NVIDIA H100 Price in 2026: Buy, Build or Rent

The short answer

An NVIDIA H100 has no list price. NVIDIA sells most of its data center GPUs to server makers and large buyers on negotiated terms, so every figure in circulation is an analyst estimate, a reseller listing or a buyer’s report. The most cited number is still the one CNBC published in March 2024: $25,000 to $40,000 per chip, according to analyst estimates. In 2026 the range has not collapsed. Retail listings for a new 80GB card cluster around $31,000, used SXM5 modules trade between $22,000 and $27,000 according to Compute Exchange data quoted by IntuitionLabs, and used PCIe cards sit lower, at $16,000 to $20,000.

Almost nobody buys one H100. The unit of purchase is a server with eight of them, and the unit of operation is a rack, a row or a whole hall. A complete eight-GPU HGX H100 system costs $250,000 to $320,000 in 2026, about $285,000 in a typical configuration, according to CloudZero’s market guide. That is roughly $35,600 per GPU once the chassis, CPUs, memory, network cards and storage are counted.

Renting is the other way in, and it is the market we measure. At the Compute Census index of 16 Oct 2026, the median on-demand price of an H100 80GB is $3.49 per GPU-hour, the cheapest accepted offer is $1.30 and the most expensive list price is $12.29. A published 12-month reserved rate stands at $3.26.

The rest of this guide takes each number apart: what a card, a server and a DGX cost, what an H100 costs to keep running per hour, what the rental market charges and why it moves, and where the break-even between buying and renting falls. Every external figure is dated and listed on our Sources page; every rental figure comes from our own price index and updates when the index does.

What one H100 costs to buy

Three products under one name

“H100” covers three different products, and their prices differ for good reasons.

• H100 SXM5. The module that sits on NVIDIA’s HGX baseboard, eight to a board. NVIDIA’s specification sheet lists 80GB of HBM3 memory, 3.35 TB/s of memory bandwidth, NVLink at 900 GB/s and a configurable power limit of up to 700 W. This is the part inside nearly every large training cluster built in 2023 and 2024, and the one most rental offers mean when they say “H100”.
• H100 PCIe. A dual-slot card for ordinary servers. NVIDIA’s product brief gives it 80GB of HBM2e, not HBM3, a peak bandwidth of 2,000 GB/s and a maximum power of 350 W. It costs less, runs cooler and fits more servers, but it moves data more slowly and links to fewer neighbours.
• H100 NVL. A PCIe card with 94GB of HBM3 and 3.9 TB/s, rated at 350 to 400 W, sold mainly for inference. NVIDIA bundles its AI Enterprise software licence with the NVL, which changes the total price of ownership in a way that a card price alone hides.

A buyer comparing quotes therefore has to ask first which of the three is on offer. A “cheap H100” listing is very often a PCIe card being compared with SXM5 prices.

New, used and why the range is so wide

The new-card market is thin because the big buyers do not buy cards. Hyperscalers, neoclouds and enterprises buy complete systems from Dell, Supermicro, HPE, Lenovo and others, and NVIDIA prices the GPU into those systems. What reaches resellers as a loose card is a small share of supply, and its price reflects scarcity as much as cost.

That scarcity peaked in 2023. In April of that year, TechRadar counted H100 cards listed on eBay between $39,995 and just under $46,000, and in late 2023 Dell quoted a 39-week lead time for H100 servers, according to The Register. By April 2024 Dell’s lead time had fallen to eight to twelve weeks. Supply caught up, Blackwell arrived, and secondary prices began to fall.

The used market now carries the H100’s price history. IntuitionLabs, quoting Compute Exchange, gives indicative third-quarter 2026 ranges of $22,000 to $27,000 for used SXM5 modules and $16,000 to $20,000 for used PCIe cards, and puts used units generally at 60 to 70 percent of the new price. Used modules come out of decommissioned clusters with a known service history, or out of failed projects with none. A buyer pays less and inherits the risk of a part that may have run at full power for two years.

Why the price holds

NVIDIA sells into a market where demand has outrun supply for three years. It reported data center revenue of $89.0 billion for the quarter that ended on 26 July 2026, up 117 percent from a year earlier, with Blackwell systems shipping and the Vera Rubin platform in full production. The H100 is no longer the product driving that growth, but its price was set in the same seller’s market, and the used market still prices it off that history rather than off what it cost to make.

What a complete server costs

HGX systems from server makers

A complete H100 machine is an eight-GPU server built around NVIDIA’s HGX baseboard: two server CPUs, one to two terabytes of memory, eight high-speed network adapters for the cluster fabric, local NVMe storage and power supplies sized for the GPUs. CloudZero’s 2026 guide prices a complete eight-way HGX H100 system at $250,000 to $320,000, about $285,000 typical. Reseller listings for Dell’s PowerEdge XE9680 with eight H100 SXM modules run from about $157,000 to $259,000, depending on memory, storage and networking.

That spread is not noise. Two servers with the same eight GPUs can differ by tens of thousands of dollars on the network cards alone, and by more on memory. The GPU is the largest line on the invoice, but it is not the whole invoice.

The DGX H100

NVIDIA’s own system, the DGX H100, is the reference version of that server. SemiAnalysis reported in 2023 that NVIDIA sold the DGX H100 for around $270,000, and one reseller lists a configured 640GB system at $375,000. The value of the DGX lies in the package: a validated design, NVIDIA support and a known path into a DGX SuperPOD cluster.

The physical numbers matter as much as the price. NVIDIA’s user guide lists eight H100 GPUs with 640GB of GPU memory in total, two Xeon 8480C processors, two terabytes of system memory and eight ConnectX-7 InfiniBand ports. The system draws up to 10.2 kW from six 3,300 W power supplies, occupies eight rack units and weighs up to 130.45 kg. NVIDIA’s SuperPOD design guide recommends four DGX H100 systems per rack, which puts a single rack at roughly 40.8 kW, several times what an ordinary enterprise rack was built to carry.

What is missing from a server quote

A server price buys compute. It does not buy a cluster. Training jobs that span more than one server need a low-latency fabric between them, and that fabric is expensive. The Next Platform reported in 2024 that InfiniBand accounts for about 20 percent of the cost of an AI cluster, against 10 percent or less for clusters built on Ethernet. Storage, management networks, cabling and the facility itself come on top.

SemiAnalysis’s model of a 100,000-GPU H100 cluster gives the scale: more than $4 billion of server capital spending and around 150 MW of critical IT power. Few buyers build at that size, but the proportions hold for a single rack. The GPUs are most of the cost, never all of it.

What an H100 costs to run

Electricity

An H100 SXM5 can draw up to 700 W, and the server around it draws more. SemiAnalysis estimates about 575 W of non-GPU server power per H100 in a large cluster, for CPUs, memory, network cards, fans and power conversion. Call it 1,275 W per GPU at the wall of the server.

The facility then adds its own overhead, measured as power usage effectiveness, or PUE: total facility power divided by IT power. Uptime Institute’s 2026 survey found an industry average PUE of 1.52, falling to 1.36 when the average is weighted by capacity, because large sites run more efficiently. Google reports a fleet-wide PUE of 1.09 for 2025. Most H100 owners will sit between the hyperscaler number and the industry average.

Electricity prices finish the calculation. The US Energy Information Administration reports average prices for July 2026 of 14.53 cents per kWh for commercial customers and 9.77 cents for industrial ones. At 1,275 W, a PUE of 1.36 and the industrial price, one H100 costs about 17 cents an hour in electricity while it runs. That is real money across a cluster, but it is a small share of the hourly cost of an H100. The bigger line is the hardware itself.

Hosting

Most buyers do not own a data center that can take a 40 kW rack. They lease space and power in someone else’s facility. CBRE’s North America report for the second half of 2025 gives an average asking rate of $195.94 per kW per month for 250 to 500 kW requirements in primary markets, up 6.5 percent in a year, with vacancy at a record low of 1.4 percent. At 1.275 kW per GPU, that is about 34 cents per GPU per hour, paid whether the GPU is busy or idle. Contracts differ in what that rate includes, so a buyer should read the electricity terms line by line. For comparing colocation offers themselves, wattandrack.com publishes provider prices per kW.

People, spares and money

The remaining costs are harder to put on one line and easy to forget. Clusters need engineers who can debug a failed node at night, spare parts on site, monitoring, security and insurance. Buyers who borrow to fund hardware pay interest; buyers who pay cash give up the return they could have earned elsewhere. None of this appears in a server quote, and all of it appears in the real cost of owning H100s.

What the rental market charges

The index in one table

Renting moves every one of those costs onto the provider. The renter pays one hourly price and walks away when the job is done. The Compute Census index records that price across providers. At the snapshot of 16 Oct 2026, it counts 8 providers with a published H100 price and 9 provider price points, with the ceiling sitting 9.5× above the floor.

The spread is the first thing to understand about the rental market. At the bottom sit marketplace and pooled offers, where capacity comes from many hosts and availability changes by the hour. In the middle sit the neoclouds, specialist GPU clouds that publish a per-hour price for dedicated nodes. At the top sit the hyperscalers’ list prices, which bundle a full cloud platform and are rarely what large customers actually pay after discounts. The median is the most useful single number because it sits where most self-serve buyers land.

How the price has moved

Against 1 Oct 2025, the H100 median has moved +79%. The longer history explains the turn. Through 2024 and 2025, rental prices fell as new capacity came online and Blackwell systems began shipping. AWS cut on-demand prices for its P5 H100 instances by 44 percent from 1 June 2025, with savings plans cut by up to 45 percent. Many observers read that as the start of a long slide.

The slide stopped. SemiAnalysis reported in April 2026 that on-demand capacity was sold out and that H100 contracts were being renewed at the same rates signed two to three years earlier, some for four years through 2028. Its one-year H100 contract index rose from $1.70 per GPU-hour in October 2025 to $2.35 by March 2026. Providers then raised list prices: Nebius lifted its on-demand H100 rate to $4.50 per GPU-hour from 1 October 2026, about 17 percent higher, its second increase in three months according to Silicon UK. AWS now lists Capacity Blocks for its eight-GPU H100 instance at $47.757 per instance-hour, or $5.970 per GPU-hour.

SemiAnalysis attributes the turn to demand that outran the arrival of new capacity: on-demand pools sold out, and older GPUs that were expected to cheapen stayed booked on long contracts. Our own index records the same turn in the median, which is why the history above is worth more than any single quote.

On-demand, reserved and the rest

The index records several kinds of offer, and they are not interchangeable. On-demand capacity can be started and stopped by the hour. Reserved capacity is committed for six months to three years in exchange for a lower rate, and the published 12-month rate sits −7% against the on-demand median at this snapshot. Spot and preemptible capacity is cheaper again, but the provider can take it back. Serverless offers bill by the second for inference. Our floor and median leave spot and preemptible offers out, so a cheap interruptible rate never pulls the headline down.

For a team that needs 512 GPUs for a year, none of these list prices is the real price. Large blocks are quoted individually, and the terms depend on volume, duration, location and how soon the capacity is needed.

Rent or buy: where the break-even falls

Hours of rent that equal one purchase

The simplest test divides the purchase price by the hourly rate. Using CloudZero’s typical eight-GPU HGX system price, about $35,600 per GPU installed, Table 2 shows how many rented hours that money buys at each point of the index, and how many months that is at full and at half use.

The table answers a narrow question. It says how long a renter can run before spending the price of the hardware. It ignores everything an owner spends after the purchase, and everything an owner gets back by selling the hardware later. For a fair comparison, both have to be counted.

The owner’s cost per used hour

Table 3 builds the owner’s side hour by hour. The purchase is written off over three to six years. Hosting is charged on the full power draw every hour of the month, because a lease does not pause when the GPU is idle. Electricity is charged only while the GPU works. The result is the cost of each hour of useful work, at full, 70 percent and 50 percent utilisation, set against the rental index.

Two lessons come out of the table. First: ownership is cheap per hour only when the hardware stays busy. Every idle hour is paid twice by an owner: once in depreciation and once in hosting, while a renter pays nothing for hours not used. The second is that the depreciation period does most of the work. An owner who writes the hardware off over three years carries a much higher hourly cost than one who assumes six, and the right assumption depends on how long the H100 stays useful, which is the subject of the next section.

The table also leaves things out on purpose. If InfiniBand is 20 percent of a cluster’s cost, as The Next Platform reports, the fabric adds a quarter on top of the servers it connects, and the capital line in the table grows by the same share. Staff, spares, insurance and the cost of capital come on top of that, and a resale value at the end comes off. A buyer should fill each of those lines with real quotes before trusting any break-even, including ours.

When buying wins, and when renting does

Buying makes sense when four things are true at once: the workload is steady for years, utilisation will stay high, a facility with enough power is already available, and the team can operate the hardware. Large model developers and providers that resell capacity meet all four. For them, owning is how margin is made.

Renting makes sense when any one of those fails. Research teams with bursty jobs, start-ups whose plans may change in six months, companies testing whether a model is worth deploying, and anyone without a 40 kW rack waiting all pay less by renting, even at a higher hourly price. Reserved contracts sit in between: the buyer commits to a term and gets most of the owner’s price without the owner’s risk.

How long an H100 stays worth owning

What the accounts say

Depreciation policy is where companies state, in audited filings, how long they expect servers to last. Alphabet depreciates servers and network equipment over six years. Meta moved most servers and network assets to 5.5 years from January 2025. Amazon went the other way: from the same date it shortened the useful life of a subset of its servers and networking equipment from six years to five, adding about $1.4 billion of depreciation in a year, mainly in AWS. Microsoft’s filings give servers and network equipment a life of two to six years.

GPU servers sit inside those averages, and the debate is about whether they deserve the full period. Microsoft’s chief executive, Satya Nadella, said in a 2025 interview that he did not want to be stuck with four or five years of depreciation on one GPU generation. A widely repeated 2024 claim that data center GPUs last only one to three years at high utilisation traced back to an anonymous post and was never confirmed.

What the market says

The rental market has so far argued for the longer life. If H100s were obsolete, their rental price would have collapsed when Blackwell systems arrived. Instead, SemiAnalysis reported four-year renewals at original rates, and our own index shows the H100 median moving +79% in the year to this snapshot. A GPU that still earns its original rent three years after delivery is a GPU whose economic life is longer than its owners assumed.

That can change quickly. NVIDIA’s chief executive called himself the “chief revenue destroyer” at GTC 2025 when he compared Hopper with Blackwell, and NVIDIA reported that its next platform, Vera Rubin, was in full production by August 2026. Each new generation does more work per watt, and in a power-limited data center, work per watt eventually decides which GPUs stay switched on.

The supply line is closing

New H100 systems are getting harder to order. Cisco set 19 July 2025 as the last order date for its UCS C885A M8 server with HGX H100 GPUs, with support continuing to 2030, and pointed buyers to configurations with H200 or AMD MI300X GPUs. Distributors have published last-time-buy dates for H100 variants. We found no end-of-life notice from NVIDIA itself for the H100, so the H100 is not formally discontinued, but buyers in 2026 should expect fewer new systems and more refurbished ones.

Buying H100s outside the United States

The H100 is an export-controlled product. US rules adopted in October 2022 barred sales of the A100 and H100 to China without a licence, and in March 2023 NVIDIA responded with the H800, a China version with reduced interconnect bandwidth. The October 2023 rules removed the interconnect test and added a performance-density threshold instead, which closed the route the H800 had used. In April 2025 the US added a licence requirement for the H20, NVIDIA’s later China product, and NVIDIA took a $4.5 billion charge for H20 inventory and purchase commitments in the quarter that followed.

For buyers elsewhere, the rules have shifted more than once. The Biden administration’s AI Diffusion Rule, which would have capped GPU exports to most countries by tier, was due to take effect on 15 May 2025. The Commerce Department rescinded it on 13 May 2025, announcing at the same time new guidance on diversion through third countries, including Malaysia and Thailand. In January 2026 the Bureau of Industry and Security moved licence reviews for H200-class and smaller chips to China and Macau from a presumption of denial to case-by-case review, under conditions on testing and customer checks.

The practical consequence is paperwork. Legitimate sellers ask for end-user and end-use statements, and shipments to some destinations need licences that take weeks. A price that comes without those questions is a warning sign, not a bargain.

Where the H100s run: power at campus scale

An H100 bought in 2026 has to live somewhere with enough power. That is the second series we publish: a census of AI data center campuses, counted from public records. As of 16 Oct 2026, it lists 634 AI campuses in 55 countries. Those under construction add up to 48.3 GW of published capacity, announced projects to another 31.6 GW, and 24 single campuses are rated at a gigawatt or more.

Put the SemiAnalysis estimate next to those numbers. A cluster of 100,000 H100s needs around 150 MW of critical IT power. The median AI campus that publishes a figure in our census is rated at 66 MW, so a cluster of that size fills a large campus on its own. Smaller deployments fit in a single hall, but they compete for the same scarce resource: power that is connected, contracted and available this year rather than in 2029.

That scarcity links the two series. Rental prices rose in 2026 partly because new power arrived more slowly than new GPUs. A buyer who has the hardware but not the power has bought a very expensive inventory problem.

Five mistakes in reading H100 quotes

Most bad H100 decisions start with a comparison that was never like for like. These are the five we see most often in requests that reach us.

Comparing a PCIe price with an SXM5 price

A PCIe card and an SXM5 module share a name and very little else. The SXM5 part has faster memory, a far faster link to its neighbours and twice the power budget. A quote for “H100” without the form factor is not a quote, and a cheap one is usually the PCIe card. Ask for the exact part number before comparing anything.

Comparing a node price with a GPU price

Cloud providers often price the whole eight-GPU instance, while marketplaces and neoclouds price one GPU. AWS lists its H100 Capacity Block at $47.757 per instance-hour, which is $5.970 per GPU-hour once divided by eight. The same confusion runs through hardware quotes, where a server price can be read as a card price by mistake. Normalise every quote to one GPU-hour or one GPU before putting it in a table.

Taking the list price of a hyperscaler as the market

The ceiling of our index is a hyperscaler list price, and it sits 9.5× above the floor. Large customers of those clouds rarely pay list, and smaller buyers rarely need the platform the list price includes. The median is a better guide to what a self-serve buyer pays, and an individual quote for a block is a better guide again.

Forgetting the fabric, the storage and the site

A server quote prices eight GPUs and the box around them. A working cluster also needs switches, cables, storage and a hall with enough power and cooling, and those lines can add a quarter or more to the hardware bill before the first job runs. Renters get all of it in the hourly price; buyers have to price it themselves.

Buying hardware before securing power

The last mistake is the most expensive. A 40 kW rack needs a facility that can deliver and cool 40 kW today, not in eighteen months. Vacancy in primary North American markets was 1.4 percent at the end of 2025, according to CBRE. Buyers who order servers first and look for a home second can wait months with idle hardware that loses value every week. Secure the site, or rent until the site is ready.

How to source H100 capacity in 2026

Decide what you are buying

Start with the job, not the GPU. Training a large model across many servers needs SXM5 systems on a fast fabric. Fine-tuning and inference often run well on PCIe or NVL cards, or on newer GPUs that do more per dollar. Compare the H100 against its neighbours before committing.

Memory is often the deciding line. The H200 carries 141GB against the H100’s 80GB, so a model that needs two H100s may fit on one H200, and the price per gigabyte of memory in Table 5 can tell a different story from the price per GPU.

Questions for a seller of hardware

• Which variant is it: SXM5 on an HGX board, PCIe, or NVL?
• New, refurbished or used, and with what service history and warranty?
• What exactly is in the server: CPUs, memory, network cards, storage, power supplies?
• What are the lead time, delivery terms and export paperwork for your country?
• Where will it be hosted, and is the power there contracted today?

Questions for a provider of rented capacity

• Is the price for dedicated nodes, shared pools, spot or serverless capacity?
• What is the minimum term, and what are the exit terms?
• Which fabric links the nodes, and is it non-blocking at your cluster size?
• What happens when a node fails: replacement time, credits, spare capacity?
• Where is the data stored, and what does it cost to move it out?

We handle both routes. GPU capacity is available for rent and hardware for purchase, through different partners, and the terms are agreed individually for each request: there is no fixed commission and no fixed minimum. Tell us the variant, quantity, location and timing in the form below, and we come back with what is available and on what terms.