The short answer
A single H100 80GB PCIe card last listed at $32,411.99 at CDW as retrieved on 21 September 2026, but the listing is marked Discontinued, so that is a last-seen price and not an offer. A complete 8-GPU H100 server last listed between $234,200 and $440,619.75 depending on the build, which is $29,275 to $55,077 per GPU when you divide by eight, with the CPUs, memory and network cards included. Renting one costs $2.50 per GPU-hour on Hyperstack (checked ) right now. Buying beats that only if the GPU stays busy most of the month, and the break-even section below turns that into one division.
What the card costs
NVIDIA does not publish list prices for its data center GPUs. CNBC reported on 19 March 2024 that what a buyer pays depends on volume and on whether they buy a complete system from NVIDIA or through Dell, HP, Supermicro and similar vendors. Every H100 "price" is therefore a reseller listing, an analyst estimate or a reported transaction.
Reseller listings on 21 September 2026
CDW listed the PNY H100 80GB PCIe card (NVH100TCGPU-KIT) at $32,411.99 on 21 September 2026, with stock status "Discontinued". CDW listed the PNY H100 NVL 94GB card (NVH100NVLTCGPU-KIT) at $32,382.99 on the same date, also "Discontinued". Both are card-only prices, and neither can be ordered.
Why the listings are discontinued
Cisco announced end-of-sale for its UCS C885A M8 servers with HGX H100 on 24 April 2025, with a last order date of 19 July 2025. Reseller Viperatech relayed an NVIDIA end-of-life notice on 17 June 2025 for the H100 NVL: final forecast deadline 1 September 2025, last-time-buy 1 November 2025, successor H200 NVL. We did not find the NVIDIA notice itself, only the reseller relay.
The Hopper card a reseller can still quote in late 2026 is the H200 NVL. Our NVIDIA GPU prices page covers that listing, and H100 vs H200 covers the difference.
The list price never really fell
The first public H100 price appeared in April 2022, when Japanese retailer GDep Advance took pre-orders at a base price of 4,313,000 yen, or $32,955 before tax, as Tom's Hardware reported on 29 April 2022. In April 2023, CNBC found at least eight H100s on eBay asking $39,995 to just under $46,000. Silicon Data, which analysed listings from June 2024 to December 2025, described new retail listings as "typically ranging from $25,000 to $40,000" and flat (30 January 2026).
The CDW last-listed price sits within 2% of that April 2022 base price. New-card listings stayed in the same band. What moved was resale.
Used cards
Silicon Data reports that the median resale listing fell from the low $40,000s in mid-2024 toward the mid $20,000s by late 2025. Used H100s listed at a median of about 61% of the new price in 2024 and about 69% in 2025. Refurbished units held about 85% and then 84%. These are listing prices, not closed sales, and we found no reliable source of completed-sale data.
What a server costs
New SXM modules are sold on HGX baseboards inside complete systems, not as loose parts. The realistic purchase unit is an 8-GPU server. The per-GPU column is nothing more than the price divided by eight.
| System | Price | Divided by 8 | Source and date | Status |
|---|---|---|---|---|
| H100 server, hyperscaler volume, server only | about $190,000 | $23,750 | SemiAnalysis estimate, 20 August 2025 | Analyst estimate |
| Same, with storage, networking and other items | about $250,000 | $31,250 | SemiAnalysis estimate, 20 August 2025 | Analyst estimate |
| Supermicro SYS-521GE-TNRT, 8x H100 80GB PCIe, no NVSwitch | $234,200 | $29,275 | Newegg, 21 September 2026 | Out of stock, last listed |
| Supermicro HGX system, 8x H100 80GB, dual Xeon 8490H, 2TB DDR5 | $306,672 | $38,334 | Newegg, 21 September 2026 | Out of stock, last listed |
| DGX H100 640GB, education SKU, 5 years support | $273,056.55 | $34,132 | ShopBLT, 21 September 2026 | Education pricing |
| DGX H100, 8x H100, dual Xeon 8480C, 2TB RAM | $440,619.75 | $55,077 | Broadberry US, 21 September 2026 | Default configured reseller price |
DGX H100 price
NVIDIA never announced a list price for the DGX H100. SemiAnalysis wrote on 29 May 2023 that "Nvidia sells DGX H100's at around $270,000", which is $33,750 per GPU. That is an analyst estimate. The ShopBLT row is an education SKU, discounted against commercial pricing. The Broadberry row is a default configured reseller price, and reseller quotes are negotiable.
Why the per-GPU figure misleads
Dividing by eight tells you what you paid per GPU. It does not tell you what a GPU costs, because you cannot buy one eighth of the machine.
NVIDIA's DGX H100 user guide lists what is in the box: eight H100 80GB GPUs, two Xeon Platinum 8480C CPUs, 2TB of system memory, four NVSwitch chips linking the GPUs, eight ConnectX-7 network cards for the cluster fabric, ten NVMe drives and six 3.3 kW power supplies. The same guide states that the system "is not shipped with network cables or adaptors". Switches, optics, cables, shared storage and racks are all extra. SemiAnalysis puts those extras at about $60,000 per server for a hyperscaler, which is the gap between its $190,000 and $250,000 figures.
The fabric shows up in the price. In the two Newegg listings above, the PCIe build is $72,472 cheaper than the HGX build ($306,672 minus $234,200). The builds also differ in CPUs and storage, so not all of that gap is the fabric. The two are not equivalent for multi-GPU training, because the PCIe server has no NVSwitch.
List prices mislead in the other direction. Lenovo's 2025 TCO paper (23 May 2025) priced a ThinkSystem SR675 V3 with eight H100 NVL cards at about $833,806 "considering no sales discounts". That is a list price, not a street price.
Prices turned up in 2026
If you planned to wait for hardware to get cheaper, the 2026 evidence points the other way.
- Tom's Hardware reported on 23 August 2026, citing Bloomberg, that NVIDIA told some of its largest customers that prices of servers with its AI chips will rise by more than 15% in many cases, on Grace Blackwell and Vera Rubin systems shipping from early 2027. NVIDIA has not confirmed it.
- CoreWeave CEO Mike Intrator said on the company's Q2 2026 earnings call that "pricing for prior generation SKUs is at or above where it was years ago" (Tom's Hardware, 12 August 2026).
What renting costs today
The cheapest on-demand H100 across the providers we track is $2.50 per GPU-hour on Hyperstack (checked ). The H100 comes in SXM, PCIe and NVL variants, and they do not rent for the same price:
| Variant | VRAM | Cheapest $/GPU-hr | Provider | Providers in stock |
|---|---|---|---|---|
| H100 PCIe | 80 GB | $2.50 | Hyperstack | 7 |
| H100 SXM5 | 80 GB | $2.89 | Massed Compute | 5 |
| H100 NVL | 94 GB | $3.11 | Massed Compute | 2 |
Hourly prices hide the monthly bill. This table multiplies the live price out to a day and to a 730-hour month:
| GPU | $/GPU-hr | Per day (×24) | Per month (×730) | Provider |
|---|---|---|---|---|
| H100 | $2.50 | $60.00 | $1,825 | Hyperstack |
Two things matter before you compare that with a purchase. First, contract rents have turned up. The SemiAnalysis GPU index (retrieved 21 September 2026) tracks the H100 one-year contract range. It fell from the first half of 2023 to a low in October 2025. By April 2026 the low end of the range was about 45% above the October 2025 low end, and the high end about 38% above. The top of the April 2026 range equals the bottom of the first-half 2023 range. The index had no rows after April 2026 as retrieved, and its on-demand column showed "Sold Out" for February, March and April 2026. For current numbers, use the daily GPU price index and the H100 rent page instead of any figure quoted in an article.
Second, older GPUs are holding their value in rental. CNBC reported on 14 November 2025 that a batch of 2022-vintage H100s at CoreWeave came off an expired contract and was "immediately booked at 95% of their original price".
Break-even, step by step
The rent vs buy GPU calculator does these sums with live prices filled in. Do it once by hand so you know what drives the answer.
Step 1: pick a purchase cost per GPU
Take a row from the server table. For a small buyer, $38,334 (the Newegg HGX listing) is a reasonable middle case. Use $31,250 only if you buy at hyperscaler volume. Use $55,077 if you want a DGX with NVIDIA support at reseller list.
Step 2: add colocation
CBRE's H2 2025 report (25 February 2026) puts the average asking rate for a 250 kW to 500 kW requirement in primary North American markets at about $195 per kW per month. NVIDIA's user guide gives the DGX H100 a maximum power of 10.2 kW, and colocation is contracted on maximum power.
10.2 kW x $195 = $1,989 a month per server, or $249 per GPU per month.
A single-server buyer should expect to pay more per kW than that average. This is base rent, with metered power billed on top.
Step 3: add electricity
10.2 kW running all year is 89,352 kWh. The US EIA's 2025 annual average (Electric Power Monthly Table 5.3, retrieved 21 September 2026) is 13.41 cents per kWh commercial and 8.62 cents industrial.
- Commercial: 89,352 x $0.1341 = $11,982 a year, or $125 per GPU per month.
- Industrial: 89,352 x $0.0862 = $7,702 a year, or $80 per GPU per month.
Small colocation tenants pay at or above the commercial rate, so use $125. Maximum power overstates average draw, so this is an upper bound. It also excludes facility overhead: SemiAnalysis wrote on 4 December 2023 that most colocation facilities run a PUE of about 1.4, so multiply by 1.4 if your contract passes cooling overhead through.
Step 4: total it over the period you will own the hardware
Running cost is $249 + $125 = $374 per GPU per month. Over 36 months that is $13,464. Over 60 months it is $22,440.
| Purchase cost per GPU | Total over 36 months | Total over 60 months |
|---|---|---|
| $31,250 | $44,714 | $53,690 |
| $38,334 | $51,798 | $60,774 |
| $55,077 | $68,541 | $77,517 |
Lenovo's 2026 TCO paper (updated 24 July 2026) amortises over five years with zero recovery value. If you expect to resell, subtract a residual: Silicon Data's listings data (30 January 2026) has three-year-old H100s at 20% to 30% of the historical peak new price.
Step 5: divide by the live hourly price
Break-even hours = total cost of owning / live price per GPU-hour
Take the hourly price from the live table above, not from this text. The result is how many rented GPU-hours the same money buys.
Step 6: compare with the hours you will really use
A month is 730 hours. Your usable hours are months x 730 x utilisation.
| Utilisation | Hours in 36 months | Hours in 60 months |
|---|---|---|
| 90% | 23,652 | 39,420 |
| 50% | 13,140 | 21,900 |
| 20% | 5,256 | 8,760 |
If your break-even hours from step 5 are lower than the hours in your row, buying wins. If they are higher, renting wins. The 90% row has 4.5 times the hours of the 20% row, so a lightly used GPU needs a rental price 4.5 times higher to justify purchase.
What this leaves out
The sum favours buying, because it omits real costs: staff time, financing, and switches and storage if your purchase price was server only. Support is another. ShopBLT listed one-year education-tier DGX H100 support renewals at $10,360.18 and $14,279.71 on 21 September 2026, and commercial tiers cost more. Then failures: during Meta's 54-day Llama 3 405B run on 16,384 H100s there were 419 unexpected interruptions, of which 148 were GPU failures and 72 were HBM3 memory failures (Tom's Hardware, 27 July 2024). Owners carry that risk. Renters hand the node back.
Vendor break-even claims need the same care. Lenovo's paper finds about 5.2 months for an 8x H200 server, but against hyperscaler on-demand list rates at 24x7 use. Against a three-year reservation its own figure is 13.4 months.
Decision rule
Buy H100 systems only if all four of these are true. Your break-even hours from step 5 come in under the 50% row for 36 months, so you have margin if usage slips. You need at least one full 8-GPU node, because that is the unit the hardware is sold in. You already have colocation space that can take 10.2 kW per server, and people to run it. And you have a real source of stock, because the mainstream card listings are discontinued.
Rent in every other case: fewer than eight GPUs, bursty usage or usage below half the month, no three-year commitment, or a likely move to H200 or Blackwell within that time. If you are between the two, price a one-year reserved contract on the cluster page before you buy. It fixes your rate and leaves the resale and failure risk with the provider. To pick hardware before you pick a purchase model, start with the best GPU for LLM work.
Sources
Source pages were retrieved with automated research tools on the dates shown and each figure was traced back to its source before publishing. Rental prices on this page are not typed: they are read live from GPUperhour's own data.
All sources accessed 21 September 2026.
- CDW: PNY NVIDIA H100 80GB PCIe (NVH100TCGPU-KIT), listing retrieved 21 September 2026
- CDW: PNY NVIDIA H100 NVL 94GB (NVH100NVLTCGPU-KIT), listing retrieved 21 September 2026
- CNBC: Nvidia's Blackwell AI chip will cost more than $30,000, CEO says, 19 March 2024
- CNBC: Nvidia's H100 AI chips selling for more than $40,000 on eBay, 14 April 2023
- CNBC: AI GPU depreciation, CoreWeave, Nvidia and Michael Burry, 14 November 2025
- Tom's Hardware: Nvidia Hopper H100 80GB price revealed, 29 April 2022
- Silicon Data: H100 GPU market value trends, 30 January 2026
- Cisco: end-of-sale notice for UCS C885A M8 with NVIDIA HGX H100, 24 April 2025
- Viperatech: upcoming EOL notifications, NVIDIA H100 NVL, L40 and A2, 17 June 2025
- SemiAnalysis: H100 vs GB200 NVL72 training benchmarks, 20 August 2025
- SemiAnalysis: AI server cost analysis, 29 May 2023
- SemiAnalysis: GPU cloud economics explained, 4 December 2023
- SemiAnalysis GPU index, retrieved 21 September 2026
- Newegg: Supermicro 8x H100 80GB HGX system, listing retrieved 21 September 2026
- Newegg: Supermicro SYS-521GE-TNRT, 8x H100 80GB PCIe, listing retrieved 21 September 2026
- ShopBLT: NVIDIA DGX H100 education SKU, listing retrieved 21 September 2026
- ShopBLT: NVIDIA H100 search results, including DGX support renewals, retrieved 21 September 2026
- Broadberry: NVIDIA DGX H100, listing retrieved 21 September 2026
- NVIDIA: DGX H100/H200 user guide, introduction, retrieved 21 September 2026
- Lenovo Press LP2225: on-premise vs cloud generative AI TCO, 2025 edition, 23 May 2025
- Lenovo Press LP2368: on-premise vs cloud generative AI TCO, 2026 edition, updated 24 July 2026
- Tom's Hardware: Nvidia reportedly warns biggest customers of 15 percent price hikes on AI servers, 23 August 2026
- Tom's Hardware: CoreWeave CEO says it has signed an A100 contract running into 2029, 12 August 2026
- Tom's Hardware: faulty H100 GPUs and HBM3 memory caused half of the failures during Llama 3 training, 27 July 2024
- CBRE: North America data center trends H2 2025, 25 February 2026
- US EIA: Electric Power Monthly, Table 5.3, retrieved 21 September 2026