The NVIDIA H100 GPU has two prices in 2026: what it costs to own and what it costs to rent. Buying a single card runs $25,000-$35,000 on channel listings checked on 20 Jul 2026, before you add a server to put it in. Renting the same silicon starts around $2.39/hr and runs to $12.29/hr per GPU depending on where you rent it, a spread of roughly 5x for identical hardware. You can check the live rate for H100 on Spheron at any time; this post breaks down where every major provider lands, why the spread is so wide, and when buying actually beats renting.
TL;DR: H100 GPU Price by Provider
Spheron's two rows are live rates, pulled when this page was last regenerated. The rest are each provider's own published figure, checked on 20 Jul 2026, listed roughly cheapest to dearest.
| Provider | Tier | Per-GPU $/hr | Notes |
|---|---|---|---|
| Spheron | Spot | $2.04/hr | Interruptible, can be reclaimed |
| Spheron | On-demand | $2.39/hr | Per-minute billing, single-GPU granularity |
| Lambda Labs | On-demand | ~$2.49 | Region and config dependent |
| Scaleway | H100 SXM on-demand | ~$2.87-3.31 | Paris and Warsaw, SecNumCloud positioning |
| Runpod | Secure Cloud PCIe / SXM5 | ~$2.89 / ~$2.99 | Community Cloud dips lower, host-dependent |
| Hyperbolic | On-demand | ~$3.19 | Repriced up from $1.50 within ten weeks |
| Nebius | H100 on-demand | ~$3.85 | Preemptible ~$2.15 |
| Crusoe | On-demand, single GPU | ~$3.90 | Single-GPU granularity |
| CoreWeave | On-demand, 8-GPU bundle | ~$6.16 | No single-GPU option; spot ~$2.44-$2.46 |
| AWS | p5 on-demand | ~$6.88 | 8-GPU instance |
| Oracle OCI | BM.GPU.H100.8 | ~$10.00 | Flat rate, service-limit approval required |
| GCP | A3 High on-demand | ~$10.98 | Spot ~$3.69; 1-yr CUD ~$8.78 |
| Azure | ND H100 v5 on-demand | ~$12.29 | Spot ~$2.25-$3.69, quota process applies |
Pricing fluctuates based on GPU availability. Spheron rates above are live as of 09 Sep 2026; other providers reflect their own published rates, checked on 20 Jul 2026, and may have changed. Check current GPU pricing → for live rates.
The headline finding: the same H100 costs several times more on Azure than it does on a marketplace provider. None of that spread is silicon. All of it is packaging, minimum purchase size, and how much operational overhead the provider prices in.
NVIDIA H100 Specs: What You Are Paying For
Before comparing rates, it is worth knowing what separates an H100 from the cards above and below it, because the spec sheet explains most of the price tiering.
| Spec | H100 SXM5 | H100 PCIe | H100 NVL |
|---|---|---|---|
| Memory | 80GB HBM3 | 80GB HBM2e | 94GB HBM3 |
| Memory bandwidth | 3.35 TB/s | 2.0 TB/s | 3.9 TB/s |
| TDP | 700W | 350W | 400W per card |
| Interconnect | NVLink 900GB/s | PCIe 5.0 | NVLink bridge |
| FP8 Tensor | ~3,958 TFLOPS | ~3,026 TFLOPS | ~3,341 TFLOPS |
The Hopper architecture's Transformer Engine is the reason H100 pricing held up so long against Blackwell. It runs FP8 natively with per-layer scaling, which roughly doubles throughput over the A100's FP16 path on transformer workloads without the accuracy loss that naive quantization causes. Our H100 specs deep dive covers the full architecture, and the A100 vs H100 comparison has the generational benchmarks.
What an H100 Costs to Buy
The card itself lists between $25,000 and $35,000 on channel listings checked on 20 Jul 2026, with SXM5 modules at the top of that range and PCIe cards toward the bottom. That number is the entry ticket, not the bill. A production deployment adds the host server, NVLink or InfiniBand networking if you scale past one card, redundant power, cooling capable of handling 700W per GPU, and the operational time to keep drivers and firmware current.
The second-hand market has softened prices as Blackwell supply ramps, but H100 resale values still hold better than most enterprise hardware because inference demand keeps absorbing used cards. If you are weighing a purchase seriously, our GPU rent vs buy TCO breakdown runs the full ownership model, including depreciation and utilization scenarios.
What an H100 Costs to Rent
The rental market splits into three bands.
Marketplace and neocloud rates (roughly $2-$4/hr on published rate cards checked on 20 Jul 2026). This is where single-GPU, per-minute rentals live. Spheron prices H100 at $2.39/hr on-demand and $2.04/hr spot. Worth checking both before you book: the two tiers track different pools of capacity, so neither is reliably the lower one. Runpod's Secure Cloud sits at $2.89-$2.99/hr; the full tier structure is in our Runpod H100 pricing breakdown. Lambda's on-demand rate lands around $2.49/hr in supported regions, and Nebius prices on-demand at roughly $3.85/hr with preemptible capacity near $2.15/hr. Crusoe sells single GPUs at about $3.90/hr, which matters more than the headline number if you only need one card.
This band is also the most volatile. Hyperbolic moved its H100 rate from $1.50/hr to $3.19/hr inside ten weeks in 2026, more than doubling, which is a useful reminder that a cheap posted rate is not a commitment. Marketplace pricing tracks capacity, and capacity moves with model release cycles.
European and sovereign options ($2.87-$3.31/hr, checked on 20 Jul 2026). Teams with a data-residency requirement pay a smaller premium than most expect. Scaleway runs $2.87-$3.31/hr from Paris and Warsaw with SecNumCloud positioning for regulated workloads. Our OVHcloud H100 pricing breakdown covers the other major EU sovereign option. Check the operating status of any smaller EU provider before you commit to it: several of the cheap sovereign options quoted in 2025 roundups have since wound down.
Specialist clouds with node minimums ($6-$10/hr effective, checked on 20 Jul 2026). CoreWeave prices an HGX H100 node at $49.24/hr, which is $6.16 per GPU whether you need one GPU or eight. Our CoreWeave GPU pricing analysis covers the full tier structure and the 2026 repricing. Oracle's flat $10/GPU/hr sits in the same band.
Hyperscalers ($6.88-$12.29/hr on their own published pricing, checked on 20 Jul 2026). AWS p5 at ~$6.88/hr is the cheapest of the big three; the AWS H100 pricing guide covers capacity blocks and reserved options. GCP's A3 High runs ~$10.98/hr on-demand, and Azure's ND H100 v5 tops the chart at ~$12.29/hr, with quota queues on top; see the Azure H100 pricing comparison for the hidden-cost math. Egress fees and storage add 5-15% to real hyperscaler bills.
For how these H100 numbers sit inside the wider market, including H200 and B200 rates, the GPU cloud pricing comparison tracks all tiers across 5+ providers.
The Node Minimum Is the Hidden Cost
Per-GPU rates make providers look comparable when they are not. CoreWeave at $6.16/GPU reads like an ordinary mid-market rate until you notice it only sells 8-GPU nodes: the smallest possible bill is $49.24/hr whether you need one card or eight. A single Spheron card is $2.39/hr. For a fine-tuning job that fits on one or two GPUs, the node minimum costs you more than the hourly rate does.
So ask a different question. Not "what is your per-GPU rate" but "what is the smallest amount I can rent, and for how long". Node minimums, billing increments, and contract terms shape the real bill more than the posted figure ever does.
SXM5 vs PCIe vs NVL: Why the Form Factor Moves the Price
The cheapest "H100 price" you see quoted is almost always a PCIe card. PCIe tops out at lower power and lacks the 900 GB/s NVLink fabric of SXM5 boards, which is why SXM5 carries a premium over PCIe. The NVL variant, two bridged cards aimed at inference with 94GB of HBM3, currently rents on Spheron as spot capacity only. If your workload is single-GPU inference, PCIe is usually the right buy; multi-GPU training wants SXM5. The H100 form factor guide covers when each one pays off.
Buy vs Rent: The Break-Even Math
Take the mid-range purchase price of $30,000 for one SXM5 card, from channel listings checked on 20 Jul 2026. At Spheron's on-demand rate of $2.39/hr, that money buys several thousand GPU-hours, on the order of a year of running the card 24/7. And that comparison is generous to ownership, because it prices the card at zero for chassis, networking, power, cooling, and operations.
The utilization question decides it. A card you own earns its keep only while it runs. Rented capacity bills only while you use it. Training teams with bursty schedules, inference workloads that scale with traffic, and anyone still iterating on model size will come out ahead renting. The case for buying starts at sustained utilization above roughly 60% for multiple years, and even then the H100-specific risk is generational: Blackwell cards already undercut H100 on cost-per-token for many inference workloads, which pressures the resale value your TCO model depends on.
What Moves H100 Prices Next
Two forces push in opposite directions. Blackwell supply growth pulls H100 rates down: as B200 capacity spreads, providers reprice Hopper to keep it attractive. Inference demand pushes back up: H100s remain the workhorse for production serving, and spot markets tighten whenever a model launch spikes demand. The Hyperbolic repricing above shows how fast that can move in the wrong direction for buyers.
The practical takeaway is not to time the market but to avoid locking a long commitment at today's rate for hardware whose price trend is downward. Per-minute billing with no contract keeps that option open. If you are choosing between generations rather than providers, the H100 vs H200 comparison covers where the extra HBM3e earns its premium.
H100 SXM5, PCIe, and NVL are all live on Spheron with per-minute billing, from $2.39/hr as of 09 Sep 2026, no quota queue and no 8-GPU minimum.
Frequently Asked Questions
A single H100 card lists in the $25,000-$35,000 range on channel listings checked on 20 Jul 2026, depending on form factor (PCIe vs SXM5) and channel. That figure covers the card only: a usable deployment adds server chassis, networking, power, and cooling on top. Most teams evaluating a purchase should run the break-even math against renting first, because at current cloud rates a card has to run near-continuously for the better part of a year before ownership wins.
As of 09 Sep 2026, H100 rental runs from about $2.39/hr to $12.29/hr per GPU depending on provider and tier. Spheron lists H100 at $2.39/hr on-demand and $2.04/hr spot; on a marketplace the two tiers are quoted from different pools and move independently, so read both rather than assuming which one is lower. On their own published rates, checked on 20 Jul 2026, Runpod Secure Cloud runs about $2.89-$2.99/hr, Nebius about $3.85/hr, CoreWeave works out to $6.16/GPU-hr in mandatory 8-GPU bundles, AWS p5 is about $6.88/hr, GCP A3 High about $10.98/hr, and Azure ND H100 v5 about $12.29/hr. Rates move with availability, so check live pricing before budgeting.
Three reasons. First, form factor: SXM5 boards with NVLink command more than PCIe cards. Second, packaging: hyperscalers sell H100s inside managed instances with egress fees and quota processes priced in, while marketplaces sell closer to bare metal. Third, purchase model: on-demand costs more than spot or interruptible capacity, and some providers only sell 8-GPU nodes, which raises the effective minimum spend even if the per-GPU rate looks similar.
Rent, for most teams. A $30,000 card, mid-range on channel listings checked on 20 Jul 2026, at Spheron's on-demand rate of $2.39/hr buys several thousand GPU-hours, on the order of a year of 24/7 utilization, before counting the server, power, cooling, and the engineer who babysits it. Ownership only wins with sustained multi-year, high-utilization workloads. If your utilization is bursty or below roughly 60%, renting is the cheaper path.
The H100 pairs 80GB of HBM3 (94GB on the NVL variant) with 3.35 TB/s of memory bandwidth on SXM5, a 700W TDP, and fourth-generation Tensor Cores with a Transformer Engine that runs FP8. Against the A100 it delivers roughly 3x the training throughput on large language models and up to 30x inference speedup on some workloads. The 900 GB/s NVLink fabric on SXM5 boards is what makes multi-node training practical, and it is the single biggest reason SXM5 rents at a premium to PCIe.
