GPU Cloud Pricing November 2026: H100, H200, B200, GB200 and MI300X Hourly Rates Across AWS, Azure, Google Cloud and Neoclouds

GPU Cloud Pricing November 2026: H100, H200, B200, GB200 and MI300X Hourly Rates Across AWS, Azure, Google Cloud and Neoclouds

GPU Cloud Pricing November 2026: H100, H200, B200, GB200 and MI300X Hourly Rates Across AWS, Azure, Google Cloud and Neoclouds

On 20 November 2026, the same eight-GPU H100 server costs $55.04 an hour on AWS on-demand, $98.32 on Azure in East US, and $19.71 as spot capacity on CoreWeave. That is a 5x spread for nearly identical silicon. This gpu cloud pricing gap is not noise. It is the product of purchase model, interconnect, region, bundled CPU and storage, and how hard each provider wants to sell a given generation. It matters now because Nvidia reported record data center revenue three days ago, Blackwell capacity is still mostly sold by commitment, and at least one neocloud has published H100 and H200 on-demand increases for 1 December. A headline dollars-per-GPU-hour figure is the least useful number in a GPU quote, yet it is the one everyone compares. You will leave with a verified price table pulled from primary pricing pages this morning, a normalization method you can reuse, and worked math that turns hourly rates into cost per million tokens and cost per training run.

What this covers: verified hourly rates for H100, H200, B200, B300, GB200 and MI300X across AWS, Azure, Google Cloud, Oracle and seven neoclouds; purchase models; interconnect; effective-cost math; and a decision guide by workload. This page is updated monthly.

Context and Background

Three years ago the GPU rental market had one question: can you get H100s at all? By late 2026 the question has changed. Hopper (H100 and H200) is abundant on most clouds, Blackwell (B200, B300, GB200 and GB300) is the generation providers are racing to fill, and AMD’s Instinct MI300X has a real but narrower foothold, mainly on Azure, Oracle and a few neoclouds. The term “neocloud” describes GPU-first providers such as CoreWeave, Lambda, Crusoe, Nebius and Together AI that sell accelerated compute without the full hyperscaler service catalog.

Demand has not cooled. NVIDIA’s third-quarter fiscal 2027 results, released on 17 November 2026, reported revenue of $115.8 billion, up 103% year on year, with data center revenue of $107.6 billion, up 110%. The company guided to $130.0 billion, plus or minus 3%, for the fourth quarter. That release talks about Vera Rubin as “off to an explosive start”, which is a forward signal: Hopper prices are now set by the arrival of two newer generations, not by scarcity.

GPU cloud pricing reflects that. Hyperscalers still list H100 on-demand rates that look expensive next to neoclouds, but they discount heavily through reservations, committed use discounts and short-term capacity products. Neoclouds publish simpler per-GPU rates and lean on spot and reserved clusters. Marketplaces such as Vast.ai sit at the bottom of the range, with variable hosts and variable reliability.

We published an earlier November snapshot of cloud GPU per-hour costs that focused on headline rates. This edition is deliberately different. It adds GB200 and B300, pulls hyperscaler numbers from machine-readable price APIs rather than marketing pages, and argues a specific thesis: for most teams in late 2026, the purchase model and the interconnect class move total cost more than the choice of cloud does. If you want the hardware background first, our GB300 NVL72 Blackwell Ultra architecture deep dive explains why a rack-scale NVLink domain is priced as a different product from an eight-GPU node.

One framing note. Every number below is a list price. Large buyers negotiate, and private pricing can be far below list. A living benchmark of list prices is still useful, because list prices anchor negotiations, set the cost of burst capacity, and are the only numbers anyone can verify.

GPU Cloud Pricing in November 2026: The Verified Reference Table

GPU cloud pricing in November 2026 ranges from about $2.46 per H100-hour (CoreWeave spot) to $12.29 (Azure East US on-demand), and from about $4.26 per B200-hour (CoreWeave spot) to $14.24 (AWS on-demand). GB200 lists at $10.50 per GPU-hour on CoreWeave, $10.58 in AWS Capacity Blocks and $27.04 on Azure on-demand.

GPU cloud pricing normalization chain from list price per instance hour to cost per useful token

Figure 1: How a list price per instance-hour becomes an effective cost per useful token or per training run.

Figure 1 is the method behind every table on this page. We start from the provider’s own list price per instance-hour, divide by the number of GPUs in the instance, and record that as the headline per-GPU rate. Everything to the right of that box changes the real cost: the purchase model, the add-ons, the hours you pay for but do not use, and the throughput the hardware actually delivers. The dotted inputs, interconnect class and software stack, change delivered throughput, which is why two identical headline rates can produce very different costs per token.

Hyperscalers: AWS, Azure, Google Cloud and Oracle

The hyperscaler rows below come from machine-readable sources where possible: the AWS EC2 on-demand price file (publication stamp 2026-11-20T00:50Z), the Azure Retail Prices API, and the Google Cloud accelerator-optimized pricing page. Oracle rows come from Oracle’s public price-list API. Derived per-GPU rates are the instance rate divided by GPU count, computed in Python and rounded to cents.

Provider and instance GPUs Region Model $/instance-hr $/GPU-hr
AWS p5.48xlarge 8 x H100 us-east-1 On-demand 55.04 6.88
AWS p5.4xlarge 1 x H100 us-east-1 On-demand 6.88 6.88
AWS p5e.48xlarge 8 x H200 us-east-2 On-demand 63.30 7.91
AWS p5en.48xlarge 8 x H200 us-east-1 On-demand 63.30 7.91
AWS p6-b200.48xlarge 8 x B200 us-east-1 On-demand 113.93 14.24
AWS p6-b300.48xlarge 8 x B300 us-east-1 On-demand 142.42 17.80
AWS u-p6e-gb200x72 72 x GB200 GPUs Dallas Local Zone Capacity Block 761.90 10.58
AWS p5.48xlarge 8 x H100 multiple Capacity Block 47.76 5.97
AWS p6-b200.48xlarge 8 x B200 multiple Capacity Block 113.67 14.21
Azure ND96isr H100 v5 8 x H100 East US Pay as you go 98.32 12.29
Azure ND96isr H100 v5 8 x H100 East US 3-year reserved 43.16 5.40
Azure ND96isr H200 v5 8 x H200 East US 2 Pay as you go 84.80 10.60
Azure ND128isr NDR GB200 v6 4 x GB200 GPUs East US Pay as you go 108.16 27.04
Azure ND128isr NDR GB200 v6 4 x GB200 GPUs East US 3-year reserved 47.59 11.90
Azure ND96isr MI300X v5 8 x MI300X East US 2 Pay as you go 48.00 6.00
Google a3-highgpu-8g 8 x H100 page default On-demand 88.49 11.06
Google a3-megagpu-8g 8 x H100 page default On-demand 93.40 11.68
Google a3-ultragpu-8g 8 x H200 page default On-demand 84.81 10.60
Google a4-highgpu-8g 8 x B200 page default On-demand not listed not listed
Google a4-highgpu-8g 8 x B200 page default DWS Flex-start 64.44 8.06
Oracle BM.GPU.H100.8 8 x H100 price list Pay as you go 80.00 10.00
Oracle BM.GPU.B200.8 8 x B200 price list Pay as you go 112.00 14.00
Oracle BM.GPU.GB200 4 x GB200 GPUs price list Pay as you go 64.00 16.00
Oracle BM.GPU.MI300X.8 8 x MI300X price list Pay as you go 48.00 6.00

Retrieved 20 November 2026. Azure reserved rows are the API’s total term price divided by term hours. Oracle publishes a per-GPU rate; the instance column is that rate multiplied by GPUs per shape, following Oracle’s own stated formula.

A few rows need explaining. Google’s page lists no on-demand price for the B200-based A4 machine; it lists DWS Flex-start ($64.44), DWS Calendar Mode ($90.22), spot ($43.60) and resource-based committed use discounts ($88.93 for one year, $56.71 for three). In other words, Google sells Blackwell by commitment or by its Dynamic Workload Scheduler, not as plain on-demand. Google’s page also does not state which region the default figures apply to, so we label them “page default”. The A4X (GB200) machine did not appear on the pricing page at all at retrieval time.

AWS lists the GB200 NVL72 UltraServer only as a Capacity Block in the Dallas Local Zone, at $761.904 per hour for 72 GPUs or $380.952 for 36. AWS states that Capacity Block prices are “scheduled to be updated next in January, 2027” and are charged at the rate in effect at purchase. Oracle also publishes a GB300 rate of $18.00 and a B300 rate of $15.00 per GPU-hour, and an MI355X rate of $8.60.

Neocloud GPU pricing and marketplaces

Neoclouds publish per-GPU or per-node rates on public pages. The table normalizes them to per-GPU-hour.

Provider GPU Unit shown On-demand $/GPU-hr Spot or preemptible Reserved
CoreWeave GB200 NVL72 4-GPU instance, $42.00 10.50 not listed contact sales
CoreWeave HGX B200 8-GPU, $68.80 8.60 4.26 contact sales
CoreWeave HGX B300 8-GPU contact sales 4.48 contact sales
CoreWeave HGX H200 8-GPU, $50.44 6.31 2.62 contact sales
CoreWeave HGX H100 8-GPU, $49.24 6.16 2.46 contact sales
Lambda HGX B200 per GPU 7.69 not listed 1-Click Cluster 8.87 to 9.86
Lambda HGX H100 per GPU 4.59 not listed contact sales
Nebius B300 / B200 per GPU 9.50 / 8.50 not listed contact sales
Nebius H200 / H100 per GPU 5.40 / 4.50 not listed contact sales
Crusoe, until 30 Nov H200 / H100 / MI300X per GPU 4.29 / 3.90 / 3.45 not listed contact sales
Crusoe, from 1 Dec B200 / H200 / H100 per GPU 8.50 / 5.40 / 4.50 not listed contact sales
Together AI HGX B200 per GPU 9.29 4.65 7.99 to 8.69
Together AI HGX H200 / H100 per GPU 5.99 / 3.99 2.99 / 1.99 4.15 to 4.99 / 3.45 to 3.69
RunPod Secure Cloud B200 / H200 / H100 SXM per GPU 7.99 / 5.29 / 4.39 not listed contact sales
RunPod Community Cloud H200 / H100 SXM per GPU from 3.59 / from 2.69 not listed n/a
Vast.ai marketplace B200 / H200 / H100 SXM per GPU, median of live offers 8.74 / 4.61 / 2.34 interruptible, varies varies

Retrieved 20 November 2026 from each provider’s public pricing page; Vast.ai from its public offer search API (on-demand, rentable, unrented offers; n = 18 B200, 16 H200, 13 H100 SXM). Together AI reserved ranges cover 3 to 90 day terms. CoreWeave GB200 instances are two Grace Blackwell Superchips, each with one Grace CPU and two Blackwell GPUs, per CoreWeave’s footnote.

Two things jump out. First, the floor for Hopper on-demand at mainstream neoclouds is now roughly $4 to $4.60 per H100-hour, not the $2-and-change rates common in 2025 marketing. Crusoe’s page shows a scheduled rise from $3.90 to $4.50 for H100 and from $4.29 to $5.40 for H200 on 1 December. Second, Blackwell on-demand at neoclouds clusters around $7.70 to $9.30 per B200-hour, while hyperscaler B200 sits at $14.00 to $14.24 on AWS and Oracle. The hyperscaler premium for Blackwell is larger than it is for Hopper.

How we normalize GPU cloud pricing, and why the per-GPU column is only a starting point

We divide instance price by GPU count because buyers compare per GPU. But an AWS p5.48xlarge bundles 192 vCPUs, 2 TiB of RAM, eight 3.84 TB NVMe drives and 3,200 Gbps of EFA networking. A marketplace H100 may come with a fraction of that. A GB200 “GPU” on Azure is one quarter of a VM that is itself a slice of a 72-GPU NVLink domain. Treating all of these as one “GPU-hour” ignores what you get alongside the silicon.

The normalization rule we use: compare per-GPU rates only within the same interconnect class and the same purchase model. Then convert to cost per unit of work. The next section shows how.

Why $/GPU-hr Misleads: Purchase Models, Interconnect and Effective Cost

In GPU cloud pricing, headline rates are inputs, not answers. Three factors decide what a GPU actually costs you: how you buy it, how it is wired to its neighbours, and how much useful work you extract from each billed hour.

GPU cloud pricing purchase models compared: spot, on-demand, capacity blocks and reserved commitments

Figure 2: Purchase models by duration and certainty, and the hidden cost each one carries.

Figure 2 maps the four purchase models to the question that selects between them: how long do you need the capacity, and how certain is that need? Each branch ends in the risk you are trading for the discount. Spot trades price for interruption. On-demand trades price for flexibility but carries availability risk for full eight-GPU nodes. Capacity Blocks trade flexibility for a guaranteed window. Reservations trade flexibility for the deepest committed discount, and you pay for idle hours.

Purchase models, quantified

On-demand is the anchor everyone quotes and few large buyers pay. On Azure, the H100 v5 one-year reservation works out to $62.93 per instance-hour against $98.32 pay-as-you-go, a 36% discount. The three-year term is $43.16, a 56% discount. Google’s a3-highgpu-8g shows a similar ladder: $88.49 on-demand, $61.38 with a one-year CUD (31% off), and $38.86 with a three-year CUD (56% off).

Capacity Blocks and calendar reservations are the 2024 to 2026 innovation. AWS EC2 Capacity Blocks for ML let you reserve GPU instances for a fixed future window and pay up front. For p5.48xlarge the block rate is $47.757 per hour, 13% below on-demand. For p6-b200 the block rate ($113.666) is almost identical to on-demand ($113.933), which tells you AWS is not using blocks to discount Blackwell; it is using them to ration it. Google’s DWS Calendar Mode and Flex-start play the same role. Flex-start H100 on a3-highgpu-8g is $38.32 per instance-hour, 57% below on-demand.

Spot and preemptible capacity is the cheapest list price and the most misunderstood. CoreWeave spot H100 at $2.46 per GPU-hour is 60% below its own on-demand rate. Azure lists ND H100 v5 spot at $21.80 per instance-hour in North Central US, or $2.73 per GPU-hour. But some Azure API rows list “spot” GPU prices that equal pay-as-you-go, for example GB200 v6 spot in East US at $108.16. We treat those rows as listed but not meaningfully discounted. Google’s spot price for a3-highgpu-8g ($52.70) is higher than its Flex-start price ($38.32), a reminder that spot prices are set by demand and can move daily.

Reserved clusters at neoclouds are mostly “contact sales”. Together AI is the exception: it publishes B200 at $8.69 per GPU-hour for 3 to 30 days and $7.99 for 31 to 90 days, against $9.29 on-demand. Lambda publishes 1-Click Cluster B200 pricing from $9.86 per GPU-hour at 16 GPUs down to $8.87 at 256 or more.

Interconnect: the line item you cannot see in $/GPU-hr

GPU cloud pricing and interconnect: HGX eight-GPU nodes with EFA or InfiniBand versus a GB200 NVL72 rack-scale NVLink domain

Figure 3: Two interconnect classes. An HGX node has an eight-GPU NVLink island; a GB200 NVL72 rack is one 72-GPU NVLink domain.

Figure 3 contrasts the two dominant designs. In an HGX node, the eight GPUs share an NVLink switch inside the server, and every byte that leaves the node goes over a scale-out network: Elastic Fabric Adapter (EFA) on AWS, InfiniBand on Azure and many neoclouds, or RoCE Ethernet elsewhere. AWS lists p5.48xlarge with 3,200 Gbps of EFA and p6-b300.48xlarge with 6,400 Gbps. In a GB200 NVL72 rack, all 72 Blackwell GPUs sit in a single NVLink domain, so tensor-parallel and expert-parallel traffic for very large models never touches the slower scale-out fabric.

That is why a GB200 GPU-hour and a B200 GPU-hour are not the same product even at similar prices. CoreWeave lists GB200 at $10.50 per GPU-hour against B200 at $8.60, a 22% premium. For a dense 8B model that fits on one GPU, that premium buys nothing. For a trillion-parameter mixture-of-experts model served with wide expert parallelism, the rack-scale NVLink domain can be the difference between a viable and an unviable deployment. Our comparison of NVLink, InfiniBand and RoCE for NCCL collectives explains the all-reduce maths behind that claim.

Interconnect also explains part of the hyperscaler premium. An AWS p5 or Azure ND H100 v5 node comes with a non-blocking, multi-terabit fabric designed for thousands of GPUs. A marketplace H100 typically does not. If your job is a single-node fine-tune, you are paying for fabric you do not use. If it is a 512-GPU pretraining run, a cheap GPU without a fabric is not a substitute at any price.

From hourly rate to cost per million tokens (illustrative)

Inference buyers care about cost per token. The conversion is:

cost per million tokens = $/GPU-hr ÷ (tokens per second per GPU × 3,600 ÷ 1,000,000 × utilization)

Throughput depends on model, precision, batch size, sequence lengths and serving engine, so we use illustrative round numbers, not measurements: 1,500 output tokens per second per GPU for a mid-sized model on H100 or MI300X, and 4,000 per GPU on B200 using FP8 or FP4. Treat these as placeholders; plug in your own benchmark.

Scenario (illustrative throughput) $/GPU-hr At 100% utilization At 50% utilization
H100, AWS on-demand, 1,500 tok/s 6.88 $1.27 / M tokens $2.55 / M tokens
H100, CoreWeave on-demand, 1,500 tok/s 6.16 $1.14 / M tokens $2.28 / M tokens
H100, Vast.ai median, 1,500 tok/s 2.34 $0.43 / M tokens $0.87 / M tokens
MI300X, Azure pay as you go, 1,500 tok/s 6.00 $1.11 / M tokens $2.22 / M tokens
B200, CoreWeave on-demand, 4,000 tok/s 8.60 $0.60 / M tokens $1.19 / M tokens
B200, AWS on-demand, 4,000 tok/s 14.24 $0.99 / M tokens $1.98 / M tokens

Illustrative. Throughput figures are assumptions for arithmetic, not benchmarks.

Two conclusions survive any reasonable change to the assumptions. First, utilization dominates: halving utilization doubles cost per token, which is a bigger swing than switching from CoreWeave to AWS. Second, if Blackwell really delivers 2.5x or more Hopper’s tokens per second for your model, a B200 at $8.60 beats an H100 at $4.50 on cost per token. Whether it does depends on low-precision formats; see our explainer on NVFP4 versus MXFP4 for why FP4 support is the main source of Blackwell’s inference advantage. For the serving-side metrics that set real utilization, read our guide to LLM inference SLOs, goodput and capacity planning.

From hourly rate to cost per training run (illustrative)

Training cost follows the standard approximation that a dense transformer needs about 6 × N × D floating-point operations, where N is parameters and D is training tokens. Take an 8-billion-parameter model trained on 1 trillion tokens: 6 × 8e9 × 1e12 = 4.8e22 FLOPs.

Nvidia’s data sheet lists H100 SXM at about 989 dense BF16 teraFLOPS. Assume 40% model FLOPs utilization (MFU), a common but workload-dependent figure. That gives about 33,700 H100 GPU-hours. For B200 we assume roughly 2,250 dense BF16 teraFLOPS (reported for HGX B200; check your SKU) at 35% MFU, giving about 16,900 GPU-hours.

Scenario (illustrative) GPU-hours $/GPU-hr Compute cost
H100, AWS on-demand 33,700 6.88 ~$231,900
H100, CoreWeave on-demand 33,700 6.16 ~$207,400
H100, Nebius on-demand 33,700 4.50 ~$151,700
H100, Together AI on-demand 33,700 3.99 ~$134,500
H100, CoreWeave spot, no interruption 33,700 2.46 ~$83,000
B200, AWS on-demand 16,900 14.24 ~$241,100
B200, CoreWeave on-demand 16,900 8.60 ~$145,600
B200, Lambda instance 16,900 7.69 ~$130,200

Illustrative. MFU, peak FLOPS and run length are assumptions; storage, networking, egress and failed runs are excluded.

The pattern matters more than the exact numbers. On a neocloud, B200 and H100 land at similar total cost for this run, but B200 finishes in about half the wall-clock time: roughly 11 days on 64 B200s versus 22 days on 64 H100s. On AWS on-demand, B200 is slightly more expensive in total, because the Blackwell list premium there is larger than the speed-up we assumed. Spot looks spectacular on paper, but a multi-week synchronous training job on preemptible capacity pays a rework tax every time a node disappears. Our guide to spot and preemptible GPU training with checkpointing quantifies that tax.

Methodology caveats for this benchmark

This GPU cloud pricing benchmark uses list prices only, and it has limits you should know before you quote it.

  • Retrieval date and volatility. All prices were retrieved on 20 November 2026. Spot and marketplace prices can change daily or hourly. Crusoe has already announced changes effective 1 December.
  • Regions differ. Azure ND H200 v5 is $84.80 in East US 2 but $110.24 in West US and $131.44 in Japan West. We show one representative region per row; your region may be 20% to 55% higher.
  • Bundles differ. Hyperscaler instances include CPUs, RAM, local NVMe and fabric. Some neocloud per-GPU rates include similar bundles; marketplace offers vary per host.
  • Scraped tables can mislead. Crusoe’s pricing table was read from flattened page text and matched by row order; verify on the live page before contracting. Google’s page does not state the region for its default figures.
  • Marketplace samples are small. Vast.ai medians come from 13 to 18 live offers per GPU type at one moment, and the totals may include small storage charges.
  • No negotiated prices. Enterprise agreements, credits and private offers are excluded by design.

Trade-offs, Gotchas, and What Goes Wrong

The cheapest GPU-hour is often the most expensive project, and most GPU cloud pricing mistakes are utilization mistakes. Here are the failure modes we see most.

Chasing spot for synchronous training. A 256-GPU data-parallel job stalls when any one node is reclaimed. If preemptions and restarts waste 20% of the hours you pay for, a $2.46 spot H100 effectively costs about $3.08, and you also lose calendar time. Spot is excellent for embarrassingly parallel batch inference, evaluation sweeps and hyperparameter searches. It is a poor default for long synchronous runs without elastic training support.

Reserving for a peak you never hit. A three-year Azure H100 reservation at $5.40 per GPU-hour looks cheap until utilization falls to 50%, at which point you are paying $10.79 per useful GPU-hour, close to the on-demand rate. Reservations convert price risk into utilization risk. If your fleet is Kubernetes-based, measure real utilization first with GPU-aware cost allocation; our OpenCost guide to Kubernetes cost allocation, rightsizing and showback covers the mechanics.

Buying a whole GPU for a small model. Many inference workloads use a fraction of an 80 GB H100. Paying for eight full GPUs to serve several small models wastes most of the bill. Partitioning with MIG, MPS or time-slicing can multiply effective utilization; our GPU sharing on Kubernetes comparison of MIG, MPS, time-slicing and vGPU explains the isolation trade-offs.

Ignoring the software tax on MI300X. MI300X lists at $6.00 per GPU-hour on Azure and Oracle, and $3.45 on Crusoe, with 192 GB of HBM per GPU. That is among the lowest on-demand list prices per gigabyte of HBM in the table. But your cost per token depends on ROCm kernel maturity for your specific model and serving engine. Budget engineering time for validation; our ROCm versus CUDA comparison for MI300X and MI350 lists where the gaps remain.

Forgetting egress and storage. Moving a multi-terabyte checkpoint set out of a hyperscaler can cost more than a day of GPU time. Neoclouds often waive or discount egress, but check the contract. Our cloud data transfer cost architecture guide covers the patterns.

Assuming on-demand means available. An on-demand price for p6-b200 or ND GB200 v6 does not guarantee you can launch one today. Blackwell capacity at hyperscalers is still largely allocated through reservations and blocks. Plan for allocation, not just price.

Practical Recommendations

GPU cloud pricing decision guide by workload: prototyping, batch inference, online inference, pretraining and enterprise tenancy

Figure 4: A decision guide mapping workload type to the purchase model and GPU class that usually minimizes effective cost.

Figure 4 condenses the analysis into a decision tree. Start from the workload, not the provider. Prototyping and small fine-tunes belong on per-second-billed neocloud or marketplace instances. Batch and offline inference belong on spot Hopper with checkpointing. Online inference splits by whether the model fits in one node: if it does, an H200 or B200 node with a reserved baseline is usually cheapest; if it is a very large MoE model, a slice of a GB200 NVL72 domain can justify its premium. Pretraining above a few hundred GPUs needs a reserved cluster with a proper fabric. Regulated workloads usually stay on a hyperscaler with reservations plus spot burst.

Three principles sit behind the tree, and they hold regardless of how GPU cloud pricing moves next month. Price the unit of work, not the GPU-hour. Match the purchase model to the certainty of the demand. Pay for interconnect only when the parallelism strategy needs it. For training cluster scheduling choices that affect utilization, see our Slurm versus Kubernetes guide for AI training. For offline inference, our offline batch LLM inference architecture guide compares self-hosted GPUs with provider batch APIs.

A short checklist before you sign:

  • Benchmark tokens per second or MFU for your model on two GPU classes before comparing prices.
  • Measure fleet utilization for at least two weeks before committing to a reservation.
  • Compare per-GPU rates only within the same interconnect class and purchase model.
  • Ask every neocloud for a reserved quote; list on-demand is a ceiling, not a price.
  • Check the region: the same Azure SKU varies by more than 50% across regions.
  • Price egress, storage and support alongside the GPU line.
  • Re-check spot and marketplace prices on the day you launch.

Frequently Asked Questions

What is the H100 price per hour in November 2026?

List H100 prices on 20 November 2026 range widely. On-demand, AWS p5.48xlarge works out to $6.88 per GPU-hour in us-east-1, Azure ND H100 v5 to $12.29 in East US, and Google a3-highgpu-8g to $11.06 at the page’s default region. Neoclouds are lower: Together AI $3.99, Nebius $4.50, Lambda $4.59 and CoreWeave $6.16. Spot and marketplace rates go lower still: CoreWeave spot $2.46 and a Vast.ai median of about $2.34. Reservations can cut hyperscaler rates by 36% to 56%.

How much does B200 cloud pricing cost per GPU-hour?

B200 on-demand ranges from about $7.69 per GPU-hour (Lambda instances) to $14.24 (AWS p6-b200.48xlarge). CoreWeave lists $8.60, Nebius $8.50, Together AI $9.29 and RunPod Secure Cloud $7.99. Oracle’s price list shows $14.00. Google does not list on-demand for its B200-based A4 machine; it lists DWS Flex-start at $8.06 per GPU-hour and a three-year commitment at $7.09. Spot B200 is listed at $4.26 on CoreWeave and $4.65 preemptible on Together AI.

Why are hyperscaler GPUs so much more expensive than neocloud GPUs?

Part of the gap is real value: hyperscaler instances bundle large CPU and memory allocations, local NVMe, multi-terabit fabrics, enterprise identity, compliance attestations and integration with managed services. Part is pricing strategy: hyperscalers set high on-demand anchors and expect large buyers to commit, which brings effective rates much closer to neoclouds. Azure’s three-year H100 reservation, for example, works out to $5.40 per GPU-hour against $12.29 on-demand. Neoclouds compete on simpler, lower list prices for GPU-first workloads.

Is a GB200 GPU-hour worth more than a B200 GPU-hour?

Only for workloads that use the rack-scale NVLink domain. A GB200 NVL72 rack links 72 Blackwell GPUs in one NVLink domain, so very large models can use tensor or expert parallelism across many GPUs without crossing a slower network. CoreWeave lists GB200 at $10.50 per GPU-hour versus $8.60 for B200, and AWS sells GB200 only through Capacity Blocks at $10.58. For models that fit on eight GPUs, the premium buys little; for trillion-parameter MoE serving, it can be decisive.

Are AWS Capacity Blocks cheaper than on-demand?

Sometimes, and not by much. On 20 November 2026, p5.48xlarge Capacity Blocks were $47.757 per instance-hour against $55.04 on-demand, a 13% saving, plus a guaranteed window. For p6-b200, the block rate of $113.666 was essentially equal to on-demand. Capacity Blocks are mainly a way to secure scarce capacity for a fixed period, not a discount product. AWS says block prices are next scheduled to update in January 2027, and you pay the rate in effect at purchase.

How often is this GPU cloud pricing page updated?

This page is a living benchmark, updated monthly. Each edition re-pulls hyperscaler prices from machine-readable sources (the AWS price file, the Azure Retail Prices API and Oracle’s price-list API) and neocloud prices from their public pricing pages, then recomputes every derived per-GPU figure. Spot and marketplace prices can change daily, so treat those rows as a snapshot. If you are about to sign a contract, re-check the provider’s live page and ask for a reserved quote.

Further Reading

By Riju — about

Comments

No comments yet. Why don’t you start the discussion?

Leave a Reply

Your email address will not be published. Required fields are marked *