Same GPU, up to 28× the price: the provider spread on 2026-10-10

Most GPU cost discussions start with the chip: H100 or A100, B200 or L40S. The data says the provider can matter just as much. Rent the same card as 1 GPU on-demand, and on 2026-10-10 the gap between the cheapest and most expensive tracked price ranged from 1.3× on the B200 to 28× on the V100.

This article reports those gaps. It does not explain them, and it compares nothing beyond the hourly rate. A marketplace listing and a provider's own instance are not necessarily the same product. That difference has its own section below.

The spread table, 2026-10-10

For each chip: the cheapest and most expensive tracked price for 1 GPU on-demand, the multiplier between them, and how many providers we track for that card. Prices as of 2026-10-10.

The same multipliers as bars, on one common scale (bar length proportional to the spread):

Spread, cheapest vs most expensive (1 GPU, on-demand, 2026-10-10)

V100 16GB       ████████████████████████████  28×
A100 SXM 40GB   █████▌                        5.5×
A10 22GB        █████▎                        5.3×
RTX A4000 16GB  █████                         5.1×
RTX A5000 24GB  █████                         5.1×
RTX A6000 48GB  ███▊                          3.8×
L40S 45GB       ███                           3×
B200 192GB      █▎                            1.3×

Two caveats before reading anything into it:

• Some rows rest on only 2 providers (A10, A5000). There, the spread is just the gap between two quotes.
• The multiplier only describes the two ends. The providers in between are on each chip's page.

The rows are ordered by spread, not by chip generation. This sample is too small to support a trend by age, so the sections below group chips by what their rows contain, not by a story about them.

V100, A100 and A10 spreads

The V100 is the extreme case. In the chart below, bar length is proportional to the hourly price, on a common scale within the chart.

V100 16GB, 1 GPU on-demand, 2026-10-10

Vast.ai (marketplace)  █                             $0.082/h
RunPod                 ██▎                           $0.19/h
Verda                  ██▌                           $0.206/h
Lambda                 █████████▌                    $0.79/h
Paperspace             ████████████████████████████  $2.30/h

The cheapest is 96% less than the most expensive. The gap is not one outlier against a cluster: the prices form a low group (Vast.ai, RunPod, Verda), then Lambda, then Paperspace on its own.

The A100 SXM 40GB shows a 5.5× spread (82% less) across 6 providers:

A100 SXM 40GB, 1 GPU on-demand, 2026-10-10

Vast.ai (marketplace)  █████                         $0.362/h
Paperspace             ████████████████▎             $1.15/h
Hyperstack             ███████████████████           $1.35/h
RunPod                 █████████████████████████▎    $1.79/h
Verda                  ██████████████████████████▌   $1.89/h
Lambda                 ████████████████████████████  $1.99/h

On this chip, Paperspace is the cheapest offer after the Vast.ai listing. On the V100, on the same date, it is the most expensive. A provider's rank is not stable across chips, so check per card rather than assume a provider is "the cheap one".

The A10 22GB has a 5.3× spread (81% less), but only between two quotes, both in the table above.

RTX A4000, A5000 and A6000 spreads

On the V100, A100 and A10, the cheapest price is a Vast.ai listing that our data labels as marketplace. That makes it tempting to read the spread as "marketplace vs everyone else". The workstation rows do not fit that reading neatly: none of them contains a listing identified as a marketplace in our data, and the spreads are still wide.

The chart puts all ten workstation quotes on one common scale, so lengths can be compared across the three cards:

Workstation cards, 1 GPU on-demand, 2026-10-10

A4000  Hyperstack  ██▎                           $0.15/h
A4000  RunPod      ███▊                          $0.25/h
A4000  Paperspace  ███████████▎                  $0.76/h

A5000  RunPod      ████                          $0.27/h
A5000  Paperspace  ████████████████████▌         $1.38/h

A6000  Hyperstack  ███████▌                      $0.50/h
A6000  RunPod      ████████▊                     $0.59/h
A6000  Verda       ██████████▎                   $0.692/h
A6000  Lambda      ████████████████              $1.09/h
A6000  Paperspace  ████████████████████████████  $1.89/h

That is 5.1× (80% less) on the A4000, 5.1× (80% less) on the A5000 and 3.8× (74% less) on the A6000. Wide spreads also appear in rows without a listing identified as a marketplace in the supplied data. That does not tell us how each provider operates; it only tells us the gap is not limited to rows with a labelled marketplace listing.

B200 and L40S spreads

The B200 192GB sits at the other end: five providers within a 1.3× band, or 25% less from top to bottom. Same rule as above, bars start at zero on a common scale:

B200 192GB, 1 GPU on-demand, 2026-10-10

Hyperstack             █████████████████████         $6.00/h
Vast.ai (marketplace)  ██████████████████████        $6.25/h
Lambda                 ███████████████████████▌      $6.69/h
Verda                  ██████████████████████████    $7.42/h
RunPod                 ████████████████████████████  $7.99/h

On the B200, the marketplace listing is not the cheapest; Hyperstack is. We report the observation only: this sample does not tell us why the B200 range is tight while the V100 range is wide, and we will not guess. At this price level, even a narrow spread is worth a look.

The L40S 45GB sits in between with a 3× spread (67% less), with RunPod at $1.09/h between the two ends shown in the table.

What the hourly price doesn't tell you

The table compares one thing: the listed hourly price for 1 GPU, on-demand. It does not compare conditions of service, and those can differ.

Our data labels only the Vast.ai prices as marketplace listings. It does not classify the other offers, and we do not score any provider's service model here. The comparison is on price only.

Before you treat the cheapest row as a straight substitute, check these points on the provider's own terms:

• what availability and interruption terms apply to the listing;
• what the host, network and storage setup looks like, and what storage and egress cost on top;
• what support, uptime commitments and data handling you get;
• whether the exact configuration (memory, interconnect, region) matches what you need.

A wide spread is a signal to look closer, not a verdict on which offer is better value for your workload.

Prices move too

The spread is a snapshot, and prices at a single provider can move meaningfully within a month. Verda's on-demand prices between 2026-09-10 and 2026-10-10:

A B200 cost estimate built on the 2026-09-10 Verda price would already be out of date a month later. The same applies to any budget built from a one-time comparison.

How to use it

1. Start from the chip page, not the provider page. Open the page for your card to see every tracked provider side by side: V100, A100, A10, RTX A4000, A5000, A6000, L40S, H100, H200, B200. Prices are collected daily, so check before each rental rather than relying on what you saw last time.
2. Shortlist by price, then filter by conditions. Take the cheapest few offers and go through the checklist above. Drop anything that fails your requirements.
3. Feed the real rate into your cost model. The hourly rate is one input. Idle hours, utilisation and the option of using an API decide the rest. Is it cheaper to self-host, or just use an API? and What does it really cost to run an LLM? walk through that arithmetic.

Takeaway

Pick the chip for the workload. Then pick the provider from current prices on the chip page, and check the service conditions yourself before you commit.