and Google Cloud excluded
Share of price variation · h100_sxm_80gbLive
The price we can't explain, before accounting100.0%
−where it is · REGION60.2 left
−how it's rented · COMMITMENT38.1 left
−who sells it · PROVIDER21.4 left
−what's bundled · BUNDLE17.7 left
=Unexplained17.7%
17.7 of every 100 · unaccounted

Where the machine is. How it’s rented. Who sells it. What comes bundled with it. Everything observable, accounted for, and still a share of the price has no explanation.

Change the order of the factors and the four credits move. The remainder does not. That is why the remainder is the headline. In market-priced segments it has ranged 13% to 61% across the last 31 days, so basis risk is segment- and time-conditional.

07 · Findings of record

Three sheets that survived review.

Subject · observable boundBound, not victory

We built a richer model to kill the finding. It did worse.

Forty-five features, day-based validation, and a leakage guard that fails the run.

Four factors · in-sample56.3%
45 features · out-of-sample45.4%
Δ −10.9ppas of Jul 31
Show the method

Splits fall on ordered days, never rows. The final 10 days never enter selection. Scoring is day-demeaned, so the model gets no credit for knowing roughly what an H100 costs this month. A permuted-target holdout above 0.05 kills the run; this one scored -0.19. The gap bounds what observables can do. It says nothing about what nobody publishes.

Subject · host identityIdentity, not specs

The leftover is not noise. It sticks to specific machines.

0.55Intraclass correlation · 61 hosts · 10-day tenure

Over half of what survives the subtraction tracks who the host is, day after day. In a market of independent operators, price is substantially a function of identity. That is very nearly the opposite of a commodity.

Show the sensitivity
  • ≥5d0.551
  • ≥20d0.531

Published side by side, across every tenure threshold we tried. No threshold was chosen for flattery.

Subject · the moving shareA range, not a favourite

The number moves. We publish the range it moves in.

13%61%

Unexplained share in market-priced segments across the last 31 days. A single figure would be a snapshot pretending to be a constant, so the file quotes both ends and dates them.

  • Limitation L1Quoted prices, not transactions. This is a lower bound.
  • Limitation L2A marketplace and a hyperscaler are not like-for-like populations.

Somebody is about to writethe rules for how computegets priced.

You can’t build financial plumbing on a price you can’t explain. There is growing interest in treating AI compute like a commodity, with indexes, futures and contracts on top of it.3 All of that assumes a GPU-hour has a knowable market price. That unexplained remainder is the risk any benchmark would silently absorb.

  • It’s a study, not a product

    Nothing for sale, no paid feed as a required input. Public quotes and a method you can rerun.

  • Every number has a receipt

    Headline share, contributing offers, raw response, exact rules applied. Four clicks, no exceptions.

  • Honest about limits

    One collection outage found, published, root-caused and turned into a standing alarm rather than smoothed away.

The file is open

It is not asking to be believed. It is asking to be checked.

Open the dashboardRead the methodology →

I was just bored and curious. So here it is:

— Raj

Sources
  1. Hardware identity within a canonical SKU: normalization rules and variant separation — Basis methodology §3 (canonical schema).
  2. Commodity basis, standard definition — CME Group education materials on basis and hedging.
  3. Compute-as-commodity framing — Ornn AI public materials; the essays referenced in the original Basis proposal.

The GPU spread itself is our own live data. See the dispersion page.