pricinghardware

H200, B200, A100 Prices: You Buy a Server, Not a GPU

Verified purchase listings for NVIDIA H200, B200, B300, GB200 NVL72 and A100 with seller and date, the misquoted figures corrected, and live rental prices.

By Faiz Ahmed•
•15 min read

The only GPU on this page that you can order as a single card from a mainstream US reseller is the H200 NVL, which CDW listed at $33,731.99 on 21 September 2026 with a lead time of "4-6+ Weeks". B200 and B300 are sold only inside 8-GPU systems that list from $384,228 to $785,606.50, GB200 and GB300 are sold as racks that analysts put at about $3 million and up, and the A100 is discontinued. NVIDIA publishes no list price for any of them except the DGX A100 from 2020, so every figure below is a reseller listing or an analyst estimate, and is labelled as one.

If what you need is compute and not hardware, here is what the same GPUs rent for right now:

GPUCheapest $/GPU-hrProviderProviders in stock
H200$3.43QuantaCloud5
B200$3.75Packet.ai3
B300$7.89RunPod1
A100$0.68LeaderGPU7
Cheapest in-stock on-demand price per GPU-hour, from providers with live stock tracking. Latest stock observation: .

Verified listings at a glance

Each row is a listing we could read directly, with the seller and the date. "Last listed" means the item was out of stock or discontinued, so the price is a last-seen price and not an offer.

ProductWhat it isPriceSellerDate
PNY H200 NVL 141GBSingle PCIe card$33,731.99CDW, "4-6+ Weeks"21 September 2026
Supermicro SYS-821GE-TNHR8x H200 HGX server, starting price$319,421.64Supermicro eStore21 September 2026
Lenovo ThinkSystem, 8x H200Server, usual customer sale price$397,801.60Lenovo Press LP2368as of 15 June 2026
DGX H200NVIDIA system, 8x H200$467,311.30Broadberry US, default configuration21 September 2026
Supermicro SYS-A22GA-NBRT-G18x B200 HGX server with 400G NICs$384,228Newegg, last listed21 September 2026
Supermicro SYS-A21GE-NBRT-G18x B200 HGX server with 400G NICs$394,406Newegg, last listed21 September 2026
Lenovo ThinkSystem SR680a V3, 8x B200Server, usual customer sale price$550,475.10Lenovo Press LP2368as of 15 June 2026
DGX B200NVIDIA system, 8x B200$562,154.09Broadberry US, default configuration21 September 2026
Supermicro SYS-822GS-NB3RT8x B300 HGX server, starting price$536,771.05Supermicro eStore21 September 2026
Supermicro AS-8126GS-NB3RT-01-G28x B300 HGX server, fixed configuration$662,845.95Supermicro eStore, shown in stock21 September 2026
DGX B300NVIDIA system, 8x B300$646,878.46Broadberry US, default configuration21 September 2026
Lenovo ThinkSystem SR680a V4, 8x B300Server, usual customer sale price$785,606.50Lenovo Press LP2368as of 15 June 2026
Exxact DGX Station, one GB300 superchipDesktop workstationfrom $94,930Exxact, reported by Tom's Hardware23 August 2026
PNY A100 80GB PCIeSingle PCIe card$13,223.99CDW, "Discontinued", last listed21 September 2026
DGX A100NVIDIA system, 8x A100, launch pricefrom $199,000NVIDIA press release14 May 2020

There is no GB200 NVL72 row because no public listing is usable as a price. That section explains why.

Specifications for the five families, from our spec table:

SpecA100H100H200B200B300
VRAM40 to 80 GB80 to 94 GB141 GB180 to 192 GB262 to 288 GB
Memory bandwidth2,039 GB/s3,350 GB/s4,800 GB/s8,000 GB/s8,000 GB/s
FP16 (dense)312 TFLOPS989 TFLOPS989 TFLOPS2,250 TFLOPS2,250 TFLOPS
FP16 (with sparsity)624 TFLOPS1,979 TFLOPS1,979 TFLOPS4,500 TFLOPS4,500 TFLOPS
FP8 (dense)Not published1,979 TFLOPS1,979 TFLOPS4,500 TFLOPS4,500 TFLOPS
FP8 (with sparsity)Not published3,958 TFLOPS3,958 TFLOPS9,000 TFLOPS9,000 TFLOPS
TDP400 W700 W700 W1,000 W1,400 W
Launch year20202022202420242025
Figures from the vendor datasheets: A100, H100, H200, B200, B300, checked 13 Sep 2026. With-sparsity figures assume 2:4 structured sparsity and are twice the dense figure, so compare dense with dense. "Not published" means the vendor gives no figure.

H200

What is sold: a PCIe card (H200 NVL), HGX 8-GPU servers from Supermicro, Dell, Lenovo and others, and NVIDIA's own DGX H200.

The CDW card listing is $1,349 above the last price CDW showed for the discontinued H100 NVL ($32,382.99), which is about 4% more for a card with 141 GB of memory and 1.4 times the memory bandwidth of the H100 in our spec table. Compute figures are the same for both. H100 vs H200 has the detail.

For the SXM version there is no reputable US per-unit price, because SXM modules ship only on HGX baseboards inside systems. The one reported per-chip figure is for China: Reuters, citing unnamed sources, reported on 1 January 2026 that NVIDIA priced the H200 for Chinese customers at around $27,000 per chip. That is volume pricing subject to a 25% US fee, not a US price.

Per-GPU arithmetic on the three system listings:

  • Supermicro starting price: $319,421.64 / 8 = $39,928. The configuration behind the starting price is not itemised on the page, so read it with care.
  • Lenovo usual sale price: $397,801.60 / 8 = $49,725.
  • Broadberry DGX H200: $467,311.30 / 8 = $58,414.

A correction: CDW also shows HPE's HGX H200 8-GPU accelerator option at $1,038,021.99, marked "Discontinued" (21 September 2026). That is an OEM option list price. Do not read it as a market price.

To rent instead, the cheapest H200 is $3.43 per GPU-hour on QuantaCloud (checked ).

B200 and DGX B200

What is sold: HGX B200 8-GPU servers and the DGX B200. Not cards, and not loose modules.

What Jensen Huang actually said

On 19 March 2024, Jensen Huang told CNBC's Jim Cramer that Blackwell "will cost between $30,000 and $40,000 per unit", and estimated about $10 billion in R&D. Later the same day he told CNBC's Kristina Partsinevelos that the cost "isn't just about the chip but also about designing data centers and integrating into other company's data centers."

He said "Blackwell". He did not say "B200" and named no SKU. Tom's Hardware noted the same day that the figure is approximate because NVIDIA prefers to sell whole systems, and lists only DGX B200 systems and SuperPODs, not B200 modules. Articles that quote his range as "the B200 price" misreport it.

A related misquote: HSBC's estimate of $60,000 to $70,000, reported by Tom's Hardware on 14 May 2024, is for a GB200 superchip, which is one Grace CPU plus two B200 GPUs. It is not a per-GPU price.

HGX B200 servers

Newegg showed two Supermicro Gold Series 10U servers on 21 September 2026, both out of stock. Both include the HGX B200 baseboard, eight 400G network cards and a BlueField-3 DPU.

  • SYS-A22GA-NBRT-G1, dual Xeon 6960P, 2.3TB DDR5: $384,228 / 8 = $48,029 per GPU.
  • SYS-A21GE-NBRT-G1, dual Xeon Platinum 8570, 3TB DDR5: $394,406 / 8 = $49,301 per GPU.

Lenovo's usual customer sale price for an 8x B200 ThinkSystem SR680a V3, as of 15 June 2026, was $550,475.10, or $68,809 per GPU.

DGX B200 price

"NVIDIA's official DGX B200 list price is $515,000" is wrong on two counts. The figure comes from a Broadberry reseller listing at $515,410.43, which Wccftech reported on 12 October 2024. NVIDIA published nothing. And the same Broadberry listing read $562,154.09 on 21 September 2026, about 9% higher, which is $70,269 per GPU. CDW lists the DGX B200 with three years of support, no public price and a status of "4-6+ Weeks" (21 September 2026).

No reputable source quotes a US price for a single B200 or B300 module. The ranges repeated elsewhere carry no citations.

To rent instead, the cheapest B200 is $3.75 per GPU-hour on Packet.ai (checked ). Compare B200 vs H200.

B300

What is sold: HGX B300 8-GPU servers and the DGX B300. NVIDIA's DGX B300 user guide describes eight B300 GPUs at 288GB each and a power consumption of 14.5 kW.

The B300 has the firmest price on this page. Supermicro's own eStore showed the Gold Series AS-8126GS-NB3RT-01-G2, a fixed configuration with dual EPYC 9575F, in stock at $662,845.95 on 21 September 2026. That is $82,856 per GPU from the manufacturer.

The other listings bracket it. Supermicro's configurable SYS-822GS-NB3RT starts at $536,771.05 ($67,096 per GPU). Broadberry's DGX B300 default configuration is $646,878.46 ($80,860 per GPU). Lenovo's usual sale price for an 8x B300 ThinkSystem SR680a V4 was $785,606.50 as of 15 June 2026 ($98,201 per GPU). CDW lists the DGX B300 without a public price at "4-6+ Weeks".

A correction: Tom's Hardware, summarising the Financial Times on 24 June 2026, said the DGX B300 "retails in the U.S. for nearly $400,000". That matches no US listing we found, which run from about $537,000 to $786,000.

To rent instead, the cheapest B300 is $7.89 per GPU-hour on RunPod (checked ).

GB200 and GB300 NVL72

What is sold: a liquid-cooled rack. An NVL72 rack holds 36 GB200 superchips, which means 72 GPUs and 36 Grace CPUs, as Supermicro's specification confirms. Articles that say "72 GB200 superchips" have it wrong.

There is no verified listing, only estimates:

  • SemiAnalysis, 20 August 2025: the GB200 NVL72 rack "alone costs $3.1M for a typical hyperscaler", and about $3.9 million with networking and storage. Divided by 72, that is $43,056 per GPU for the rack and $54,167 all-in. SemiAnalysis puts the all-in capital cost per GPU at about 1.6 to 1.7 times that of an H100.
  • Wolfe Research, 30 January 2026: about $3 million for a GB200 NVL72 and $4.3 million for a GB300 NVL72.
  • For GB300 the estimates disagree. A single unnamed source quoted by Tom's Hardware on 24 March 2026 put it at $6 million to $6.5 million, and the article adds that "Nvidia has never confirmed the list prices." We do not know which is closer.

A correction: Broadberry lists an "NVIDIA DGX GB200" with 72 GPUs at $9,833,605.20 (21 September 2026). That is roughly three times every analyst estimate for a GB200 NVL72. It is an outlier.

Prices may rise further. Tom's Hardware reported on 23 August 2026, citing Bloomberg, a rise of more than 15% on Grace Blackwell and Vera Rubin servers shipping from early 2027, which NVIDIA has not confirmed. The H100 price article covers the 2026 price turn in full.

Price is not the main barrier. Supermicro specifies its GB200 NVL72 at 132 kW total power, liquid cooled. Tom's Hardware (12 August 2026) sets that against roughly 20 kW per rack in legacy halls.

On the rental side we track the GB300: no in-stock on-demand offer right now. For rack-scale capacity, use cluster and reserved pricing.

A100

What is sold: nothing new. The A100 is end-of-life, and what remains is last-listed reseller pages and a used market.

The A100 is the one case with an official price. NVIDIA's press release of 14 May 2020 said "NVIDIA DGX A100 systems start at $199,000", which is $24,875 per GPU for a full 8-GPU system.

For the card, Tom's Hardware recorded the A100 80GB at $13,224 at CDW in April 2022. CDW still showed the PNY A100 80GB PCIe at $13,223.99 on 21 September 2026, marked "Discontinued", with a listing age of 1,851 days. It is a last-listed price only.

On end of life: integrator One Stop Systems issued a notice on 24 January 2024 for the HGX A100 4-GPU 80GB and the PCIe A100 80GB, with a final order date of 26 April 2024 and a final factory ship date of 31 December 2024. We did not find NVIDIA's own notice, only integrator and reseller relays. On software support, NVIDIA staff wrote on the developer forum on 21 February 2023 that "we support typically for at least 3 years after last time buy ... most of the time MUCH longer." That is a forum post, not a policy.

On the used market, the data is thin. Hashrate Index's overview of 8 April 2026, a secondary summary that shows no transaction data, describes visible commercial listings from roughly $7,800 for a refurbished A100 40GB to about $18,900 for an A100 80GB PCIe. Those are asking prices, not transactions. The top of that range is above the 2022 new price.

Operators say the same. CNBC reported on 14 November 2025 that CoreWeave's A100s are "all fully booked", and that NVIDIA CFO Colette Kress said A100s sold six years ago still run at full utilisation. Tom's Hardware reported on 12 August 2026 that CoreWeave signed an A100 contract running into 2029.

To rent instead, the cheapest A100 is $0.68 per GPU-hour on LeaderGPU (checked ). Against newer parts, see A100 vs H100 and A100 vs H200.

Renting instead

Renting skips the reseller lead times and the power build-out. The table at the top of the page shows the cheapest live price for each GPU and whether providers have it in stock today. This one turns the hourly price into a 730-hour monthly figure, the number to hold against a purchase quote:

GPU$/GPU-hrPer day (×24)Per month (×730)Provider
H200$3.43$82.32$2,504QuantaCloud
B200$3.75$90.00$2,738Packet.ai
B300$7.89$189.36$5,760RunPod
A100$0.68$16.32$496LeaderGPU
Cheapest in-stock on-demand price per GPU-hour, for one GPU running the whole time. Per day = hourly price × 24 hours. Per month = hourly price × 730 hours (365 days × 24 ÷ 12), rounded to the nearest dollar. Storage, data transfer and tax are not included. Latest stock observation: .

A purchase price is not the whole cost of owning. NVIDIA's user guides give maximum system power of 10.2 kW for a DGX H200 and 14.3 kW for a DGX B200, and colocation is contracted on that figure. The H100 price article linked above walks through the break-even arithmetic with sourced colocation and electricity inputs, and the rent vs buy GPU calculator runs it for any GPU on this page. Day-to-day movement is on the GPU price index.

Decision rule

A100: rent. You cannot buy it new, the used market shows only asking prices, and the hardware has no FP8 support in our spec table.

H200: the only sensible small purchase here. Buy H200 NVL cards if you already own PCIe servers, need a fixed amount of on-premise inference capacity, and will keep the cards busy for years. Otherwise rent.

B200 and B300: buy only as complete 8-GPU systems, which means $384,228 to $785,606.50 per node on the listings above, plus a site that can supply more than 14 kW per server. If you cannot tick both boxes and show high utilisation in the calculator, rent.

GB200 and GB300 NVL72: not a purchase for anyone reading a price guide. No listing is reliable, the racks need liquid cooling and well over 100 kW, and a price rise is reported for 2027. Rent rack-scale capacity on a contract.

Whatever you choose, quote a price only with its seller and date. Apart from the 2020 DGX A100 launch price, none of these numbers is an NVIDIA list price, and most will have moved by the time you read this.

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.

Frequently asked questions

What is the NVIDIA H200 price?▾

The H200 NVL PCIe card is the only one of these GPUs with a single-card listing at a mainstream US reseller, and the article gives that CDW listing with its date. The SXM version of the H200 has no reputable per-unit price because it ships only on 8-GPU baseboards inside servers.

Did Jensen Huang give a price for the B200?▾

No. On 19 March 2024 he gave a per-unit range for Blackwell without naming a SKU, and clarified the same day that the cost is not just about the chip. Quoting his range as the B200 price is a misreport.

Is the widely quoted DGX B200 price an official NVIDIA list price?▾

No. The figure traces to a Broadberry reseller listing from October 2024. NVIDIA published no list price, and the same listing is higher today.

What does a GB200 NVL72 cost?▾

There is no usable public listing, only analyst estimates in the low millions of dollars per rack. NVIDIA has never confirmed a list price, and an NVL72 rack holds 72 GPUs on 36 GB200 superchips.

Can you still buy an NVIDIA A100?▾

Not new from mainstream channels. An integrator's end-of-life notice set the final order date in April 2024, CDW marks the card Discontinued, and what remains is a used market where only asking prices are visible.

Should I buy or rent these GPUs?▾

Most readers should rent, because B200 and B300 are sold only as complete systems and GB200 only as liquid-cooled racks. The H100 price article and the rent vs buy calculator show how to check the break-even for your own utilisation.

TAGGED WITH

#H200#B200#A100#H100#cost-comparison#cloud-gpu

Related Posts

How to Read GPU Specs for AI: 3 Numbers That Decide

VRAM, memory bandwidth and dense throughput decide an AI workload. How to find them on a datasheet, and why the headline TFLOPS figure is usually doubled.

Read more

NVLink vs PCIe vs SXM: When the Interconnect Matters

NVLink, PCIe and SXM explained for people renting multi-GPU servers: bandwidth per generation, what the H100 NVL is, and when the SXM premium pays off.

Read more

What MIG Is, and When a Fractional GPU Is Enough

MIG splits one NVIDIA data centre GPU into isolated instances with their own memory. See which GPUs support it and when a slice beats a whole cheap GPU.

Read more