On this page
The H100 label covers materially different deployment forms
NVIDIA's Hopper-generation H100 is a data-center accelerator used for AI, high-performance computing and analytics. NVIDIA's current HGX reference documentation lists the H100 SXM at 80 GB of HBM3 memory and 3.35 TB/s of GPU memory bandwidth. NVIDIA's technical architecture documentation also distinguishes the SXM implementation from the PCIe implementation, which uses 80 GB of HBM2e and lower memory bandwidth. H100 NVL is another distinct product with 94 GB of HBM3. These differences matter because a provider's H100 listing may not represent the same physical configuration as another provider's listing.
Source: NVIDIA HGX AI Factory reference architecture and NVIDIA Hopper architecture documentation.
For procurement, the practical implication is simple: record form factor and node architecture before recording price. An eight-GPU HGX H100 system with high-speed GPU interconnect is not equivalent to isolated PCIe access just because both use an H100 name. Likewise, a provider may package CPU, RAM, local storage and networking differently around the accelerator.
Current public H100 offers illustrate why billing units must be preserved
The following examples are source-linked observations from public provider pages checked on September 24, 2026. They are not a ranking and should not be read as guaranteed live capacity or negotiated enterprise pricing.
| Provider | Public H100 signal | Original billing context | Decision implication |
|---|---|---|---|
| CoreWeave | HGX H100 shown at $49.24/hour on demand and $19.71/hour spot | 8-GPU HGX H100 system | Normalize the node rate before comparing it with single-GPU listings. |
| Lambda | H100 SXM self-service instances shown at $3.99/GPU-hour; cluster pricing varies by GPU count and term | Per GPU-hour, with separate cluster offers | Self-service and committed cluster economics are different products. |
| RunPod | Its September 2026 H100 guide states H100 access starts at $1.99/hour on Community Cloud | Community Cloud starting rate | Marketplace-style availability and product conditions must be checked alongside price. |
| Crusoe | H100 80 GB HGX shown at $3.90/GPU-hour on demand | Per GPU-hour | Compare the same deployment and commitment type before drawing conclusions. |
Sources: CoreWeave pricing, Lambda pricing, RunPod H100 guide, and Crusoe Cloud pricing.
Normalize H100 pricing before comparing providers
A valid H100 price comparison should preserve the provider's original commercial unit and then calculate derived metrics. For example, CoreWeave's published HGX H100 on-demand figure is for an eight-GPU system. Dividing $49.24 by eight gives approximately $6.16 per GPU-hour, but that derived figure still describes a specific multi-GPU system with its own CPU, RAM, storage and network context. It should not be presented as identical to a provider selling one GPU in a different instance shape.
At minimum, a pricing record should capture currency, source date, billing unit, GPU count, exact H100 variant, region, CPU/RAM allocation, local storage, network context, commitment assumption, interruption status and applicable taxes or ancillary charges. The existing GPU Cloud Pricing Comparator should remain the canonical tool for normalizing changing commercial observations rather than creating a second generic pricing page inside the H100 cluster.
| Comparison field | Why it matters |
|---|---|
| GPU variant and memory | Separates SXM, PCIe and NVL products with different characteristics. |
| Billing unit | Prevents a node-hour figure from being compared directly with a GPU-hour figure. |
| Commitment | On-demand, spot and reserved capacity have different economics and risk. |
| Region and availability | A published SKU may not imply capacity in the required geography or at the required time. |
| Network and storage | Distributed training and data-heavy workloads can be limited outside the GPU itself. |
Workload fit determines whether H100 is the right economic choice
H100 remains relevant for large-model training, high-throughput inference, HPC and analytics workloads, but the correct decision depends on memory footprint, throughput target, scale and utilization. A small intermittent workload may value flexible access more than the lowest committed rate. A multi-node training run may value topology, high-speed fabric and capacity certainty more than the nominal single-GPU price. An inference deployment may care more about predictable latency, software operations and sustained utilization.
The decision should also compare H100 against adjacent accelerator generations rather than assume H100 is universally optimal. H200 offers materially more memory per GPU in NVIDIA's HGX reference architecture and has a dedicated H200 Cloud & Rental guide. B200 moves to Blackwell with higher memory and bandwidth and now has a separate B200 Cloud & Rental guide with its own evidence and provider set.
At sustained utilization, cloud economics should be tested against owned or colocated infrastructure using the AI Infrastructure ROI Calculator. That analysis should include utilization, financing, power, cooling, networking, staffing, support, depreciation and the opportunity cost of capacity commitments. H100 sourcing therefore connects directly to data-center infrastructure, power, cooling, technology architecture and business-model economics.
A practical H100 shortlist should start with constraints, not provider popularity
First define the technical envelope: accelerator form, GPU count, memory requirement, model size, precision, storage footprint, data movement and network sensitivity. Second define the commercial envelope: start date, duration, interruption tolerance, geography, support and budget. Third compare only offers that satisfy both envelopes. This avoids a common failure mode in which a low headline rate wins the spreadsheet even though the required capacity, topology or region is unavailable.
The GPU Cloud Providers hub provides the parent decision framework for provider types and sourcing models. H100-specific analysis then adds the accelerator constraints required to make those provider records genuinely comparable.
Methodology, limitations and freshness
This page uses official NVIDIA technical documentation and current public provider pages as primary evidence. Prices and capacity are fast-changing fields. Public pages can omit negotiated discounts, enterprise contracts, regional inventory, taxes, minimum spend, support terms and private capacity commitments. A published product should therefore not be interpreted as guaranteed immediate availability.
Pricing observations on this page were checked on September 24, 2026. AIDataCenterHQ should treat H100 price, region and availability data as F2 information requiring periodic revalidation. Material changes should update the page and sitemap lastmod only when the page itself is substantively updated.
Adjacent technology, IP and commercialization resources
H100 sourcing can intersect with accelerator IP, AI infrastructure contracting, technology commercialization and international protection. Relevant specialist resources include Patent Business Lawyer, GIP Research, GIPResearch.org, Tech Law Attorney, US Tech Law Attorney, Patent Business Attorney, International Patents, TechCorpLegal and Advocate Rahul Dev. These contextual links do not imply endorsement or provider relationships.
Frequently asked questions
How much does an H100 cost to rent in the cloud?
There is no single market price. Current public listings use different products and billing units, including per-GPU-hour rates, complete multi-GPU node rates and committed cluster terms. Normalize the exact configuration before comparing figures.
Is H100 SXM the same as H100 PCIe?
No. NVIDIA documents different memory subsystems and bandwidth characteristics for the SXM and PCIe implementations. The provider's exact H100 variant should therefore be recorded.
When is an eight-GPU H100 node useful?
It can be useful for workloads needing tightly connected multi-GPU compute, but suitability depends on model architecture, parallelism strategy, memory needs and interconnect-sensitive scaling.
Should I choose the cheapest published H100 rate?
Not without checking configuration, capacity, region, interruption risk, networking, storage, support and workload-level cost. The cheapest headline rate can be unsuitable for the actual deployment.