How much does a single NVIDIA H100 cluster node cost to lease in 2027?

The two (or more) options compared

The 2027 H100 leasing market presents three distinct procurement paths, each with fundamentally different cost structures and risk profiles. On-demand leasing from hyperscale clouds and specialized GPU providers charges by the GPU-hour or node-hour with zero commitment, offering maximum flexibility at the highest unit price. A single node on AWS P5 instances, for example, might run $28-$35 per GPU-hour for an 8-GPU node, translating to roughly $224-$280 per node-hour, with no termination penalties but no price protection either. This option suits experimental workloads, short-term spikes, and teams that cannot forecast utilization beyond a few weeks.

Reserved capacity contracts—typically one to three years—slash the per-unit cost substantially in exchange for a fixed monthly payment. A one-year reservation on a dedicated 8-GPU H100 node from a specialized GPU cloud in 2027 often lands between $18,000 and $25,000 per node-month, while a three-year commitment can push below $15,000 per node-month. The tradeoff is lock-in: you pay whether you use the node or not, making accurate utilization forecasting critical. Reserved leases dominate for production training pipelines, fine-tuning operations, and inference serving with predictable load patterns.

The third option, hardware purchase with colocation, remains viable only for organizations with significant capital budgets and operational depth. Buying eight H100 GPUs plus the HGX server, networking, and supporting infrastructure in 2027 might cost $250,000-$350,000 upfront, with colocation space, power, and cooling adding $3,000-$6,000 per month. At very high sustained utilization—above roughly 70-80% over three years—this path yields the lowest per-GPU-hour cost, but it demands capital allocation, hardware lifecycle management, and the ability to absorb technological obsolescence risk.

How to decide between them

The decision between on-demand, reserved, and buy-and-colocate hinges on a single variable: your sustained utilization rate over the planning horizon. The mermaid diagram below captures the standard decision logic that RevOps teams apply when modeling GPU infrastructure spend for a single H100 cluster node.

The break-even calculation is straightforward: divide the reserved contract's total cost by the number of GPU-hours you realistically expect to use. If that effective hourly rate falls below the on-demand rate you would otherwise pay, reservation wins. Because idle reserved capacity is pure waste, honest utilization forecasting—not optimistic projections—is the foundation of a sound decision. Many teams adopt a hybrid model: a reserved baseline covering 60-70% of peak demand, with on-demand burst capacity for the remainder.

Concrete numbers behind each option

In 2027, the market has settled into recognizable pricing bands that reflect the competitive dynamics between hyperscalers, specialized GPU clouds, and marketplace aggregators. For a single 8-GPU H100 HGX node with NVLink and InfiniBand fabric connectivity, on-demand hourly rates from the major hyperscalers typically range from $28 to $45 per GPU-hour, or $224 to $360 per node-hour. Specialized GPU clouds undercut this by 20-35%, with on-demand rates around $18 to $28 per GPU-hour for similar hardware, though the surrounding service ecosystem is thinner.

Reserved pricing tells a different story. A one-year commitment on a dedicated 8-GPU node from a specialized provider in a moderate-cost power region—such as the US Midwest or parts of Europe—generally lands between $18,000 and $25,000 per node-month. A three-year reservation can drop to $12,000-$18,000 per node-month, reflecting the provider's ability to amortize hardware over a longer period and the market's expectation of continued price declines. Hyperscaler reserved pricing tends to run 15-25% higher than specialized clouds for equivalent reservations, though the gap narrows when enterprise SLAs and managed services are factored in.

Marketplace aggregators, which match idle capacity from multiple operators, often produce the lowest headline rates—sometimes as low as $12-$16 per GPU-hour for short-term commitments—but with significant caveats. The hardware may be PCIe-based rather than NVLink-connected, the fabric may be shared or best-effort, and support responsiveness can vary wildly. For workloads that genuinely tolerate these constraints, marketplace rates can be compelling, but the risk of performance degradation or capacity unavailability is real.

The buy-and-colocate path requires a capital outlay of roughly $250,000-$350,000 for the node and supporting infrastructure, plus $3,000-$6,000 monthly for colocation space, power, and cooling. At 80% utilization over three years, the effective cost per GPU-hour lands around $4-$6, roughly half the best reserved lease rates. However, this ignores the cost of capital, hardware maintenance, staffing, and the risk that newer accelerators make the H100 obsolete before the depreciation schedule completes.

Implementation details and sequencing

Leasing a single H100 cluster node in 2027 involves a sequence of steps that determine whether you capture value or leave money on the table. The process begins with workload profiling: understand your model's memory footprint, interconnect sensitivity, and training duration to confirm that an 8-GPU H100 node is the right unit of compute. Some workloads benefit from smaller nodes with faster interconnects, while others scale efficiently across larger configurations. The mermaid diagram below maps the implementation sequence from need assessment to live production.

The pilot phase is non-negotiable. Advertised specs and delivered performance diverge often enough that a short paid test—running your actual model or a representative benchmark—reveals whether the node's interconnect bandwidth, storage I/O, and network latency match what the lease promises. A PCIe-based node advertised as "H100" will fail to saturate the GPUs on large multi-GPU training jobs, wasting both time and lease dollars. Confirm the node is genuine NVLink-connected HGX hardware, not a cheaper PCIe variant, and validate that the InfiniBand or RDMA Ethernet fabric delivers the throughput needed for multi-node scaling.

Contract structuring matters as much as hardware validation. In a market where H100 prices are still declining, locking into a three-year reservation at today's rates means you miss future savings. Shorter initial terms—one year with renewal options—preserve flexibility while still capturing the reservation discount. Include provisions to revisit pricing at renewal, and negotiate caps on egress fees and storage costs that can otherwise balloon unpredictably. The revenue impact of a poorly structured lease—idle capacity, overpriced fabric, surprise egress bills—can erode the cost advantage that made leasing attractive in the first place.

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

Related on PULSE

• [How much does AI Infra cost in 2027?](/knowledge/ai445)
• [The 10 Best AI Data Version Control Systems in 2027](/knowledge/ai0460)
• [The 10 Best AI Feature Stores for Production in 2027](/knowledge/ai0459)
• [The 10 Best AI Model Registries in 2027](/knowledge/ai0458)