Blog ▸ Wholesale GPU Cluster Price Benchmark: October 2026
GPU Infrastructure
B300 clusters are quoting below B200 on committed terms this month despite 50% more memory. Committed cluster price bands across every tier.
Wholesale GPU Cluster Price Benchmark: October 2026
GPUaaS.com Team
GPU Infrastructure
October 1, 2026
No items found.
B300 clusters are quoting below B200 clusters on committed terms this month, across the whole band.
The newer part carries 50% more memory and costs less. It is also the most requested GPU on the platform right now.
October 2026 at a glance
B300 committed bands sit below B200 across the full range, despite the larger memory footprint
GB300 NVL72 rack-scale is quoting at the bottom of the Blackwell Ultra range, not the top
Blackwell-class parts dominate buyer requests. Hopper has become the minority ask
Committed bands run well below published hyperscaler on-demand rates across every tier
Regional variation is minimal this month. The same bands hold across most sourcing regions
◆ COMMITTED CLUSTER PRICE BANDS, OCTOBER 2026
GPU
Band per GPU-hour
Typical use
RTX 4090
$0.50
Small-model inference, dev
RTX 5090
$0.50 - $0.75
Dev, small-batch inference
A100 40GB
$1.00 - $1.50
Fine-tuning, legacy pipelines
RTX PRO 6000
$1.25 - $1.75
Mid-tier inference, 96GB
H100 PCIe
$1.75 - $2.00
Single-node training, inference
H100 SXM
$2.50 - $3.00
Multi-node training
H200 SXM
$2.75 - $3.50
Long-context inference, training
GB300 NVL72
$5.00 - $5.50
Rack-scale, single NVLink domain
B300 SXM
$5.00 - $6.50
Frontier training, large inference
B200
$7.00 - $8.00
Frontier training, large inference
Committed-term cluster bands across the GPUaaS.com network, October 2026. Bands rather than fixed rates, because term length, cluster size, region and interconnect all move the final number. Contact for commitment terms.
◆ THE INVERSION
More memory, lower price
B300 carries 288GB of HBM3e against B200's 192GB, and this month it quotes below B200 at both ends of the band. That is not how generational pricing usually works, and it is worth checking before a spec gets locked.
For a buyer the implication is direct. Specifying B200 because it is the familiar part means paying more for less memory. The sizing arithmetic already favoured B300 on memory-bound workloads. This month the price does too.
$5.00 - $6.50
the B300 SXM committed band this month, below the B200 band at every point, on a part carrying 50% more memory
GPUaaS.com network, October 2026
◆ RACK SCALE IS NOT THE PREMIUM TIER
GB300 NVL72 at the bottom of the Ultra range
GB300 NVL72 puts 72 Blackwell Ultra GPUs into one NVLink domain, liquid cooled, sold as a unit. Conventionally that configuration prices above node-level because the fabric is the product.
This month it quotes at the bottom of the Blackwell Ultra range rather than the top. Where the allocation is available, rack-scale is not carrying the premium a buyer would expect to pay for it. That makes the usual rack-versus-node question less about budget and more about whether the workload can use a single memory domain.
◆ WHAT BUYERS ARE ACTUALLY ASKING FOR
Newest silicon, not cheapest
Across requests coming through the network this quarter, B300 is the most asked-for GPU, ahead of B200. Blackwell-class parts together account for the clear majority of specified requests, while Hopper has become a minority ask rather than the default.
Below that, RTX PRO 6000 and GB300 NVL72 appear at similar frequency, and a small but persistent stream of requests names AMD MI300-series parts. A handful of requests already specify Vera Rubin, which is not shipping.
The pattern worth noting is that demand concentrates on the newest available generation at the same time as that generation prices below its predecessor. Buyers are not trading down to save money this quarter.
◆ AGAINST THE PUBLISHED REFERENCE
Committed terms and on-demand list are different products
Published hyperscaler on-demand medians in late September ran near $7.89 for H100, $10.85 for H200 and $14.24 for B200. Specialist GPU clouds published roughly $5.00, $4.50 and $7.88 for the same parts.
Those are on-demand numbers and the bands above are committed terms, so the comparison is directional rather than like for like. Committed capacity trades flexibility for rate, which is why it prices lower, and a buyer who needs to spin capacity up and down weekly should be reading the on-demand column instead.
Spot sits lower again. Reported H100 spot near $2.50 looks competitive with a committed band until the interruption profile enters the comparison, which is the whole substance of the two-price market.
◆ HOW TO READ A BAND
Four things move a quote inside it
Term length moves it most. A twelve-month commitment lands lower in the band than three months on the same hardware, because the provider is pricing certainty rather than silicon.
Cluster size moves it next. Contiguous capacity at scale prices differently from the same GPU count scattered across nodes, and a request that can tolerate non-contiguous allocation usually lands lower.
Interconnect tier and region fill in the rest. InfiniBand-connected capacity sits above Ethernet at the same GPU count, and regional variation was minimal this month, with similar bands holding across most sourcing regions.
This is the first edition of a monthly benchmark. The intention is a dated reference that can be checked against rather than a marketing number, which means publishing bands that sometimes move the wrong way and noting when they do.
Next edition publishes in early November, with what moved and why.
Get a quote inside these bands.
Submit a spec and quotes come back within a few hours, sourced across 20+ vetted partners. For single GPUs, packet.ai handles self-serve access with 24/7 human support.
B300 SXM committed cluster capacity is quoting at $5.00 to $6.50 per GPU-hour this month. Where a quote lands inside that band depends on term length, cluster size, interconnect tier and region.
This month, yes. B300 SXM bands sit at $5.00 to $6.50 against $7.00 to $8.00 for B200, on a part carrying 288GB against 192GB. Worth confirming with a current quote before a spec is locked, since bands move month to month.
Because a single rate would be wrong for most requests. Term length, cluster size, interconnect tier and region each move the final number materially, so a band describes the real range a buyer should expect and a fixed figure would not.
They are different products. Committed capacity trades flexibility for rate and prices lower than published on-demand list, which in late September ran near $7.89 for H100 and $14.24 for B200 at hyperscalers. A workload that scales up and down weekly should price against on-demand instead.
B300, ahead of B200. Blackwell-class parts together make up the clear majority of specified requests this quarter, with Hopper now a minority ask. Demand is concentrating on the newest available generation at the same time as that generation prices below its predecessor.
Last reviewed: 2 October 2026. Committed cluster bands across the GPUaaS.com network, October 2026, expressed as ranges because term length, cluster size, interconnect tier and region move the final quote. Demand direction drawn from buyer requests received through the network this quarter, reported as ranking only. Published on-demand reference figures from the GetDeploying price tracker via the DataStorage State of GPU Cloud September 2026 report, and spot reference from Silicon Data. Bands move month to month and should be confirmed with a current quote before use in a model. Browse current GPU cluster availability on GPUaaS.com.
