This website uses cookies

Read our Privacy policy and Terms of use for more information.

Last Updated: November 20, 2026

An Nvidia B200 rents for $7.99 to $14.21 per GPU-hour on the official cloud price lists we checked on November 19, 2026. Buying is a different scale of money: Nvidia publishes no list price for the chip, analysts put it at roughly $30,000 to $40,000 per GPU, and a complete 8-GPU HGX or DGX B200 server sells for about $400,000 to $575,000.

Most businesses will never buy a B200 on its own. Nvidia sells it inside 8-GPU servers and 72-GPU racks, and most companies meet it in the cloud. The useful question is what a month of B200 capacity costs, and when owning beats renting.

B200 Price at a Glance

The B200 comes in four price forms: the GPU itself, an 8-GPU server, a 72-GPU GB200 NVL72 rack, and cloud rental. Here is each one in a single view.

How you pay for a B200

Price

Source and date

One B200 GPU (purchase)

About $30,000 to $40,000 (no official list price)

Jensen Huang (then Nvidia CEO) to CNBC, March 2024; Silicon Analysts estimate of ~$40,000 (buy prices last verified June 2026)

8-GPU HGX B200 server

$400,000 to $500,000

Silicon Analysts, based on OEM and reseller listings (last verified June 2026)

DGX B200 (Nvidia's own 8-GPU system)

$515,410

Broadberry listing reported by Wccftech, October 2024

GB200 NVL72 rack (72 GPUs)

About $3 million (2024 estimate)

Morgan Stanley via United Daily News, July 2024; repeated by Barron's

Cloud rental, one GPU

$7.99 to $14.21 per GPU-hour on-demand

RunPod, Nebius, CoreWeave and AWS price pages, November 19, 2026

Market rental index

$6.94 per GPU-hour

Silicon Data B200 index, November 19, 2026

The purchase figures are estimates and listings, not Nvidia prices. Treat them as a budgeting range and get written quotes.

How Much Does It Cost to Rent a B200 per Hour?

Renting a B200 costs $7.99 to $14.21 per GPU-hour on-demand across the four providers below, or about $5,800 to $10,400 per GPU per month if it runs around the clock. The cheapest rates come from specialist GPU clouds, and AWS sits at the top.

Provider and option

Price as listed

Per GPU-hour

One GPU, 730 hours

RunPod Secure Cloud (180 GB)

$7.99/hr per GPU

$7.99

$5,833

Nebius HGX B200, on-demand

$8.50 per GPU-hour

$8.50

$6,205

CoreWeave HGX B200, 8 GPUs (180 GB each)

$68.80/hr per instance

$8.60

$6,278

AWS GB200 UltraServer Capacity Block, 72 GPUs

$761.904/hr per UltraServer

$10.58

$7,725

AWS p6-b200.48xlarge Capacity Block, 8 GPUs

$113.666/hr per instance

$14.21

$10,372

Monthly figures assume 730 hours, a full average month. Tax, storage and data transfer are extra. The AWS rates are for Capacity Blocks, which reserve GPUs for a fixed window and are charged at the price in force when you book. AWS says those prices are next scheduled to update in January 2027.

Cheaper capacity exists if your work can be interrupted. CoreWeave lists a spot price of $34.11 an hour for its 8-GPU B200 instance in North America, about $4.26 per GPU, and Nebius lists preemptible B200 capacity "from $0.99" per GPU-hour, with prices that move with available capacity. Spot capacity can be taken back, so it suits batch jobs and experiments, not customer-facing services.

Google Cloud offers the B200 as its A4 machine type, but its pricing page showed no clear on-demand A4 rate when we checked, and we could not verify Azure or Lambda rates from their own pages, so all three are left out.

B200 rental prices are rising, not falling

A B200 has not gotten cheaper with age. Silicon Data's B200 rental index, a daily blend of neocloud and marketplace rates, read $6.94 per GPU-hour in its dated price table for November 19, 2026, up from $6.14 on October 21. The same index stood at $5.62 on September 1. Data Gravity reported that Ornn's B200 price index rose 79% in three months to $8.01 an hour by September 24, and that Nebius raised its B200 list price from $7.15 to $8.50 on October 1.

That changes the default advice. With older GPUs, waiting usually meant cheaper capacity. With the B200 in late 2026, waiting has meant paying more.

What Does an 8-GPU HGX or DGX B200 Server Cost?

An 8-GPU B200 server costs about $400,000 to $575,000 to buy, depending on the vendor and configuration. Nvidia's DGX B200 and partner HGX B200 systems from companies such as Supermicro both carry eight GPUs with 180 GB of memory each, or 1,440 GB per system according to Nvidia's DGX B200 page.

  • Analyst range: Silicon Analysts puts an 8-GPU HGX B200 server at $400,000 to $500,000 based on OEM and reseller listings.

  • DGX B200: Wccftech reported in October 2024 that server reseller Broadberry listed the DGX B200 at $515,410.43, including Nvidia AI Enterprise software.

  • Current reseller listing: AVADirect listed a Supermicro AS-A126GS-TNBR with eight B200s at $574,348.48 on November 19, 2026, excluding tax and shipping.

What about a single GPU? Tom's Hardware reported in March 2024 that Jensen Huang, Nvidia's CEO at the time, told CNBC Blackwell would sell for $30,000 to $40,000, while noting that Nvidia prefers to sell complete systems rather than standalone chips. Silicon Analysts' 2026 estimate is about $40,000 per GPU, within a $30,000 to $55,000 range, against an estimated manufacturing cost of about $6,750.

Eight GPUs at $40,000 is $320,000, so a large share of a server's price goes to CPUs, memory, networking, storage and the chassis, before power and cooling.

Hands sliding a server into a data center rack

Buying means installing and running the hardware yourself. Photo: Getty Images / Unsplash+

What an 8-GPU B200 node costs to rent instead

Option

Hourly, 8 GPUs

Monthly (730 hrs)

Yearly (8,760 hrs)

RunPod Secure Cloud (8 x $7.99)

$63.92

$46,662

$559,939

Nebius on-demand (8 x $8.50)

$68.00

$49,640

$595,680

CoreWeave HGX B200

$68.80

$50,224

$602,688

AWS p6-b200.48xlarge Capacity Block

$113.67

$82,976

$995,714

One year of a full B200 node on a specialist cloud costs more than buying the server outright. That is why large, steady users look hard at ownership and long contracts.

How Much Does a GB200 NVL72 Rack Cost?

A GB200 NVL72 rack, which links 72 Blackwell GPUs and 36 Grace CPUs into one system, has been estimated at about $3 million, a figure that dates back to 2024. Nvidia has never published a rack price. Tom's Hardware reported in July 2024 that a Morgan Stanley report, cited by United Daily News, projected the NVL72 to start at $3 million and the 36-GPU NVL36 at about $2 million. A 2026 Data Gravity analysis cites Barron's at about $3 million per rack and a Tom's Hardware range of $2.8 million to $3.4 million.

AWS lists its 72-GPU GB200 UltraServer at $761.904 an hour as a Capacity Block, about $556,000 for a 730-hour month. CoreWeave rents GB200 NVL72 capacity in 4-GPU instances at $42.00 an hour, or $10.50 per GPU. At AWS's rate, renting the full rack for about 164 days around the clock costs as much as the 2024 estimate of $3 million, before power, cooling and staff.

Notice the per-GPU math on AWS: a B200 inside the GB200 UltraServer costs $10.58 an hour, less than the $14.21 per GPU on its 8-GPU p6-b200 instance. Our Nvidia AI chips guide explains how the GB200, B200 and other systems fit together.

Should a Business Rent or Buy B200s?

Most businesses should rent. Buying only pays off if you can keep eight GPUs busy nearly every hour for most of a year, and have the power and cooling to host them. Here is how long it takes for rent to add up to Silicon Analysts' $400,000 to $500,000 server range:

Rental option (8 GPUs)

Hourly rate

Hours to reach $400K / $500K

Days at 24/7

AWS p6-b200 Capacity Block

$113.67

3,519 / 4,399

147 / 183

CoreWeave HGX B200

$68.80

5,814 / 7,267

242 / 303

Nebius on-demand

$68.00

5,882 / 7,353

245 / 306

RunPod Secure Cloud

$63.92

6,258 / 7,822

261 / 326

This simple break-even flatters buying. Nvidia lists the DGX B200's maximum power draw at about 14.3 kW, which is roughly 10,400 kWh in a month at full load. Every 10 cents per kWh you pay adds about $1,040 a month, and cooling, data center space, staff and depreciation come on top. Many offices and older server rooms cannot host that kind of load at all, which is part of the wider fight over who pays for AI data center power.

Eight cooling fans inside a server rack

Power and cooling are the costs a simple break-even leaves out. Photo: Getty Images / Unsplash+

A practical rule for a business leader:

  • Rent on demand for pilots, fine-tuning and anything under a few months. Start with one GPU before booking a full node.

  • Reserve or commit when the workload is steady and proven. Nebius says large multi-month reservations cost less than on-demand but publishes no figures, so get quotes from several providers.

  • Buy only with near-constant use, a data center partner that can handle the power, and a plan for when newer chips arrive.

Free download: The AI Stats Sheet 2026. 36 key AI numbers for business on four printable pages, with the source under every figure. Join 2,000+ business leaders who read AI Business Weekly and get it in your inbox right away.

H200 vs B200: Which Should You Pick?

Pick the B200 when you need the most memory and speed per GPU, and the H200 when cost per hour matters more. On the same provider, a B200 rents for about 36% to 107% more than an H200, but it carries 180 GB of memory against 141 GB.

Factor

H200

B200

GPU memory

141 GB HBM3e

180 GB HBM3e (as listed by CoreWeave and RunPod)

Rent, official price lists (Nov 19, 2026)

$3.59 to $10.60 per GPU-hour

$7.99 to $14.21 per GPU-hour

Same-provider premium

Baseline

+36% CoreWeave, +51% RunPod, +57% Nebius, +107% AWS

Market index (Silicon Data, Sept 7, 2026)

$3.29 per GPU-hour

$5.69 per GPU-hour, about 1.7x

Reported purchase price per GPU

About $27,000 (planned China price, Reuters sources, Dec 2025)

About $30,000 to $40,000 (estimates)

Nvidia performance claim

Up to 1.9x H100 on Llama 2 70B inference

DGX B200: 3x training, 15x inference vs DGX H100 (projected)

Best for

Serving mid-size models on a budget, memory-bound work that fits in 141 GB

Large models, heavy training, high-volume inference where speed per GPU cuts total cost

The performance figures are Nvidia's own and compare against the older H100, so treat them as a ceiling. The honest test is cost per job: rent both for a day, run your model, and divide the bill by the work done. A chip that costs 57% more per hour but finishes twice as fast is the cheaper choice.

For provider-by-provider H200 rates, monthly server costs and AWS's 2026 price increases, see our full H200 price guide.

How Much Does the B300 Cost?

The B300, Nvidia's Blackwell Ultra chip, rents for $8.99 to $16.15 per GPU-hour on the official price lists we checked, about 12% to 14% more than a B200 on the same provider. An 8-GPU B300 server costs about $550,000 in the United States, according to sources cited by Reuters in April 2026.

Provider

B200 per GPU-hour

B300 per GPU-hour

B300 premium

RunPod Secure Cloud

$7.99

$8.99 (288 GB)

+13%

Nebius on-demand

$8.50

$9.50

+12%

AWS Capacity Blocks (p6-b200 vs p6-b300)

$14.21

$16.15

+14%

CoreWeave HGX B300 (270 GB)

$8.60

Contact sales (spot $35.84 per 8-GPU instance)

n/a

On the purchase side, Reuters reported that an eight-GPU B300 server was priced at about $550,000 in the US, up from around $500,000 late last year, while restricted B300 servers were selling for about $1 million in China. For the rack version, Data Gravity reported that recent GB300 NVL72 purchase orders came in just under $5.0 million, with a fully deployed rack closer to $5.7 million, excluding the building, core switching and storage.

The B300 premium pays off only if its extra memory lets you run a model on fewer GPUs. Otherwise the B200 is the better-value Blackwell chip.

Why Is the B200 So Expensive?

The B200 is expensive because Nvidia can charge far more than it costs to make, and demand still outruns supply. Silicon Analysts estimates a B200 costs about $6,750 to build and sells for about $40,000, a gross margin near 83%. High-bandwidth memory alone makes up about 48% of that build cost, by the same estimate.

Demand shows up in Nvidia's results. The company reported third-quarter fiscal 2027 revenue of $115.8 billion on November 17, 2026, with $107.6 billion from data centers, up 110% from a year earlier. Our Nvidia statistics page tracks those numbers over time.

The next generation is already shipping. In the same release, Nvidia said Vera Rubin NVL72 racks have shipped and are running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius. New chips usually push older ones down in price, but B200 rents rose through the autumn anyway, so do not budget on a sudden drop.

Sameer's take: Sameer thinks Nvidia's AI chip lead is safe for years, thanks to CUDA and its scale. For the chips trying to challenge that lead, see our guides to Nvidia's competitors and TPU vs GPU.

Frequently Asked Questions

What is the cost of a B200 GPU?
Nvidia does not publish a list price, but estimates put one B200 at about $30,000 to $40,000. Renting one costs $7.99 to $14.21 per GPU-hour on the official cloud price lists we checked on November 19, 2026.

Can you buy a single B200?
Not easily. Nvidia sells the B200 mainly inside 8-GPU HGX and DGX systems and 72-GPU racks, and Tom's Hardware reported that Nvidia prefers selling complete systems over standalone chips. A single B200 is easiest to get by renting one from a cloud such as RunPod or Nebius.

Is B200 faster than H200?
Generally yes. The B200 has 180 GB of memory against the H200's 141 GB, and Nvidia claims the DGX B200 delivers 3 times the training and 15 times the inference performance of the older DGX H100, while it describes the H200's gains over the H100 as coming mostly from memory. Real gains depend on your model, so benchmark before committing.

Is B200 better than H100?
For large AI workloads, yes: it has more than twice the H100's 80 GB of memory and much higher throughput. It also costs more, with Silicon Data's index putting B200 rent at about 2.2 times H100 rent in September 2026, so the H100 can still be the cheaper choice for smaller models.

The Bottom Line

A B200 costs about $8 to $14 an hour to rent, around $47,000 to $83,000 a month for a full 8-GPU node, and roughly $400,000 to $575,000 to buy as a server. Rental prices have been climbing, not falling, through late 2026.

For most companies the right move is to rent one GPU, measure cost per job against an H200, and commit only once the workload is proven. Check each provider's page on the day you sign, because these prices move. If renting GPUs is more than you need, our AI pricing hub covers what AI tools and APIs cost.

  • H200 Price: rental rates, 8-GPU server costs and the H100 vs H200 comparison.

  • Nvidia AI Chips: the full Nvidia data center lineup in one place.

  • Nvidia Competitors: AMD, Google, Amazon and the other companies building alternatives to Nvidia's chips.

  • TPU vs GPU: Google's TPUs and Nvidia's GPUs compared for business use.

  • AI Stocks: the public companies behind the AI chip boom. Not investment advice.

⭐ Get more AI Business Weekly in Google: add us as a preferred source on Google.

Written by Sameer Khan, with support from the AI Business Weekly Editorial Team. AI-assisted and fact-checked against the sources linked above.