How much does a single NVIDIA H100 cluster node cost to lease in 2027?
It depends — leasing a single NVIDIA H100 cluster node in 2027 hinges on contract length, whether GPUs are dedicated or on-demand, and how much networking, storage, and support the provider bundles in. As a rule, on-demand hourly rentals from cloud marketplaces cost the most per GPU-hour, while multi-year committed reservations from specialized "neocloud" providers cost dramatically less. Expect prices to keep drifting downward through 2027 as H200 and Blackwell-class supply expands and older Hopper capacity gets repriced.
A single "node" is not a single GPU. A standard H100 cluster node is an 8-GPU HGX server, so any lease quote you compare must be normalized to the same unit — per-GPU-hour, per-node-hour, or per-node-month — before the numbers mean anything. This guide breaks down what actually drives the price, how the market is structured in 2027, and how to negotiate a lease that reflects real capacity rather than headline sticker rates.
What exactly counts as a single H100 "node," and why does the unit matter?
In data-center terms, an H100 node almost always refers to an 8-GPU HGX H100 baseboard server: eight SXM5 H100 GPUs linked by NVLink and NVSwitch inside one chassis, paired with high-core-count CPUs, several terabytes of system RAM, local NVMe, and multiple high-speed network interface cards. When someone quotes a "node," confirm the GPU count, the interconnect (SXM/NVLink versus PCIe), and the memory variant (80GB is standard, with 94GB and higher-bandwidth SKUs in the field by 2027). A PCIe-based 8×H100 box without NVLink is a different — and cheaper — animal that will underperform on large multi-GPU training jobs.
The unit matters because leases are advertised in inconsistent denominations. A marketplace may show a per-GPU-hour rate that looks tiny until you multiply by eight and add the always-on premium for reserved capacity; a colocation-style provider may quote a flat per-node-month figure that already folds in power, cooling, and cross-node InfiniBand fabric. Normalizing everything to the same basis — ideally per-node-month for committed use and per-GPU-hour for burst use — is the single most important discipline when you compare quotes. For a deeper primer on sizing GPU infrastructure to workloads, see the walkthrough at https://pulserevops.com/knowledge/gpu-node-sizing.
What drives the price of an H100 node lease in 2027?
Six variables move the number more than anything else, and they compound. The first is commitment term: on-demand and hourly spot capacity carries the highest unit price because the provider absorbs idle risk, while one-, two-, and three-year reservations unlock progressively steeper discounts. The second is dedicated versus shared: a reserved bare-metal node you hold 24/7 costs more in absolute dollars but far less per useful GPU-hour than paying on-demand rates for the same utilization. The third is cluster networking — nodes wired into a non-blocking InfiniBand or high-speed Ethernet fabric with RDMA cost meaningfully more than isolated servers, because that fabric is what makes multi-node training viable at all.
The remaining drivers are region and power cost, support tier and SLA, and ancillary services. Electricity and cooling dominate a GPU node's operating cost, so providers in low-power-cost regions can undercut those in expensive metros; the tradeoff is often latency and data-residency constraints. Enterprise SLAs, managed orchestration, and premium support add a layer on top of raw compute, while egress fees, high-performance parallel storage, and dedicated IP allocations show up as line items that can quietly rival the GPU cost itself. The diagram below sketches how these inputs stack into a final lease figure.
Because these factors interact, two quotes with the same headline per-GPU-hour rate can differ by a wide margin in true total cost of ownership once you account for utilization, fabric, and egress.
How is the 2027 H100 leasing market structured?
By 2027 the market has settled into several distinct layers, each with a different price posture. Hyperscale clouds (the largest general-purpose providers) sit at the premium end: they bundle H100 capacity with a full managed ecosystem, deep integrations, and enterprise contracts, and they charge accordingly. Specialized GPU clouds — the "neocloud" tier that exists primarily to rent accelerators — typically undercut hyperscalers substantially for comparable raw compute, especially on committed reservations, because their whole business is dense GPU utilization rather than a broad service catalog.
Below those sit marketplaces and aggregators that match idle capacity from many operators to renters, often producing the lowest hourly rates but with more variance in reliability, network quality, and support. Finally, colocation and self-hosting remains an option for organizations large enough to buy hardware outright and rent only rack space and power; by 2027 this can be the cheapest per-GPU-hour path at high sustained utilization, but it demands capital, operational expertise, and a multi-year horizon to amortize. The competitive dynamic across these layers — plus expanding supply of newer Hopper and Blackwell parts — has kept steady downward pressure on H100 pricing throughout 2027. A market-tier comparison lives at https://pulserevops.com/knowledge/gpu-cloud-tiers.
Should you lease on-demand, reserve capacity, or buy?
The right structure follows directly from your utilization curve. The decision tree below captures the standard logic RevOps and infrastructure teams apply when they model GPU spend.
The general principle: if your nodes will run above a break-even utilization threshold — often estimated somewhere in the range where reserved capacity's fixed monthly cost divides down to less than the on-demand rate you'd otherwise pay — reserving wins. Below that threshold, on-demand's flexibility is worth the premium because you avoid paying for idle silicon. Buying only makes sense at very high sustained utilization over a multi-year window, and even then only when you have the capital and the operational depth to run the hardware reliably. Many teams blend approaches: a reserved baseline for steady workloads plus on-demand burst for spikes. See the cost-modeling framework at https://pulserevops.com/knowledge/reserve-vs-ondemand.
What hidden costs turn a cheap H100 lease into an expensive one?
The GPU line item is rarely the whole bill. Data egress is the classic surprise: moving training datasets and checkpoints out of a provider can accumulate charges that, over a long run, rival the compute itself. High-performance storage is another — large-scale training needs a parallel filesystem fast enough to keep eight GPUs fed, and that tier costs far more than commodity object storage. Networking premiums for the InfiniBand fabric that makes multi-node training possible are sometimes quoted separately from the node, so a "cheap node" without fabric may be useless for your actual workload.
Beyond infrastructure, watch for minimum commitments and ramp clauses that bill you for capacity before you can use it, support-tier gating that puts real responsiveness behind a higher plan, and early-termination penalties on reserved contracts. Utilization discipline is its own hidden cost: a reserved node sitting idle is money burned, so the effective price you pay is always the contract cost divided by the useful GPU-hours you actually extract. The lowest advertised rate is frequently not the lowest total cost once these factors are modeled honestly.
How do you negotiate and validate an H100 node lease?
Start by normalizing every quote to a common unit and a common scope — same GPU count, same interconnect, same fabric, same storage and egress assumptions — so you are comparing like for like. Then benchmark against multiple tiers: get a hyperscaler number, a specialized-cloud number, and a marketplace number, and use the spread as leverage. Providers competing for committed multi-year business in 2027 have room to move, particularly on older Hopper inventory as newer accelerators ship. Ask explicitly what is bundled versus billed separately, and get egress, storage, and support in writing.
Validate before you commit at scale. Run a short paid pilot on the exact node type to confirm real interconnect bandwidth, storage throughput, and uptime match the SLA — advertised specs and delivered performance diverge more often than vendors admit. Confirm the node is genuine NVLink-connected HGX rather than a PCIe box if your workload needs it. Finally, structure the contract so utilization and pricing can be revisited: shorter initial terms with renewal options preserve flexibility in a market where prices are still falling. Pair the technical validation with a clear internal utilization forecast so the commitment matches real demand rather than optimistic projections.
Related questions
Is an H100 node the same as a single H100 GPU?
No. A standard H100 cluster node is an 8-GPU HGX server with NVLink and NVSwitch, plus CPUs, memory, storage, and networking. Always normalize quotes to per-GPU or per-node before comparing.
Are H100 lease prices rising or falling in 2027?
Generally falling. Expanding supply of newer Hopper and Blackwell-class accelerators and intense competition among specialized GPU clouds have kept steady downward pressure on H100 rates through 2027, especially for older reserved inventory.
Is it cheaper to lease H100s or buy them?
Buying is usually cheapest per GPU-hour only at very high sustained utilization over multiple years, and only if you have the capital and operations to run the hardware. Below that, leasing wins on flexibility and lower upfront cost.
What network fabric should an H100 training cluster use?
Multi-node H100 training needs a non-blocking, high-bandwidth fabric with RDMA — typically InfiniBand or high-speed RDMA Ethernet. Without it, scaling across nodes stalls, so confirm fabric is included, not billed separately.
Do H200 or Blackwell GPUs make H100 leases obsolete?
Not obsolete — repriced. Newer accelerators pull premium workloads upmarket while H100 capacity becomes a cost-effective option for many training and inference jobs, often at improving lease rates.
FAQ
How many GPUs are in a single H100 cluster node? A standard HGX H100 node contains eight SXM5 H100 GPUs interconnected by NVLink and NVSwitch, paired with high-core CPUs, several terabytes of RAM, local NVMe, and multiple network interfaces. Some providers offer 4-GPU or PCIe variants, so always confirm the exact configuration before comparing prices.
What's the difference between on-demand and reserved H100 leasing? On-demand billing charges by the hour with no long-term commitment, giving maximum flexibility at the highest unit price. Reserved leasing commits you to a term — commonly one to three years — in exchange for a substantial discount. Reserved wins when utilization is high and predictable; on-demand wins for bursty or experimental workloads.
Why do specialized GPU clouds cost less than hyperscalers? Specialized "neocloud" providers exist primarily to rent accelerators, so their whole business optimizes for dense GPU utilization rather than a broad managed-service catalog. That focus lets them offer lower raw compute rates, particularly on committed reservations, though they may bundle fewer surrounding services than a hyperscaler.
What hidden fees should I watch for in an H100 lease? The common surprises are data egress charges, high-performance parallel storage, separately billed networking fabric, minimum-commitment and ramp clauses, support-tier gating, and early-termination penalties. Model total cost of ownership — not the headline GPU rate — and get every ancillary line item in writing before signing.
Does leasing a cheaper node hurt training performance? It can. A PCIe-based node without NVLink, or a node not wired into a high-speed RDMA fabric, will bottleneck large multi-GPU and multi-node training even if the GPUs themselves are H100s. Validate real interconnect and storage throughput in a short pilot before committing at scale.
How do I calculate the break-even point between reserving and on-demand? Divide the reserved contract's fixed cost by your expected useful GPU-hours to get an effective hourly rate, then compare it to the on-demand rate. If your projected utilization pushes the effective reserved rate below on-demand, reserve. Because idle reserved capacity is wasted spend, honest utilization forecasting is essential.
Will H100 lease prices keep dropping after 2027? The trajectory points downward as newer accelerator generations expand supply and push Hopper-class hardware into value pricing, but the pace depends on overall AI compute demand. Structuring shorter initial terms with renewal options lets you capture future declines rather than locking in today's rate.
Can I mix reserved and on-demand H100 capacity? Yes, and many teams do. A reserved baseline covers steady, predictable workloads at the lower committed rate, while on-demand capacity absorbs spikes and experiments. This hybrid model captures most of the reservation discount without paying for idle silicon during quiet periods.
Sources
- NVIDIA H100 Tensor Core GPU
- NVIDIA HGX Platform
- NVIDIA DGX H100 Documentation
- AWS EC2 P5 Instances (H100)
- Google Cloud A3 GPU VMs
- Microsoft Azure ND H100 v5 Series
- NVIDIA InfiniBand Networking
- MLCommons MLPerf Training Benchmarks