The honest answer is about $14,000 to $16,500 for a new A100 80GB PCIe card, and new cards are hard to find. Both distributor listings cited below showed the card out of stock on 2026-09-28, and one server maker's end-of-life notice set its last A100 order date for April 2024. Renting one on QuantaCloud costs from $1.48/GPU-hr. At the lowest A100 price on 2026-09-27, $1.475 per GPU-hour, a card at $13,999.95 to $16,535 equals 9,491 to 11,210 hours of rental, or 13 to 15 months of round-the-clock use (our calculation, below). For a whole system, NVIDIA's eight-GPU DGX A100 started at $199,000 when it launched in May 2020.
How much an A100 costs to buy#
The price depends on which A100 you mean and when you asked. NVIDIA published a system price for this generation, and the rest comes from news reports and distributor listings.
| Date | Source | What the price covers | Price |
|---|---|---|---|
| May 2020 | NVIDIA | DGX A100 system with eight A100 40GB GPUs, 320 GB in total | Starts at $199,000 |
| April 2022 | Tom's Hardware | A100 80GB card, what it could be bought for that day | $13,224 |
| February 2023 | CNBC | One A100, memory size not stated | Roughly $10,000 |
| February 2023 | CNBC | DGX A100 with eight A100 GPUs | Suggested price of nearly $200,000 |
| 2026-09-28 | Central Computer listing, PNY kit NVA100TCGPU80-KIT | New A100 80GB PCIe card, out of stock | $13,999.95 |
| 2026-09-28 | SabrePC listing, same kit | New A100 80GB PCIe card, out of stock | $16,535.00 |
The sources agree on the order of magnitude and disagree on the details. CNBC's roughly $10,000 does not name the memory size, while all three dated prices for the 80 GB card are above $13,000. The 80 GB card's listed price has not fallen with age either: $13,224 in April 2022, then $13,999.95 and $16,535.00 at two distributors in September 2026. CDW's own listing for the PNY kit shows it discontinued on June 10, 2022. I would budget from the 2026 listings and expect to wait for stock.
What an older generation means for price and availability#
The core problem with buying an A100 in 2026 is supply, not price. NVIDIA put the A100 into full production in May 2020 and announced the 80 GB version that November. In January 2024, One Stop Systems, a server maker, published an end-of-life notice for the A100 80GB PCIe card and the HGX A100 4-GPU board: last orders on April 26, 2024, last factory shipments on December 31, 2024, and the H100, L40S and L40 named as replacements. Communications Today, a trade publication, reported in February 2025 that NVIDIA had announced end of life for the PCIe and SXM versions in January 2024. New cards now come from whatever stock is left, and if a card you own fails, its replacement comes from the same stock.
The A100 is also a generation behind on number formats. FP8 Tensor Core math needs compute capability 8.9 or newer, and the A100 is 8.0, so vLLM runs FP8 checkpoints on it weight-only (W8A16) rather than natively. For BF16 training and inference that does not matter, and 80 GB is still 32 GB more than the L40S or RTX A6000 carries. The L40S vs A100 and RTX A6000 vs A100 comparisons cover that trade-off.
The upside of an older generation is the hourly price. These live prices put both A100 versions next to the H100 PCIe:
| GPU | Memory | From | Available now |
|---|---|---|---|
| A100 PCIe 80GB | 80 GB | $1.48/GPU-hr | Yes |
| A100 SXM4 80GB | 80 GB | $1.49/GPU-hr | Yes |
| H100 PCIe | 80 GB | $2.59/GPU-hr | Yes |
A100 80GB PCIe vs SXM4#
For a single GPU, the two A100 80GB versions are close. The SXM4 module has about 5% more memory bandwidth (our calculation: 2,039 / 1,935 = 1.054) and a 400 W power limit against 300 W.
| A100 80GB PCIe | A100 80GB SXM4 | |
|---|---|---|
| Memory | 80 GB HBM2e | 80 GB HBM2e |
| Memory bandwidth | 1,935 GB/s | 2,039 GB/s |
| Maximum power | 300 W | 400 W |
| GPU-to-GPU link | NVLink bridge, 2 cards, 600 GB/s | NVLink, 12 links, 600 GB/s |
| FP8 Tensor Core | No | No |
| Sold as | Dual-slot PCIe card | Module on 4- or 8-GPU HGX A100 boards |
The SXM4 part has no card price I can quote, for the same reason as the H100 SXM: it ships on HGX boards inside servers such as the DGX A100. The difference that matters at eight GPUs is the NVLink fabric on the HGX board, which the PCIe card only gets in bridged pairs.
DGX A100 and HGX A100 prices#
The one solid system price is NVIDIA's own: DGX A100 systems started at $199,000 in May 2020, with eight A100 40GB GPUs, and CNBC still quoted a suggested price of nearly $200,000 in February 2023. The DGX A100 640GB, with eight 80 GB GPUs, followed in November 2020, and NVIDIA's announcement gave no price. I could not find a reliable, dated price for an HGX A100 server today. With the A100 past end of life, I would treat any A100 server offered now as old stock or used, and ask the seller which.
What renting an A100 costs on QuantaCloud#
An A100 80GB rents from $1.48/GPU-hr on QuantaCloud, in both the PCIe and SXM4 versions. These are the live configurations:
| GPUs | vCPU | RAM | Disk | Region | Per hour | Per GPU-hour | Launch |
|---|---|---|---|---|---|---|---|
| 1x A100 SXM4 80GB | 16 | 120 GB | 1,000 GB | us-midwest-2 | $1.50 | $1.50 | Launch |
| 1x A100 SXM4 80GB | 16 | 120 GB | 1,000 GB | us-midwest-1 | $1.50 | $1.50 | Launch |
| 1x A100 SXM4 80GB | 16 | 120 GB | 1,000 GB | us-east-1 | $1.50 | $1.50 | Launch |
| 1x A100 SXM4 80GB | 14 | 100 GB | 625 GB | us-midwest-2 | $1.50 | $1.50 | Launch |
| 2x A100 PCIe 80GB | 30 | 150 GB | 1,536 GB | us-midwest-2 | $2.95 | $1.48 | Launch |
| 4x A100 PCIe 80GB | 60 | 300 GB | 3,072 GB | us-midwest-2 | $5.91 | $1.48 | Launch |
| 8x A100 SXM4 80GB | 126 | 800 GB | 5,000 GB | us-midwest-2 | $11.92 | $1.49 | Launch |
Prices checked 5 Oct 2026, 22:36 UTC
On 2026-09-27 the offers ran from a single A100 SXM4 VM to an 8-GPU A100 SXM4 VM with 126 vCPU, 800 GB of RAM and 5,000 GB of disk, across Virginia and the Midwest. Every A100 instance is a VM with Ubuntu 22.04, the NVIDIA driver and Docker, and you connect over SSH as ubuntu. You pay from a prepaid balance with a $5 minimum deposit: the first hour at launch, each further hour as the previous one runs out, and a refund of the unused seconds when you stop. Data transfer in and out is not charged. A stopped instance is terminated and its disk deleted, so move your results off before you stop it. The A100 page covers specs and sizing, and the pricing page and billing docs cover the rules.
Break-even: how many rented hours one A100 buys#
The rule I follow is the same for every GPU: purchase price divided by hourly price, then divided by the hours a year the GPU will really work. At $1.475 per GPU-hour, the lowest A100 price on 2026-09-27 (an A100 PCIe 80GB in a 2-GPU VM at $2.95 an hour, live now from $1.48/GPU-hr), our calculation gives:
| Purchase price used | Rental hours it equals | Years at 100% use | Years at 50% use | Years at 25% use |
|---|---|---|---|---|
| $13,999.95 (A100 80GB PCIe, Central Computer, 2026-09-28) | 9,491 | 1.1 | 2.2 | 4.3 |
| $16,535.00 (A100 80GB PCIe, SabrePC, 2026-09-28) | 11,210 | 1.3 | 2.6 | 5.1 |
| $199,000 (DGX A100 at launch) against an 8x A100 SXM4 VM at $11.92 an hour on 2026-09-27 | 16,695 | 1.9 | 3.8 | 7.6 |
The arithmetic for the first row: $13,999.95 / $1.475 = 9,491 hours, and 9,491 / 8,760 hours in a year = 1.1 years at 100% use, 2.2 years at 50% and 4.3 years at 25%. For the system row, $199,000 / $11.92 = 16,695 hours. That row is not like for like: the DGX A100 at that price had eight 40 GB GPUs against the VM's eight 80 GB GPUs, and $199,000 was a 2020 launch price. If the live prices differ from the ones on 2026-09-27, divide by the live numbers.
The time column is the one to watch. At 25% use a card needs 4.3 to 5.1 years to match the rent, and the A100 itself went from full production in May 2020 to One Stop Systems' last order date in April 2024, just under four years.
What the purchase price leaves out#
The purchase price covers the GPU and nothing it runs on.
| Cost | Buying an A100 | Renting on QuantaCloud |
|---|---|---|
| Host server for PCIe cards, or a whole HGX server for SXM4 | A separate purchase | In the hourly price |
| Power per GPU | 300 W PCIe or 400 W SXM4: 2,628 or 3,504 kWh a year at full load | In the hourly price |
| Cooling and rack space | Yours: the PCIe card is passive and needs server airflow | In the hourly price |
| Spares for a discontinued card | Remaining stock or the used market | Launch another VM |
| Network, setup and staff time | Yours | In the hourly price, no egress charges |
| Depreciation and resale | Yours | None |
The energy figures are our calculation: 300 W x 8,760 hours = 2,628 kWh, and 400 W x 8,760 hours = 3,504 kWh, for the GPU alone at full load. Multiply by your electricity rate, then add the host server. I have not found a reputable, dated source for used A100 prices, so I leave resale value out rather than guess.
Buy, rent or reserve#
My rule for the A100 is to rent it unless you are extending a fleet you already run. Its best feature in 2026 is the hourly price, and owning one adds the risks of a discontinued product to the usual costs of owning hardware. Renting also lets you move up a generation the day the A100 stops fitting: the H100 price guide and H200 price guide run the same arithmetic for newer cards, and the A100 vs RTX 5090 comparison covers the consumer card people weigh against it.
Buying makes sense when your servers already take PCIe cards, you have the power and cooling, and matching the hardware you run matters more than FP8 or other newer features.
Reserving fits sustained work on dedicated hardware you do not want to own. We order and build the hardware to your spec, and the configuration, lead time and terms come back in writing. For a GPU past end of life, I would describe the workload in the brief rather than insist on the A100, so the written quote can compare it with the cards that replaced it. Send a capacity brief, or read how reserved capacity works first.
The decision rule is short: rent the A100 by the hour, measure how busy it really is, and let that number decide, not the card price. Start from the A100 page, run a week of real work, and put the busy hours into the table above.