On this page

At a glance Spread Waiting on How counted FAQ

InfratoApp / GPU price index

GPU price index

These are posted rates — what providers publish, not what trades clear at. A posted-rate spread is still the most useful signal we can measure honestly today.

The widest spread today is 3.8× on the L40S-48GB — $0.60 at Vast.ai against $2.25 at CoreWeave. The typical GPU here spreads 2.3× between its cheapest and dearest posted rate, across 13 cards sold by three or more providers.

WIDEST SPREAD 3.8× L40S-48GBTYPICAL 2.3×GPUS TRACKED 13CAPTURE STATUS ✓ ON-DEMAND LIVE · ⚑ RESERVED PARTIAL

The market at a glance

13 days of captures · median of medians

Each provider is counted once per GPU, so listing one card at eight instance sizes buys one vote, not eight. A model counts toward a change only if it was priced on both dates.

3.8×

Widest spread · L40S-48GB

2.3×

Typical spread · 13 cards

○

Past 4 weeks

needs two comparable weeks; we hold days

○

Past year

needs a year of captures

Hyperscalers vs GPU clouds$/GPU·HR

Hyperscalers$2.95 · 1 seller
GPU clouds$3.20 · 11 sellers
Serverless GPU$3.99 · 1 seller

Serverless GPU post ~35% more than Hyperscalers for an H100-SXM-80GB — a median of $3.99 against $2.95 per hour.

What a commitment buysVS THE SAME PROVIDER’S OWN ON-DEMAND

SPOT64%

Spot lists at 64% of on-demand across 30 comparable models — the discount pays for interruption risk. That is a sample size, not a history.

Spread over time — H200-SXM-141GB, per day read

Low · median · high across providers, per day · hover a column for its numbers

2026-09-10BAR SPANS CHEAPEST TO DEAREST · HAIRLINE IS THE MEDIAN2026-10-27

H200-SXM-141GB across 42 days read. Widest on 2026-10-16, when 23 providers posted rates 3.6× apart. Each column is that day's spread, not a trend.

Today’s spread, all GPUs

One rate per seller, their cheapest · open a row for every seller behind it

Cross-vendor on-demand price spread per GPU model
GPU	CHEAPEST	DEAREST	SPREAD	READ
L40S-48GB	$0.60 Vast.ai	$2.25 CoreWeave	3.8×	11 Oct

L40-48GB	$0.46 Vast.ai	$1.25 CoreWeave	2.7×	26 Oct

B200-SXM-192GB	$3.74 Shadeform	$9.86 Vast.ai	2.6×	26 Oct

H100-SXM-80GB	$2.40 Vast.ai	$6.16 CoreWeave	2.6×	26 Oct

A100	$1.09 Civo	$2.70 CoreWeave	2.5×	26 Oct

RTX-6000-ADA-48GB	$0.67 Vast.ai	$1.57 DigitalOcean	2.3×	11 Oct

RTX-4090-24GB	$0.39 Packet·ai	$0.89 RunPod	2.3×	26 Oct

H100-80GB	$2.25 TensorDock	$4.69 Lyceum	2.1×	11 Oct

A100-SXM-80GB	$1.07 Vast.ai	$2.18 Verda (DataCrunch)	2.0×	27 Oct

RTX-PRO-6000-96GB	$1.40 Vast.ai	$2.57 Verda (DataCrunch)	1.8×	27 Oct

H200-SXM-141GB	$3.49 Civo	$6.30 CoreWeave	1.8×	11 Oct

B300-SXM-288GB	$7.50 Enverge	$11.09 Lyceum	1.5×	26 Oct

H100-PCIE-80GB	$2.54 Vast.ai	$3.50 Oblivus	1.4×	26 Oct

Waiting on cleared-trade data

Each block names what it needs and what we hold today. A progress bar, not a promise.

WAITING

Index, equal-weight, rebased to 100

An index needs twelve weekly captures. We have 13 days, first captured 10 Sept — so this block states that instead of drawing a line through them.

NEEDS 12 weekly capturesHAVE 13 days · 13 captures

FIRST VERSION64%

Spot as a share of on-demand

One number, not a series: across the 30 SKUs where a vendor publishes both, spot sits at 64% of on-demand today. That is a sample size, not a history length — a handful of captures cannot make a trend.

NEEDS paired spot + list pricesHAVE 30 SKUs · 13 captures

WAITING

Reservation term structure

14 rows. A one-year against three-year curve drawn from that would assert a shape neither observation supports — so the block names what it holds and waits.

NEEDS 1y and 3y prices at volumeHAVE 14 rows

WAITING

Weekly top movers

Needs two comparable weeks and we have days. This is the block that will be most read once it works, which is exactly why it must not be faked first.

NEEDS two comparable weeksHAVE 13 days · not yet

How this is counted — and what this is not, yet

Counted: posted on-demand rates, normalized to $/GPU·hr, one per seller per card, stale readings (>21d) excluded.

Not yet: a cleared-trade index. Nobody outside the exchanges knows what H100 hours actually trade at; we won’t pretend otherwise.

Not ever: a composite “market price” blending terms, regions and SKUs into one number.

Per-GPU price tables →How we read this data →

Frequently asked

How much does the price of the same GPU vary between providers?

Substantially. On the cards rented by at least three providers here, the dearest seller commonly charges between 1.2 and 3 times the cheapest for the identical part, on-demand and before any commitment discount.

Why is the same GPU priced so differently?

Region, network and storage included in the rate, support, contract terms, and how recently the seller last looked at a competitor. The card is a commodity; everything wrapped around it is not, and pricing has not converged.

Is this a GPU price index?

Not yet. An index needs a long enough series that a movement means something, and this holds a handful of days. What is published is the cross-vendor spread, which is a fact about today and needs no history.

Are these the prices people actually pay?

No, and we say so at the top of the page. These are posted rates — what a provider publishes on its own pricing page. Negotiated and committed deals are struck privately and nobody outside the exchanges knows what GPU hours clear at, so a cleared-trade index is not something we can honestly publish.

How often are these prices re-read?

Continuously, and each row carries the date its own figure was read. A reading older than 21 days is hidden rather than shown stale, so a card with few sellers is a card we are reading honestly rather than one we have padded.