Declassified in fullauthority: public dataFootage 0000 ftCopy 01 of 01
The price of one GPU-hourFile copy · not for publication
Identical hardware.Quoted the same day.Two times apart.
$9.80Google Cloud · on demand
1.8×
$17.57Azure · on demand
Basis studies what makes identical compute cost differently — and how much of that difference observable facts still cannot explain. Every figure traces back to the raw response a provider returned.
118 on demand quotes · H100 SXM 80GB · 5 providers · USD/GPU-hour · Sep 19 · 20:01 UTC
Commodities convergetoward a price.GPU-hours don’t.
An H100 SXM 80GB has the same silicon wherever you rent it.1 Region, commitment, and contract structure explain part of the spread — but once those observable differences are accounted for, prices ought to converge. They don’t. Not even close.
Nobody had published, on public data, with a method you can rerun, how much of that dispersion is actually explainable. So we started asking twice a day, and writing down every answer.
In commodity markets, basis names the gap between a local cash price and the relevant futures price — the part geography, timing, and contract terms cannot standardize away.2 Traders have a word for the gap between a reference price and the price realized in a particular market. We borrow it for compute.
If GPU compute is becoming a commodity, it appears to carry a basis of its own. In this study, Basis is the share of quoted price dispersion that remains after provider, region, commitment, and other observable terms are accounted for — not the raw spread between two quotes, but the part normalization still cannot explain.
Median quote · same SKU & collection dayObserved quotes only. Basis names what normalization cannot explain.
04 · The method
Twice a day. Four steps.Every answer kept.
Collect → file → canonicalize → account. The full procedure is documented separately.
H100 SXM 80GB · 2026-09-19 · Sep 19 · 20:01 UTC · 500 quotes · raw dispersion before controlsLog scale: equal horizontal distance is an equal price multiple. Hover, tap or arrow through a lane to read a quote.
Raw quoted dispersion · log scale · USD per GPU-hour · before controlsn = 500 of 638 recorded
p5 $9.80p95 $17.57
Vast.ai7
AWS15
RunPod6
Google Cloud376
Azure96
$1.00$3.00$10.00$20.00
We report the middle 90% (p5–p95), not the mean — the quote distribution is skewed, and outliers can distort an average without describing the market.
The spread is real. Region, commitment, provider, and bundle explain some of it. Next, we remove what we can explain.
06 · Exhibit C
The settlement sheet. One hundredunits of disagreement, filed againstwhat sellers disclose.
Sample populationmarket-priced · Azure and Google Cloud excluded
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.
FAX TRANSMISSION · 01/03Subject · observable boundBound, not victory
A richer model on the same days still falls short.
Forty-five features, day-based validation, and a leakage guard — scored on the same held-out days as the four-factor bound.
Four factors · same holdout days56.3%
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. Both bars use the same holdout window (10 days). The gap bounds what observables can do. It says nothing about what nobody publishes.
INDEX CARD · 02/03Subject · host identityIdentity, not specs
The remainder isn’t only noise. It persists by host.
Over half of what survives the subtraction tracks which host listed the offer, day after day. That persistence is inconsistent with a fully fungible market — at least within the factors and period studied.
Show the sensitivity
≥5d0.551
≥20d0.531
Published side by side, across every tenure threshold we tried. No threshold was chosen for flattery.
ATTACHMENT · 03/03
Subject · the moving shareA range, not a favorite
The number moves. We publish the range it moves in.
0%to23%
Unexplained share in market-priced segments across the last 30 days. A single figure would be a snapshot pretending to be a constant, so the file quotes both ends and dates them.
Select a sheet to read it in full
Limitation L1Quoted prices, not transactions. Negotiated and realized prices are not observed.
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.
