An NVIDIA H100 costs $25,000 to $30,000 to buy as a single SXM5 GPU, $250,000 to $320,000 as an 8-GPU HGX server, and $3.93 per GPU-hour to rent on average across 26 cloud providers, as tracked by the Mercatus GPU Index on October 6, 2026.
None of those numbers is the whole answer. The same H100 rents for $1.79 an hour from one provider and more than $10 from another. An owned H100 costs about $1.77 an hour all-in at 70 percent utilization and about $2.79 at 40 percent. Which price applies to you depends on whether you buy or rent, who you rent from, and how hard you run the card.
This guide covers each price, shows the math behind the owned cost per hour, and gives you the formula to plug in your own assumptions. For the cluster-scale version, see Total Cost to Own 100 H100 GPUs. For the generation comparison, see A100 vs H100 vs H200.
H100 price at a glance
| What you are pricing | Price (2026) |
|---|---|
| H100 SXM5 GPU, new from OEM | $25,000 to $30,000 |
| H100 PCIe card, new | $22,000 to $27,000 |
| 8-GPU HGX H100 server | $250,000 to $320,000 |
| Used H100, 18 to 30 months old | $22,500 to $25,500 |
| Refurbished H100 | $18,000 to $22,000 |
| Cloud rental, index average | $3.93 per GPU-hour |
| Cloud rental, range across providers | $1.79 to $12.52 per GPU-hour |
| Owned, all-in at 70% utilization | About $1.77 per GPU-hour |
Rental figures are on-demand rates from the Mercatus GPU Index as of October 6, 2026. Used and refurbished figures are from H100 resale value.
The sticker price: what an H100 costs to buy
There are three ways to acquire an H100, and the headline price differs for each:
| Acquisition path | Price range (2026) | Notes |
|---|---|---|
| Single H100 SXM5 (OEM, GPU only) | $25,000 to $30,000 | Rarely sold standalone. Most go through OEM systems |
| 8-GPU HGX H100 server (Supermicro, Dell, HPE) | $250,000 to $320,000 | $31,000 to $40,000 per GPU including server, NVLink, and networking |
| H100 PCIe (single card, slot form factor) | $22,000 to $27,000 | 10 to 15 percent cheaper, but no NVLink, so single-GPU only |
For most institutional buyers, the relevant number is the 8-GPU HGX system price, because H100s are deployed in 8-GPU nodes. A typical integrated quote is about $285,000. The full breakdown of what is inside that number is in H100 server price.
Why does the same H100 list at different prices?
- Volume tier. OEMs price by quantity. A 100-server order beats a 1-server quote by 8 to 12 percent in 2026.
- Region. EU and APAC pricing typically runs 5 to 10 percent above US prices because of import duties and logistics.
- OEM choice. Supermicro is consistently cheapest. Dell and HPE charge premiums for support tiers.
- Service contract. GPU-only pricing strips support. Full service, with 4-hour replacement and on-site engineering, adds 8 to 15 percent.
These are hardware numbers. They do not include power, colocation, networking outside the box, or operations. Those come later in this guide.
The cloud price: what the same H100 actually rents for
Most teams do not buy H100s. They rent, and this is where the price varies most.
The Mercatus GPU Index puts the volume-weighted average H100 on-demand price at $3.93 per GPU-hour across 26 providers on October 6, 2026, with a range of $1.79 to $12.52 after outlier filtering. The lowest listed price at each provider in the index snapshot falls into three groups:
| Provider group | Providers | Lowest listed $/GPU-hr |
|---|---|---|
| Specialist and marketplace clouds | 17 | $1.66 to $3.50 |
| Mid-priced clouds | 5 | $3.77 to $5.99 |
| Hyperscalers (Azure, Google Cloud, Oracle) | 3 | $6.98 to $10.00 |
The same H100, with the same silicon and the same 80GB of HBM3, lists at about 7 times the price at the top of the index range as at the bottom. A single H100 running around the clock for a 730-hour month costs about $1,307 at $1.79, $2,869 at the index average, and $7,300 at $10.00.
Two more things the index shows.
The H100 rental price has been rising. The index average was $3.70 ninety days ago and is $3.93 now, up 6.1 percent, with the period high set on October 6. A GPU that launched in 2022 is getting more expensive to rent in late 2026. Demand for inference capacity is setting the rate, not the age of the silicon.
A year of average-priced rental costs about as much as the hardware. One H100 at $3.93 an hour for a full year is about $34,400, in line with the $31,000 to $40,000 all-in purchase price per GPU. At a specialist rate of $1.99 it is about $17,400. Which rental price you compare against decides whether buying looks obvious or marginal.
Reserved and committed-use pricing runs lower than these on-demand figures, and spot pricing runs lower still but can be interrupted. For the price at every provider, see GPU rental prices. For live figures, see the H100 page on the Mercatus GPU Index.
Why the cloud price is what it is
Every provider renting an H100 is covering the same underlying costs: hardware amortization, power, colocation, networking, and operations. Those costs do not differ by a factor of seven between providers. What differs is what surrounds the GPU.
Hyperscaler pricing bundles enterprise support, compliance, and integration with the rest of the platform. A buyer who needs those things is paying for them. A buyer who only needs the GPU is paying for them anyway. Specialist providers sell the GPU and little else, and marketplaces resell spare and community capacity at lower rates still.
The practical takeaway: price the specialist tier first, then add the hyperscaler premium only for a reason you can name. The full analysis is in why GPU prices differ for the same hardware.
What an owned H100 actually costs per hour
If you buy an H100, your cost per GPU-hour depends on five inputs:
- Hardware capex. $31,000 to $40,000 per GPU including the server.
- Resale value. What the hardware sells for when you are done with it.
- Utilization. The share of hours actually serving a workload.
- Power cost. Varies heavily by region, from $0.06 to $0.25 per kWh.
- Colocation and operations. Varies by deployment.
Worked example for one H100 in an 8-GPU node, at $33,000 all-in capex (typical of a multi-system order), in a US Tier 3 colocation facility, 70 percent utilization, over 3 years:
// text
Hardware: $33,000 capex
- $8,000 resale value at 36 months (about 25% retained)
= $25,000 net depreciation
/ (3 years x 8,760 hours x 0.70 utilization = 18,396 hours)
= $1.36 per GPU-hour
Power: 0.7 kW x 1.4 PUE x $0.10 per kWh = $0.10 per GPU-hour
Colo: $150 per kW per month x 1.0 kW / 730 = $0.21 per GPU-hour
Ops: failures, replacements, monitoring = $0.10 per GPU-hour
Total: about $1.77 per GPU-hour all-in
Utilization moves that number more than anything else. Holding the other inputs fixed:
| Utilization | All-in cost per GPU-hour |
|---|---|
| 90% | $1.46 |
| 70% | $1.77 |
| 50% | $2.31 |
| 40% | $2.79 |
At 70 percent utilization, an owned H100 costs less than half the index average rental price and sits level with the cheapest specialist providers. At 40 percent, it costs more than the specialist tier charges, and renting wins. The utilization side of this is covered in GPU utilization.
Resale value is the second lever. The 25 percent assumption above is a middle case. Used H100s that are 36 months old resell for anywhere from 40 to 75 percent of list depending on how fast Blackwell supply arrives. At 40 percent retained, the all-in cost in the example drops to about $1.48 per GPU-hour. At 15 percent, it rises to about $1.92. See H100 resale value and H100 depreciation.
The general formula:
// text
Effective $ per GPU-hour =
(Capex - Resale value) / (Lifespan hours x Utilization)
+ Power $ per hour
+ Colo $ per hour
+ Ops $ per hour
Plug in your own numbers. The two variables that move the answer most are utilization and resale value.
When does buying a single H100 make sense?
Almost never. Single-H100 ownership is rarely the right choice in 2026, and the reasons are structural:
- Lumpy capex. $33,000 is a meaningful capital commitment for one GPU, and finance teams scrutinize capital purchases harder than operating spend.
- Operational overhead. A single H100 needs a colocation contract, a server, networking, monitoring, and replacement parts. The operational surface is the same as for ten H100s, just amortized worse.
- No utilization smoothing. A single GPU runs the workload you have or sits idle. With cloud, you pay for what you use.
- Resale risk. An intact 8-GPU node is far easier to sell than a single card.
It makes sense in a narrow band of cases:
- Sustained utilization of 90 percent or more with high workload predictability
- Latency-sensitive workloads where sitting next to your data is critical
- Compliance constraints, such as data residency, that prevent cloud use
- You already operate a colocation footprint and adding one GPU is marginal
Outside those cases, renting from the specialist tier is usually the better answer for a single GPU. At cluster scale the economics change. See Total Cost to Own 100 H100 GPUs, Should you buy or rent GPUs, and GPU ROI.
What if you already own H100s and have spare capacity?
Most owners do not run their fleet at 90 percent utilization. If your H100s sit idle part of the time, you are paying the full ownership cost, capex amortization plus power, colocation, and operations, whether the cards are working or not. The table above shows what that does to cost per hour.
Idle capacity has commercial value. Selling it as inference capacity recovers cost without disrupting your primary workload, and every sold hour lowers your effective cost on the hours that matter. To list capacity on Mercatus, see Sell capacity.
How H100 cost translates to what buyers pay per token
If you buy tokens instead of GPUs, the H100 rental rate is the floor under your price. A provider selling inference has to earn back the GPU-hour in tokens.
The arithmetic is simple. At the index average of $3.93, an 8-GPU H100 node costs $31.44 an hour. To break even selling output tokens at $0.60 per million, that node has to produce about 52 million output tokens an hour, roughly 14,600 tokens a second sustained. At $1.98 per million, the requirement drops to about 4,400 a second.
That is why the cheapest hosts of open-weight models are the ones renting at the bottom of the range or running owned hardware at high utilization, and why the same model carries different prices at different hosts. The token side of this is in why token prices differ.
For day-to-day pricing:
- GPU prices: Mercatus GPU Index
- Token prices: Mercatus Token Index
Frequently Asked Questions
How much does an H100 GPU cost to buy in 2026?
NVIDIA H100 SXM5 OEM pricing ranges from $25,000 to $30,000 in 2026, depending on volume and supplier. The PCIe variant is 10 to 15 percent cheaper. 8-GPU HGX servers cost $250,000 to $320,000 fully built, about $31,000 to $40,000 per GPU including server overhead.
How much does it cost to run an H100 per hour?
Renting: $3.93 per GPU-hour on average across 26 providers as of October 6, 2026, with a range of $1.79 to $12.52. Owning: about $1.77 per GPU-hour all-in at 70 percent utilization on a 3-year horizon, including power and colocation, rising to about $2.79 at 40 percent utilization.
Why do H100 prices differ so much across cloud providers?
Hyperscalers bundle support, compliance, and platform integration into the rate. Specialist providers sell the GPU and little else. The hardware is identical. The rate reflects what surrounds it.
Should I buy an H100 or rent one?
For a single H100, almost always rent. For a fleet, it depends on utilization and on which rental price you compare against. Owned cost at 70 percent utilization is about $1.77 per GPU-hour, level with the cheapest specialist providers and less than half the index average.
How fast does an H100 depreciate?
Slower than typical enterprise hardware. Cards 18 to 30 months old resell at 75 to 85 percent of original list. At 36 months the range widens to 40 to 75 percent depending on Blackwell availability, and the floor at 60 months or more is roughly 25 to 35 percent.
Are H100 rental prices going down?
Not over the last 90 days. The Mercatus GPU Index average rose from $3.70 to $3.93, up 6.1 percent, with the period high on October 6, 2026.
What is the cheapest way to access an H100 right now?
Specialist and marketplace providers list on-demand H100s from $1.66 to $2.50 per GPU-hour. Spot capacity runs lower for batch workloads that tolerate interruption, and reserved contracts run lower for steady production workloads.
What about used or refurbished H100s?
Refurbished H100s from data center decommissioning sell for $18,000 to $22,000 in 2026, typically with a 12-month limited warranty.
Can I sell capacity on H100s I already own?
Yes. See Sell capacity.
Methodology
Rental pricing is from the Mercatus GPU Index and its H100 page as of October 6, 2026: on-demand rates in USD per GPU-hour, volume-weighted average with outlier filtering, 26 providers tracked and 25 in the provider snapshot. The lowest listed price in the snapshot (1.79) because the range applies outlier filtering. OEM hardware pricing reflects public quotes from Supermicro, Dell, and HPE and is unchanged from the earlier version of this page. Resale and depreciation figures are from the Mercatus H100 resale value and H100 depreciation pages. Power, colocation, and operations estimates are derived from typical US Tier 3 colocation contracts in 2026. The ownership example holds power, colocation, and operations flat across utilization levels. Token break-even figures describe required throughput, not measured throughput. Last verified: 2026-10-08.