---
title: Where to Rent a B200 SXM at the Cheapest Price (October 2026)
url: https://gpubank.io/rent-b200
description: B200 SXM rental prices checked on 16 October 2026: from $0.99 spot and $3.69 on demand to $14.24 on AWS, with node, reserved and cluster offers compared.
generated: 2026-11-02T00:20Z
publisher: GPUBank (GPU Compute OÜ, Tallinn, Estonia)
summary_for_ai: https://gpubank.io/llms.txt
full_text_for_ai: https://gpubank.io/llms-full.txt
---

> Markdown version of https://gpubank.io/rent-b200, generated from the live page. GPUBank is a GPU cloud billed by the second at one all-in price per GPU (vCPU, RAM, NVMe, a public IPv4 and traffic included), in Paris, Frankfurt, Ashburn and Singapore. Prices are in US dollars, excluding VAT.

Updated 16 Oct 2026 · 35 offers checked

# Where to rent a B200 SXM at the cheapest price

We read every B200 offer we could find on the provider's own page, separated single GPUs from whole nodes, spot from reserved and shared from dedicated, then priced the monthly bill, the cost per job against the H200 and H100, and the point where buying an HGX server pays.

**GPUBank Research** 20 min read · Prices read on 16 Oct 2026

**180 GB**HBM3e per GPU

**8 TB/s**memory bandwidth

**1.8 TB/s**NVLink 5

**1,000 W**per GPU

The short answer, 16 Oct 2026

1. $0.99 /hr **Lowest price listed.** Nebius lists its preemptible B200 from $0.99 per GPU-hour, but that is the floor of a dynamic spot price that can change every 15 minutes, rise to one cent under the $8.50 on-demand rate and stop the VM with 60 seconds of notice, and its default B200 quota is zero, so most accounts cannot book it today.
2. $5.94 /hr **Cheapest marketplace listing.** Lium's marketplace listed B200 pods from $5.94 per GPU-hour, the cheapest an 8-GPU pod in Helsinki, with the host's CPU, RAM and disk in the price and billing by the second; no marketplace undercut the cheapest managed single GPU.
3. $4.15 /hr **Whole machine, traffic included.** GPUBank's $4.15 covers 28 vCPU, 283 GB of RAM, 2 TB of NVMe, an IPv4 address and all traffic, billed per second, and falls to about $2.59 with a $1,000 top-up; on list price Deep Infra's $3.69 single-GPU container with 16 vCPU and 250 GB of RAM is lower.

[B200 SXM $4.15 /hr all-in Launch a B200 SXM Spot $2.49 · reserved $3.32
Billed per second](https://console.gpubank.io/launch?gpu=b200-sxm)

The same B200 silicon is listed this month at anything from 99 cents to more than $14 per GPU-hour, and almost none of that spread comes from the GPU. It comes from what is sold around it: one card or a mandatory eight, a machine with CPU, memory and disk or a bare GPU line, an hour that can be taken back with a minute's notice, a meter that runs per second or per started hour, a contract of a week or of four years, and stock that may or may not exist on the day you want to start. Two offers with the same headline number can be different products, and the cheapest line on an aggregator is often one you cannot book.

We read the prices on 16 October 2026 on each provider's own pricing page, documentation or public price API wherever we could, and we name the aggregator and its date where we could not. Several figures had moved since our previous survey on 29 September: Nebius raised its on-demand B200 from $7.15 to $8.50 on 1 October and turned its preemptible tier into a dynamic spot price on 8 October, Verda's single GPU went from $6.85 to $7.87, and fal's list price rose to $7.99. Read every table on this page as a dated snapshot rather than a standing quote.

This page is published by GPUBank, a GPU cloud run by GPU Compute OÜ in Tallinn that rents the B200 SXM itself, so read our own entry with that in mind. It is measured on the same terms as every other offer, and the text says where a spot tier, a marketplace, a long contract or a cheaper single-GPU container beats it, and what the buyer gives up or gains in each case.

01 / One hour, six products

## Why a B200 hour is not one product

Price per GPU-hour by buying unit, from one card to a rack

One GPU in the unit GPUBank

011 GPU on demand

DDeep Infra

**×1**

$3.69 /GPU-hr

Commitment **None**

16 vCPU, 250 GB RAM, billed per minute

021 GPU reserved

GPUBank

**×1**

$3.32 /GPU-hr

Commitment **1 week**

Prepaid, GPUs held for you, $4.15 on demand

038-GPU node on demand

CoreWeave

**×8**

$8.60 /GPU-hr

Commitment **None**

$68.80 per node-hour, 2,048 GB RAM

048-GPU node reserved

Civo

**×8**

$3.79 /GPU-hr

Commitment **3 years**

$22,133.60 a month; $4.49 on 1 year

0516-GPU cluster reserved

Lambda

**×16**

$9.86 /GPU-hr

Commitment **2 weeks**

InfiniBand; $8.87 from 256 GPUs

06GB200 NVL72 rack

AWS

**×72**

$10.58 /GPU-hr

Commitment **Capacity Block**

$761.90 per rack-hour, Dallas Local Zone

One offer for each buying unit, per GPU-hour, read on 16 Oct 2026. The squares show the smallest number of GPUs the unit can be rented with; the rack is sold whole, 72 GPUs at a time.

With most cards the unit of purchase is one GPU and the only variable is the rate. With the B200 the unit itself changes the price, and the table above shows the main ones side by side. A single GPU on demand is the newest and most flexible unit, sold by a handful of clouds from $3.69 with no commitment. An 8-GPU node on demand, the classic neocloud unit, costs $6.69 to $8.60 per GPU and makes you pay for eight cards even when your job needs one. An 8-GPU node on a contract of one to three years falls to $3.79 to $4.49 per GPU at Civo, and Vultr quoted $2.99 on 48 months. Clusters with InfiniBand start at 16 GPUs and two weeks on Lambda at $9.86, falling to $8.87 at 256 GPUs, and a full GB200 NVL72 rack on an AWS Capacity Block costs $761.90 an hour for 72 GPUs, or $10.58 per GPU.

The prices run backwards from intuition: the long contract is the cheapest hour and the short cluster the dearest, because the cluster price pays for the fabric, the guaranteed placement and the provider's risk of idle racks, while the multi-year contract transfers that risk to you. The right unit follows from the job. A model that fits in 180 GB and a workload that comes and goes point to a single GPU billed by the second; tensor-parallel serving or a fine-tune of a 70B model points to one node; pre-training or mixture-of-experts training at scale points to a cluster or an NVL72 rack, where NVLink across 72 GPUs does what InfiniBand between nodes cannot. Buying the bigger unit for a small job is the most common way to overpay for a B200.

### Three generations on the same benchmark rules

The chart below uses one MLPerf round, Inference v5.0, so that the B200, H200 and H100 rows were measured on the same benchmark rules and the same TensorRT release. The language-model rows compare each card as you would deploy it, FP4 on the B200 against FP8 on Hopper, so they show the platform's gap rather than the silicon's alone, while the Stable Diffusion XL row, at FP8 on all three, is the like-for-like test and its 1.6 times the realistic floor of what a B200 hour buys over an H100 hour.

MLPerf Inference v5.0, 8-GPU systems, each row scaled to its own best result

B200 SXM H200 SXM H100 SXM

**Llama 2 70B, server**

tokens/s, 8 GPUs

*B200 FP4, Hopper FP8*

B200

98,443 3.16× H100

H200

33,072

H100

31,106

**Llama 2 70B Interactive, server**

tokens/s, 8 GPUs

*B200 FP4, Hopper FP8*

B200

59,409 4.24× H100

H200

19,424

H100

14,009

**Mixtral 8x7B, server**

tokens/s, 8 GPUs

*B200 FP4, Hopper FP8*

B200

126,845 2.38× H100

H200

61,802

H100

53,299

**Llama 3.1 405B, offline**

tokens/s, 8 GPUs

*B200 FP4, Hopper FP8*

B200

1,526 3.50× H100

H200

561

H100

436

**Stable Diffusion XL, offline**

images/s, 8 GPUs

*FP8 on all three*

B200

30.4 1.65× H100

H200

19.0

H100

18.4

MLPerf Inference v5.0, closed division, TensorRT 10.8 and CUDA 12.8 on every row. B200 and H200 rows are NVIDIA's own systems; the H100 row is NVIDIA's DGX H100 on Llama 2 70B and the best 8x H100 SXM entry elsewhere (HPE on Interactive, Cisco on Mixtral and SDXL, QCT on 405B). On the four language-model rows the B200 ran in FP4 and Hopper in FP8, so the gap is the platform's, not the silicon's alone; Stable Diffusion XL ran in FP8 on all three and is the like-for-like row.

02 / The card

## What the B200 SXM is and who rents it

The B200 is NVIDIA's first Blackwell datacenter GPU, and in its SXM form it is not a card you can slot into a server: it ships on the 8-GPU HGX B200 baseboard that goes into NVIDIA's DGX B200 and into OEM servers from Supermicro, Dell, HPE, Lenovo and others. NVIDIA's DGX B200 page gives 1,440 GB of HBM3e and 64 TB/s of memory bandwidth for the eight GPUs, which is 180 GB and 8 TB/s per GPU, 14.4 TB/s of aggregate NVLink bandwidth (1.8 TB/s per GPU over fifth-generation NVLink), two Xeon Platinum 8570 processors, 2 TB of system memory and a maximum draw of about 14.3 kW. Its FP4 figure of 144 petaFLOPS and FP8 figure of 72 petaFLOPS for the whole system are sparse numbers, and dense throughput is half of each. NVIDIA's MLPerf write-up describes each GPU running at a 1,000 W thermal design power.

You will see both 180 GB and 192 GB quoted. The 192 GB figure comes from early HGX B200 material that some providers still reproduce, including packet.ai, Hyperstack and Spheron on their pricing pages, while shipping DGX systems and every cloud that states the usable memory, from Lambda and Verda to Deep Infra and Oracle, expose 180 GB to software. Size models, KV caches and batches on 180 GB per GPU. The renters fall into three groups: training teams that want whole nodes or multi-node clusters with InfiniBand for weeks or months, inference operators who serve large models in FP8 or FP4 on one to eight GPUs, and engineers who want one card for a few hours to test or fine-tune a model that does not fit on an H100. For years the market served only the first group, which is why so many B200 offers still start at eight GPUs, and the single-GPU on-demand offers that appeared in 2026 are where the cheapest entry prices now sit.

03 / The price ladder

## How to read the B200 market

Every price on this page is in US dollars per GPU-hour excluding VAT, and where a provider sells only an 8-GPU node we divide the node price by eight. That gives a fair per-GPU number, but you still pay for eight, so the minimum unit matters as much as the rate. A managed cloud sells a fixed machine at one hourly price, and the question is what that price leaves out: Nebius bills disks separately, Northflank bills CPU and memory as their own lines, and AWS bills EBS volumes and egress on top, while GPUBank, Beam and Deep Infra put the CPU, memory, local disk and traffic inside the hour. A marketplace lets independent hosts set their own price on machines that vary from listing to listing, and on the B200 the marketplaces were not cheaper than the cheapest managed offers in mid-October, which reverses the pattern seen on consumer cards.

The other three kinds of offer trade something for the price. Spot or preemptible capacity is the cheapest hour on the market, from Nebius's dynamic floor of $0.99 to published fixed tiers between $2.49 and the $5.25 Spheron showed on GetDeploying, and the terms that matter are the notice period and whether the disk survives. Reserved capacity brings the B200 to between $2.99 and $4.49 when the contract runs one to four years, and to $6.79 to $9.86 on the short cluster reservations of Together AI and Lambda. Serverless platforms bill by the second of work from $5.50 to about $10 and suit bursty inference rather than long jobs. Stock is the last filter: on 15 October GetDeploying marked Lambda, Massed Compute, Hyperstack's on-demand flavour, Sesterce and Runpod Community out of stock, and Nebius's default B200 quota for VMs without a reservation is zero.

**Managed on demand** A fixed machine at one hourly price.

*Deep Infra**$3.69***

*packet.ai Dynamic**$3.75***

*Beam**$4.11***

*GPUBank**$4.15***

*GMI Cloud**$5.00***

*Massed Compute**$5.43***

*Hyperstack**$6.00***

*Lambda**$6.69***

*Verda**$7.87***

*Runpod Secure Cloud**$7.99***

*Together AI**$8.19***

*Nebius**$8.50***

**Marketplace** Hosts set their own price; machines vary.

*Akash Network**$5.00***

*Lium**$5.94***

*Runpod Community Cloud**$5.98***

*Vast.ai**$6.25***

**Spot and interruptible** Cheaper hours that can be taken back.

*Nebius preemptible**$0.99***

*GPUBank interruptible**$2.49***

*Vultr**$3.20***

*Verda spot**$3.94***

*CoreWeave spot**$4.26***

**Reserved** Committed for a term, paid either way.

*Vultr**$2.99***

*GPUBank reserved**$3.32***

*Civo**$3.79***

*Together AI**$6.79***

*Lambda 1-Click Clusters**$8.87***

**Serverless** Billed by active seconds.

*Koyeb**$5.50***

*Modal**$6.25***

*Baseten**$9.98***

USD per GPU-hour, read on 16 Oct 2026. Hover a dot for the offer; crowded dots keep their label in the table below. The large clouds are listed separately.

04 / Provider by provider

## What each price buys, and what it leaves out.

The captures below were taken on 16 Oct 2026 from each provider's own B200 or pricing page, so every figure can be read next to the conditions printed beside it.

gpubank.io/pricing

**GPUBank** All-in

$4.15/hr

Managed · 1 to 8 GPUs · per second

28 vCPU 283 GB RAM 2 TB NVMe IPv4 and all traffic Spot $2.49, reserved $3.32 Frankfurt and Ashburn

GPUBank sells the B200 SXM at $4.15 per GPU-hour on demand in machines of one to eight GPUs, and the hour already includes 28 vCPU, 283 GB of RAM and a 2 TB local NVMe disk per GPU, a public IPv4 address and all inbound and outbound traffic, with a persistent disk at $0.03 per GB-month as the only other line. Billing is per second from the moment the machine passes its health check, and a failed launch costs nothing. The interruptible tier at $2.49 gives five minutes of notice and stops rather than deletes the machine, and a week or a month reserved costs $3.32. Payment is a prepaid crypto balance from $25 that never expires, with a bonus that brings the effective on-demand hour to about $2.59 at $1,000. The B200 is offered in Frankfurt and Ashburn only, and machines stop at eight GPUs, so jobs that need several nodes belong with a cluster provider.

Checked 16 Oct 2026

deepinfra.com/gpu-instances

? **Deep Infra**

$3.69/hr

Managed container · 1 to 8 GPUs · per minute

16 vCPU 250 GB RAM 2 TB SSD No egress fees SSH into a container Stock not shown

Deep Infra, better known for its inference API, rents B200 GPU instances at $3.69 an hour for one GPU, $14.76 for four and $29.52 for eight, the lowest dedicated single-GPU list price we found on a provider's own page. Each GPU comes with 16 vCPU, 250 GB of RAM and 2 TB of SSD, billing is per minute and the page says there are no egress fees. What you get is a container reached over SSH with a choice of base images rather than a full virtual machine, and the page does not say whether the disk persists after a container is deleted, so keep weights in object storage. The machine is lighter than GPUBank's 28 vCPU and 283 GB per GPU, which matters for CPU-heavy data loading. At $3.69 it is 46 cents an hour cheaper than GPUBank's list price and more expensive than GPUBank's effective price after any top-up, which starts at about $3.46 with $25.

Checked 16 Oct 2026

beam.cloud/pricing

**Beam**

$4.11/hr

Managed machine · bare metal or VM

28 vCPU 256 GB RAM 2 TB NVMe No egress fees Persistent volumes to 1 TB included Clusters on request

Beam's on-demand B200 SXM6 machine starts at $4.11 an hour with 28 vCPU, 256 GB of RAM and 2 TB of NVMe, sold as one flat price per machine on bare metal or a VM. Beam includes persistent volumes up to 1 TB and charges $0.021 per GB-month above that, and its FAQ says there are no egress or bandwidth fees on any plan, which makes this the closest like-for-like offer to GPUBank's on the market at four cents less per hour. GetDeploying did not report its stock on 15 October, so check availability in the console before planning around it. For multi-node work Beam lists 8-GPU B200 nodes with InfiniBand on request, without a published price. The page lists the machine as "from" $4.11, so larger configurations or other regions may cost more.

Checked 16 Oct 2026

nebius.com/prices

**Nebius**

from $0.99/hr spot

Managed VMs · 1 or 8 GPUs · per second

20 vCPU 224 GB RAM On demand $8.50 Default B200 quota 0

Nebius has the lowest B200 figure on any provider's page, a preemptible price "from $0.99" per GPU-hour since 8 October, but it is a floor, not a quote: the spot price changes with capacity and demand as often as every 15 minutes and can rise to one cent below the on-demand rate, which went from $7.15 to $8.50 on 1 October. Each GPU comes with 20 vCPU and 224 GB of RAM, billing is per second, and disks are billed separately at $0.071 per GiB-month for a network SSD. A preempted VM gets 60 seconds of notice through SIGTERM, keeps its attached volumes and loses dynamic public IPs. The default quota for B200 VMs without a reservation is zero in both us-central1 and me-west1, so you need a quota increase or a reservation before either price can be booked.

Checked 16 Oct 2026

lium.io/gpu/b200

**Lium**

from $5.94/hr

Marketplace · per listing · per second

Host CPU, RAM and disk Helsinki, Ashburn, Sydney Headline moves daily

Lium is the marketplace with the clearest live B200 page: on 16 October it listed B200 pods in Helsinki, Ashburn and Sydney from $5.94 per GPU-hour, billed per second. The cheapest pod was eight GPUs in Helsinki at $5.94 each, followed by five GPUs in Ashburn and a single GPU in Sydney at $6.00, and an 8-GPU Sydney pod at $6.38. The price covers the whole pod, including the host's CPU, RAM and local disk, but those vary from one host to the next, so read each listing's machine before comparing it with a managed instance. The headline had read $5.49 and $5.93 on the previous day according to our research pass, which shows how fast marketplace B200 listings change hands. At $5.94 Lium is above every single-GPU managed offer in the $3.69 to $4.15 range, so its use is a whole node when the managed clouds are sold out.

Checked 16 Oct 2026

verda.com/pricing

**Verda**

$7.87/hr

Managed · 1 to 8 GPUs · Finland

30 vCPU 170 GB RAM Spot $3.94 2-year term 25 % off

Verda, the former DataCrunch, lists a 1x B200 SXM6 180GB instance at $7.87 an hour on demand, up from $6.85 in our September survey, and $3.94 on spot, with 30 CPUs and 170 GB of RAM per GPU. It also sells the B200 as serverless containers at $8.66, or $4.33 on spot, and as instant clusters at $7.87 per GPU. Long-term discounts on self-service instances run from 2 percent for one month to 8 percent for a year and 25 percent for two years, which puts a two-year B200 at about $5.90 per GPU-hour. Verda's spot tier is one of the few fixed B200 spot prices below $4 that were actually in stock, according to GetDeploying, which makes it the main alternative to GPUBank's $2.49 interruptible hour for European workloads that can live with interruption.

Checked 16 Oct 2026

### Where GPUBank wins, and where it does not

The provider card above gives GPUBank's rates; what matters beside them is how the hour behaves. The meter starts only once a machine has passed its health check, which GPUBank gives as a median of 48 seconds from launch, a failed launch is never billed, and a stopped machine costs only its persistent disk at $0.03 per GB-month. Machines come with 1, 2, 4 or 8 GPUs and the vCPU, RAM and NVMe scale by the same multiple, any Docker image can be launched, and ready-made templates cover vLLM, SGLang, Ollama, Open WebUI, llama.cpp, Unsloth, Axolotl, JupyterLab and PyTorch 2.8 with CUDA 12.8, the stacks most B200 renters reach for first. Every account and every team member can be given a monthly spend cap, which is the simplest guard against an 8-GPU machine left running over a weekend at $33.20 an hour.

Where GPUBank wins: its $2.49 is the lowest fixed spot price with published terms we found, its $3.32 week or month is lower than every short reservation we could read, and on demand only Deep Infra's $3.69 container and packet.ai's shared $3.75 tier undercut it at list, with Beam close at $4.11. Payment is a prepaid balance topped up in crypto from $25 to $5,000, and the top-up bonus of 20 to 60 percent brings the effective on-demand hour to about $3.46 with $25, $3.07 with $100 and $2.59 with $1,000. Where it loses: Nebius's dynamic spot can be far cheaper when capacity is idle, Vultr's and Civo's multi-year contracts are lower per hour, Deep Infra is cheaper at list for one GPU, there are only two regions, machines stop at eight GPUs so multi-node training needs a cluster provider, and card payments are not accepted for now, which rules GPUBank out for some companies whatever the price.

### The rest of the market

Below GPUBank's list price, packet.ai sells a B200 Dynamic tier at $3.75 in US East, which its own pricing page describes as shared GPU pods on scheduler-isolated infrastructure where the physical GPU may be shared with other tenants, and a Dedicated pod at $6.99. Sesterce appeared at $4.11 for an 8-GPU node and Hostinger at $4.50 for one GPU on GetDeploying, the first out of stock and the second with stock not reported, and neither price could be read on the provider's own page. The same aggregator listed Lyceum's single B200 at $6.49 and its spot tier at $3.29, both out of stock, Daytona at $3.59 for the GPU alone on a waitlist, and Spheron and EmpirioLabs in stock but at $11.06 and $11.90 per GPU, close to the hyperscaler rate. GMI Cloud now lists the B200 from $5.00 with limited availability through sales, up from $4.00 in September. Massed Compute sells an 8-GPU node at $43.46 an hour, or $5.43 per GPU, with 280 vCPU and 2,800 GB of RAM, Northflank gives $5.87 for the GPU line alone, and Hyperstack's 8-GPU flavour is $6.00, marked out of stock.

The better-known neoclouds sit between $6.69 and $8.60. Lambda's own page lists the B200 at $6.69 per GPU on an 8-GPU instance and $6.99 on one GPU, both out of stock on GetDeploying, Runpod's Secure Cloud is $7.99 with 28 vCPU and 283 GB of RAM, although several of Runpod's own articles still quote $6.79, E2E Networks in India is $7.99 billed per minute, Together AI's clusters are $8.19, and Crusoe and CoreWeave sell 8-GPU nodes at $8.50 and $8.60, CoreWeave's with 128 vCPU, 2,048 GB of RAM and 61.44 TB of local storage. Shadeform's directory quoted the B200 from $6.52 with no live instance listed, and the $3.74 Boost Run node that our September survey read there is no longer visible. UpCloud's page announces new B200 servers from 20 October at €4.50 an hour on demand and €3.38 on spot for one GPU, billed by the started hour.

On the marketplaces, Akash quotes a fixed $5.00 for the GPU alone, with CPU, memory, storage and the IP lease priced in each provider's bid, a figure we could read only on ComputePrices because Akash's own page loaded without its table. Runpod Community lists $5.98, out of stock, and Vast.ai's B200 page said "No current offers" while ComputePrices recorded $6.25 listings the same day; on Vast, storage keeps billing while an instance is stopped and bandwidth is billed per byte. Other spot tiers include Together AI's preemptible $4.09, CoreWeave's $34.11 node ($4.26 per GPU) and Vultr's $3.20 node, out of stock. Among serverless offers, Koyeb is $5.50 with 15 vCPU and 240 GB of RAM, Cerebrium $6.01, Modal $6.25 for the GPU alone, fal $7.99 at list, Runpod $8.64 and Baseten $9.98.

### The large clouds, per GPU inside an eight-GPU instance

Google Cloud A4 Spot

$4.95 /GPU-hr

8-GPU node · 224 vCPU, 3,968 GB RAM, 12,000 GiB SSD

Google Cloud A4, 3-year

$7.09 /GPU-hr

8-GPU node · Same machine; no on-demand price

Oracle BM.GPU.B200.8

$14.00 /GPU-hr

8-GPU node · Bare metal, 8 x 400 Gb/s RDMA

AWS Capacity Block

$14.21 /GPU-hr

8-GPU node · p6-b200.48xlarge for booked dates, paid up front

AWS p6-b200.48xlarge

$14.24 /GPU-hr

8-GPU node · 192 vCPU, 2,048 GiB, 8 x 3.84 TB NVMe; EBS extra

The large clouds sell the B200 only as 8-GPU instances, and their per-GPU price is two to four times the neocloud rate. AWS's p6-b200.48xlarge, with 192 vCPU, 2,048 GiB of RAM and eight 3.84 TB NVMe drives, is $113.9328 an hour on demand in US East, or $14.24 per GPU, on AWS's own price list. Its Capacity Block, which reserves the instance for chosen dates and is paid up front, is $113.666 per instance-hour, or $14.21 per GPU, in Ohio, N. Virginia, Oregon, Mumbai, Hyderabad, Stockholm and Paris, up from $12.36 in our September survey, with AWS's next scheduled price change in January 2027, so a reserved block now costs about the same as on demand and buys only certainty that the capacity exists on your dates. Oracle's price list gives $14.00 per B200 GPU-hour for the BM.GPU.B200.8 bare-metal shape.

Google Cloud does not sell its A4 machine on ordinary on-demand terms: its documentation says you must reserve capacity, use Spot or Flex-start VMs, or create a resize request in a managed instance group. On Google's own pricing page the a4-highgpu-8g, with 224 vCPU, 3,968 GB of RAM and 12,000 GiB of local SSD, costs $39.6336 an hour on Spot ($4.95 per GPU), $64.44 on Flex-start ($8.06), $90.22 in calendar mode, $88.9272 on a one-year commitment ($11.12) and $56.7072 on three years ($7.09). Microsoft's retail price list carries no ND B200 v6 SKU at all; the nearest Azure product is the ND GB200 v6, a four-GPU Grace Blackwell VM at $108.16 an hour in East US, which is $27.04 per GPU. What the premium buys is the surrounding platform, from IAM and private networking to compliance programmes, and teams that do not need it pay half or less elsewhere.

Next to the B200 sits the GB200 NVL72, a rack that joins 72 Blackwell GPUs and Grace CPUs in a single NVLink domain. CoreWeave lists GB200 NVL72 instances at $42.00 an hour for four GPUs, or $10.50 per GPU, and AWS's u-p6e-gb200x72 UltraServer costs $761.904 an hour on a Capacity Block in its Dallas Local Zone, or $10.58 per GPU. That is about two and a half times a single B200 on the cheaper managed clouds, and it buys the one thing an HGX B200 cannot give: NVLink bandwidth across 72 GPUs instead of eight, which large mixture-of-experts training and very large tensor-parallel inference need. For any job that fits in one 8-GPU node, the rack is the expensive way to rent Blackwell.

05 / Every offer

## The whole comparison, sorted by kind and price.

| Provider | USD / GPU-hr | Kind | Unit | What the price includes | Billing | Stock | Basis |
| --- | --- | --- | --- | --- | --- | --- | --- |
| DDeep Infra | $3.69 | Managed on demand | 1 GPU | 16 vCPU, 250 GB RAM, 2 TB SSD, no egress fees | Per minute | Not stated | Provider page |
| ppacket.ai Dynamic | $3.75 | Managed on demand | 1 GPU | Shared GPU on a scheduler-isolated pod | Hourly or monthly plan | In stock | Provider page |
| Beam | $4.11 | Managed on demand | 1 GPU | 28 vCPU, 256 GB RAM, 2 TB NVMe, no egress fees | Not stated | Not stated | Provider page |
| GPUBank | $4.15 | Managed on demand | 1 GPU | 28 vCPU, 283 GB RAM, 2 TB NVMe, IPv4, all traffic | Per second | In stock | Provider page |
| GMI Cloud | $5.00 | Managed on demand | 1 GPU | Limited availability, through sales | Not stated | On request | Provider page |
| Massed Compute | $5.43 | Managed on demand | 8-GPU node | 280 vCPU, 2,800 GB RAM per node | Not stated | Sold out | Provider page |
| Hyperstack | $6.00 | Managed on demand | 8-GPU node | 31 vCPU, 256 GB RAM per GPU | Not stated | Sold out | Provider page |
| Lambda | $6.69 | Managed on demand | 8-GPU node | 208 vCPU, 2,900 GiB RAM, 22 TiB SSD; $6.99 for 1 GPU | Not stated | Sold out | Provider page |
| Verda | $7.87 | Managed on demand | 1 GPU | 30 vCPU, 170 GB RAM | Not stated | In stock | Provider page |
| Runpod Secure Cloud | $7.99 | Managed on demand | 1 GPU | 28 vCPU, 283 GB RAM; storage extra | Per second | In stock | Provider page |
| Together AI | $8.19 | Managed on demand | cluster | Node CPU, RAM and NVMe not itemised | Not stated | Not stated | Provider page |
| Nebius | $8.50 | Managed on demand | 1 GPU | 20 vCPU, 224 GB RAM; disks extra | Per second | On request | Provider page |
| CoreWeave | $8.60 | Managed on demand | 8-GPU node | 128 vCPU, 2,048 GB RAM, 61.44 TB local | Not stated | Not stated | Provider page |
| Akash Network | $5.00 | Marketplace | 1 GPU | GPU only; CPU, RAM, storage, IP in the bid | Not stated | Not stated | Aggregator |
| Lium | $5.94 | Marketplace | 8-GPU node | Host CPU, RAM and disk; $6.00 for 1 GPU | Per second | In stock | Provider page |
| Runpod Community Cloud | $5.98 | Marketplace | 1 GPU | 28 vCPU, 283 GB RAM; storage extra | Per second | Sold out | Provider page |
| Vast.ai | $6.25 | Marketplace | 1 GPU | GPU rate; disk and bandwidth extra | Per second | Sold out | Aggregator |
| Nebius preemptible | $0.99 | Spot and interruptible | 1 GPU | Floor of a dynamic price; 60 s notice | Per second | On request | Provider page |
| GPUBank interruptible | $2.49 | Spot and interruptible | 1 GPU | Same machine; 5 min notice, stop not delete | Per second | In stock | Provider page |
| Vultr | $3.20 | Spot and interruptible | 8-GPU node | Bare metal, 256 threads, 3 TB RAM | Not stated | Sold out | Aggregator |
| Verda spot | $3.94 | Spot and interruptible | 1 GPU | 30 vCPU, 170 GB RAM | Not stated | In stock | Provider page |
| CoreWeave spot | $4.26 | Spot and interruptible | 8-GPU node | $34.11 per node-hour | Not stated | Not stated | Provider page |
| Vultr | $2.99 | Reserved | 8-GPU node | 48-month contract | Monthly | On request | Provider snapshot |
| GPUBank reserved | $3.32 | Reserved | 1 GPU | Same machine; 1 week or 1 month, prepaid | Per term | In stock | Provider page |
| Civo | $3.79 | Reserved | 8-GPU node | 36 months; 220 cores, 2.5 TB RAM, 7 TB NVMe | Monthly | On request | Provider page |
| Together AI | $6.79 | Reserved | cluster | 91 to 180 days; $7.99 for 3 to 30 days | Not stated | On request | Provider page |
| Lambda 1-Click Clusters | $8.87 | Reserved | cluster | 256+ GPUs, 2 weeks to 1 year; $9.86 at 16 | Per term | On request | Provider page |
| Google Cloud A4 Spot | $4.95 | Large clouds | 8-GPU node | 224 vCPU, 3,968 GB RAM, 12,000 GiB SSD | Per second | Not stated | Provider page |
| Google Cloud A4, 3-year | $7.09 | Large clouds | 8-GPU node | Same machine; no on-demand price | Commitment | On request | Provider page |
| Oracle BM.GPU.B200.8 | $14.00 | Large clouds | 8-GPU node | Bare metal, 8 x 400 Gb/s RDMA | Per hour | Not stated | Provider page |
| AWS Capacity Block | $14.21 | Large clouds | 8-GPU node | p6-b200.48xlarge for booked dates, paid up front | Per term | On request | Provider page |
| AWS p6-b200.48xlarge | $14.24 | Large clouds | 8-GPU node | 192 vCPU, 2,048 GiB, 8 x 3.84 TB NVMe; EBS extra | Per second | Not stated | Provider page |
| Koyeb | $5.50 | Serverless | 1 GPU | 15 vCPU, 240 GB RAM, 320 GB disk | Per second | Not stated | Provider page |
| Modal | $6.25 | Serverless | 1 GPU | GPU only; CPU and RAM metered | Per second | Not stated | Provider page |
| Baseten | $9.98 | Serverless | 1 GPU | Dedicated replica | Per minute | Not stated | Provider page |

"Aggregator" rows could not be read on the provider's own page that day; "Provider snapshot" rows come from a page dated by the provider.

06 / The real monthly bill

## The real cost of a B200 hour

Storage is where idle weeks cost money. GPUBank bills a persistent disk at $0.03 per GB-month and nothing else while a machine is stopped, so 500 GB of weights kept for a month costs $15. Nebius stops compute charges on a stopped VM but bills a network SSD at $0.071 per GiB-month whether it is attached or not, which is $35.50 for the same 500 GiB. Runpod charges $0.10 per GB-month for a pod's volume while it runs and $0.20 while it is stopped, or $0.07 for a network volume, and Vast keeps billing disk until the instance is destroyed and charges bandwidth per byte. Egress barely matters for a training run that checkpoints to local disk, but an inference endpoint pushing images or a team pulling a 400 GB model several times a week feels it: GPUBank, Beam, Deep Infra and Runpod do not charge for traffic, while packet.ai's own FAQ quotes $0.09 per GB on AWS for comparison.

Billing granularity decides the cost of short jobs. A run of 61 minutes on one B200 costs $4.22 at GPUBank, billed for 3,660 seconds, and $3.75 at Deep Infra, billed per minute. packet.ai's own comparison pages describe hourly billing for its Dynamic tier, so the same run would be charged as two hours, or $7.50, which turns the lowest shared list price into a bill 78 percent higher than the per-second one. UpCloud also bills by the started hour, so its new €4.50 hour costs €9.00 for 61 minutes of work. On an 8-GPU node the effect multiplies by eight, and on jobs that are short and frequent an hourly-billed price has to be 30 to 50 percent lower than a per-second price before it saves anything. Prepaid balances add a different trade: GPUBank's balance never expires and its bonus acts as a discount, but the money is committed at top-up, so the bonus is worth only what you actually spend, and a reserved order above $5,000 has to be paid in several top-ups.

The most expensive B200 hour is the one that produces nothing. At GPUBank a host failure restarts the machine on a healthy host with the same disk and credits the interrupted session in full, a missed spot notice credits the last 30 minutes, and a failed launch is never billed. Nebius gives preemptible VMs 60 seconds of notice through SIGTERM, enough to flush a small checkpoint and not enough to write a terabyte of optimizer state, and a preempted VM keeps its attached volumes but loses local SSD data and dynamic public IPs. On a marketplace reliability depends on the host, and so do a slower disk, a weaker CPU or a throttled link that make a data-loading-bound job run 10 to 30 percent slower than the GPU suggests, an estimate from our own experience rather than a measurement. Price a job by its expected useful hours, adding on spot the work lost per interruption times the expected number of interruptions.

### Three workloads, priced for a month

D **Deep Infra**

176 h GPU $649.44

Storage $0.00

Other Not stated

Month $649.44

Container disk; persistence not stated

p **packet.ai Dynamic**

176 h GPU $660.00

Storage $0.00

Other None

Month $660.00

Shared GPU; local NVMe included

**Beam**

176 h GPU $723.36

Storage $0.00

Other None

Month $723.36

Volume up to 1 TB included

**GPUBank**

176 h GPU $730.40

Storage $6.00

Other None

Month $736.40

$460.25 with $1,000 top-ups

All-in

**Lium**

176 h GPU $1,056.00

Storage $0.00

Other Not stated

Month $1,056.00

1 GPU listing; host disk while running

**Nebius**

176 h GPU $1,496.00

Storage $14.20

Other Not read

Month $1,510.20

Needs a B200 quota

At list price the single-GPU containers win this month: Deep Infra costs 176 x $3.69 = $649.44 and packet.ai's Dynamic tier 176 x $3.75 = $660.00, against $723.36 at Beam and 176 x $4.15 + 200 x $0.03 = $736.40 at GPUBank. The packet.ai figure buys a shared GPU, so it suits a team that can live with another tenant's load, and Deep Infra's container is lighter at 16 vCPU and does not state whether its disk persists, which matters when 200 GB of weights must be there every morning. Paid from a balance topped up in $1,000 steps, GPUBank's $736.40 costs $460.25 in cash and becomes the cheapest by about $190. Lium at $1,056.00 and Nebius at $1,496.00 plus $14.20 of disk, once a quota is granted, are not competitive for this pattern. The ranking would change only if sessions were short and frequent enough for per-minute rounding to count, which at eight-hour sessions it does not.

07 / Measured speed

## B200 performance, measured

The B200 earns its price on three kinds of work: serving large models in FP4 or FP8, where 8 TB/s of bandwidth and native FP4 multiply tokens per second; training and full fine-tuning from about 30 billion parameters up, where 1.44 TB of HBM in one node keeps the job on one machine; and long-context inference, where 180 GB per GPU holds a KV cache that would need two H100s. It is the wrong tool for LoRA on small models, image generation at moderate volume, transcription, embeddings and anything that fits comfortably in 80 GB, where an H100 does the same work for less money per output. Our rule: if a job uses less than half of an H100's memory and is not throughput-bound, a B200 costs more and finishes only a little sooner.

MLPerf is the only cross-vendor source of B200 numbers we trust, and it tests whole 8-GPU systems at full load. In MLPerf Inference v5.0, closed division, NVIDIA's DGX B200 (entry 5.0-0056) served Llama 2 70B at 98,443 tokens per second in the server scenario, against 33,072 for NVIDIA's 8-GPU H200 system (5.0-0060) and 31,106 for its DGX H100 (5.0-0057), all on TensorRT 10.8 and CUDA 12.8. Under the interactive variant, which caps time to first token at 450 ms and time per output token at 40 ms, the figures were 59,409, 19,424 and 14,009 (HPE, 5.0-0039). The B200 ran FP4 weights and Hopper ran FP8, so part of the threefold gap is the lower precision rather than the chip: on Stable Diffusion XL, where all three ran FP8, the B200 produced 30.4 images per second against 19.0 and 18.4, a gap of about 1.6 times. Rented nodes reach the same figures: in MLPerf Inference v5.1 Nebius's HGX B200 (entry 5.1-0077) served 101,611 tokens per second in the server scenario and Lambda's 99,994, on TensorRT 10.11 and CUDA 12.9, and NVIDIA's data page shows 102,398 in a later round, so software has added only a few percent since v5.0.

Training results are thinner. NVIDIA's MLPerf Training v4.1 write-up reports 2x per GPU for GPT-3 pre-training and 2.2x for Llama 2 70B LoRA fine-tuning from DGX B200 submissions in the preview category, measured against an HGX H100 submission. Those are NVIDIA's own per-GPU ratios from a round in which the B200 was not yet a shipping product, and later rounds have mostly reported GB200 NVL72 systems: NVIDIA's Training v5.0 write-up claims up to 2.6 times Hopper per GPU, but on racks that run at a higher power envelope than an HGX B200. Treat 2 to 2.2 times as the measured range for one HGX node and anything above it as a GB200 result.

**Llama 2 70B, server scenario** MLPerf Inference v5.0, closed, 8-GPU systems, TensorRT 10.8, CUDA 12.8; B200 in FP4, Hopper in FP8

8x B200 SXM

98,443 tokens/s

8x H200 SXM

33,072 tokens/s

8x H100 SXM

31,106 tokens/s

MLCommons results 5.0-0056 (DGX B200), 5.0-0060 (H200), 5.0-0057 (DGX H100), all NVIDIA submissions, availability category

**Llama 2 70B Interactive, server scenario** MLPerf Inference v5.0, closed, 8 GPUs, 450 ms to first token and 40 ms per output token; B200 in FP4, Hopper in FP8

8x B200 SXM

59,409 tokens/s

8x H200 SXM

19,424 tokens/s

8x H100 SXM

14,009 tokens/s

MLCommons results 5.0-0056 (NVIDIA), 5.0-0060 (NVIDIA), 5.0-0039 (HPE, best H100 SXM entry)

**Stable Diffusion XL, offline** MLPerf Inference v5.0, closed, 8 GPUs, FP8 on all three, TensorRT 10.8

8x B200 SXM

30.4 images/s

8x H200 SXM

19 images/s

8x H100 SXM

18.4 images/s

MLCommons results 5.0-0056 (NVIDIA), 5.0-0060 (NVIDIA), 5.0-0010 (Cisco, best H100 SXM entry)

CHEAPEST DEDICATED HOUR

$3.69

Deep Infra, one GPU on demand

LIST PRICE SPREAD

$0.99 to $14.24

Nebius spot floor to AWS on demand

B200 VS H200, LLAMA 2 70B

3.0x

MLPerf Inference v5.0, FP4 against FP8

08 / Per job, not per hour

## B200 against the H200, H100 and MI300X, per job

Per million output tokens on a 70B model at full batch, the B200 is the cheapest card GPUBank rents. The v5.0 server result is 12,305 tokens per second per GPU, or 44.3 million an hour, which at $4.15 is $0.094 per million and $0.059 at the $2.59 effective rate. The H200 at 4,134 tokens per second per GPU makes 14.9 million an hour, $0.167 per million at $2.49, and the H100 at 3,888 makes 14.0 million, $0.132 per million at $1.85. On Llama 3.1 405B offline the gap holds: $6.04 per million tokens on the B200 against $9.87 on the H200 and $9.42 on the H100 (QCT, 5.0-0065). For image generation it reverses: at the SDXL rates above, 1,000 images cost $0.30 on the B200, $0.29 on the H200 and $0.22 on the H100, because FP8 diffusion does not use Blackwell's FP4 advantage.

For training the answer is closer. A fine-tune that takes 80 GPU-hours on H100s costs $148.00 at $1.85; at NVIDIA's measured 2.2x it needs 36.4 B200 GPU-hours, $150.91 at $4.15, so the money is about the same and the job finishes in less than half the time. If the 2.5 times NVIDIA reported for eight Blackwell GPUs in a later Training round held on an HGX node, the job would take 32 B200 GPU-hours and cost $132.80, cheaper than the H100 as well. We have no measured H200 ratio from the same round; at H100 speed the same 80 GPU-hours on the H200 would cost $199.20 at $2.49, so the B200 is cheaper per training job than the H200 unless the H200 trains at least 1.32 times faster than the H100. The MI300X, at $1.65 with 192 GB, has no single-node MLPerf v5.0 result on these benchmarks (its only Llama 2 70B entry that round used four nodes), so we do not compute its cost per job; it is the cheaper way to hold a large model in memory, and the B200 is the faster way to push tokens through it.

MI300X

**192 GB** HBM3 · 5.3 TB/s

$1.65 /hr

Very large models on a single card, on ROCm

H100 SXM

**80 GB** HBM3 · 3.35 TB/s

$1.85 /hr

AI training, fine-tuning and production LLM inference

H200 SXM

**141 GB** HBM3e · 4.8 TB/s

$2.49 /hr

Long-context LLM inference and fine-tuning above 70B

B200 SXM

**180 GB** HBM3e · 8 TB/s

$4.15 /hr

Large-scale training and frontier model inference

GPUBank on-demand prices, all-in. Memory bar scaled to 192 GB.

09 / Rent or buy

## Rent or buy an 8-GPU HGX B200

The only realistic purchase unit is an 8-GPU HGX B200 or DGX B200 server, and dated sources disagree by more than $150,000. Northflank cites Exxact's TensorEX HGX B200 Essential configuration at $390,703.50 with dual AMD EPYC and 1,536 GB of RAM, Mercatus gives $400,000 to $500,000 with $450,000 as a typical integrated quote and a DGX B200 near $515,000 (16 July 2026), and Thunder Compute gives $450,000 to $550,000 (9 October 2026). Mercatus also notes that no functioning secondary market for B200 systems exists yet, so the resale value is the weakest input in any calculation. Around the server you need high-density colocation that can feed and cool about 14.3 kW, fabric ports if you will ever cluster, and someone to swap failed parts.

Our working case is a $450,000 server resold for 40 percent of its price after 24 months, which is our assumption, so depreciation is $270,000 over two years, or $11,250 a month. Colocation at $150 per kW-month for 15 kW adds $2,250, cost of capital at 8 percent a year on $450,000 adds $3,000 and spares and support add $1,000, all three our assumptions, for $17,500 a month before power. At the US commercial average of 13.97 cents per kWh for January to July 2026, 14.3 kW under load costs $2.00 per server-hour. Renting the same eight GPUs costs $33.20 a server-hour on demand at GPUBank, $26.56 reserved and $20.75 effective with $1,000 top-ups. Owning therefore breaks even at 17,500 / (33.20 - 2.00) = 561 hours a month against on demand, which is 77 percent of the month, at about 713 hours against the reserved rate, and never against the effective rate: at 730 hours owning costs $18,958 and renting $15,147.50.

The answer moves with the inputs. At Exxact's $390,703.50 and 50 percent resale the fixed cost falls to about $14,000 a month and owning breaks even with on demand at about 450 hours; at $550,000 and 40 percent resale it rises to about $20,700 and 662 hours. Staff time is left out, and so is any discount you can negotiate on the server. Buy if you already have high-density colocation, can keep the node above about 80 percent busy for two years, need the hardware on your own premises, or need dozens of nodes. Rent if your use is bursty or below that level, if you cannot wait for an allocation, or if a B300 or Rubin generation in 18 months would leave you holding a depreciating server.

Monthly cost of one 8-GPU HGX B200 server (8 GPUs) against hours of use a month, 0 to 730. Owning: price + $96,000 host + $2,250.00 a month of colocation − $180,000 resale over 24 months, plus 14.3 kW at the wall under load. Rent: per GPU-hour × 8.

| Hours a month | Rent at $4.15 | Rent at $2.59 effective | Own | Cheaper |
| --- | --- | --- | --- | --- |
| 100 | $3,320.00 | $2,075.00 | $17,699.77 | Rent |
| 250 | $8,300.00 | $5,187.50 | $17,999.43 | Rent |
| 400 | $13,280.00 | $8,300.00 | $18,299.08 | Rent |
| 550 | $18,260.00 | $11,412.50 | $18,598.74 | Rent |
| 650 | $21,580.00 | $13,487.50 | $18,798.51 | Own beats on demand; the effective rate beats owning |
| 730 | $24,236.00 | $15,147.50 | $18,958.33 | Own beats on demand; the effective rate beats owning |

Server price: Mercatus typical integrated quote, 16 July 2026; Exxact $390,703.50 via Northflank; Thunder Compute $450,000 to $550,000, 9 October 2026. The $96,000 shown as host is our cost of capital ($72,000) and support ($24,000) over 24 months; colocation is $150 per kW-month for 15 kW.

10 / What to choose

## The cheapest B200 depends on the unit.

Serving a 70B to 400B model to many users

B200 on demand or reserved, in FP4

At full batch it produces Llama 2 70B tokens for $0.094 per million at GPUBank's list price, against $0.132 on an H100 and $0.167 on an H200.

Fine-tuning a 70B model in bursts

B200 interruptible at $2.49, or a full node on demand

A measured 2.2x over the H100 per GPU means about the same cost per job in under half the time, and spot cuts the bill by 40 percent if you checkpoint.

Pre-training across many nodes

A reserved InfiniBand cluster or a GB200 NVL72 rack

Single-node offers cannot carry the fabric traffic; Lambda's clusters start at 16 GPUs and two weeks, and NVL72 adds NVLink across 72 GPUs.

Image generation or models under 80 GB

H100 SXM at $1.85 instead

In FP8 diffusion the B200 is only 1.6 times faster, so 1,000 SDXL images cost $0.22 on an H100 against $0.30 on a B200.

11 / Questions

## What people ask before renting one.

The cheapest B200 hour on any provider's page in mid-October 2026 is Nebius's preemptible tier from $0.99, but it is the floor of a dynamic spot price that can rise to $8.49 and stops the VM with 60 seconds of notice. The cheapest fixed spot price with published terms is GPUBank's $2.49, with five minutes of notice. For an uninterrupted dedicated GPU, Deep Infra's $3.69 container is the lowest list price we found, followed by packet.ai's shared Dynamic tier at $3.75, Beam at $4.11 and GPUBank at $4.15, which falls to about $2.59 with a $1,000 top-up. On multi-year contracts Vultr quoted $2.99 for 48 months. Which of these is cheapest for you depends on how much interruption, sharing or commitment the job can accept.

Sources

## Sources

1. 01 [Nebius documentation, Compute pricing](https://docs.nebius.com/compute/resources/pricing) B200 $8.50 on demand since 1 October, dynamic preemptible from $0.99 since 8 October, disk prices · read 16 Oct 2026
2. 02 [Nebius documentation, Preemptible VMs](https://docs.nebius.com/compute/virtual-machines/preemptible) SIGTERM 60 seconds before stop; volumes kept, dynamic IPs not · read 16 Oct 2026
3. 03 [Nebius documentation, Compute quotas](https://docs.nebius.com/compute/resources/quotas-limits) Default B200 quota for VMs without reservations is 0 · read 16 Oct 2026
4. 04 [Lium, B200 marketplace page](https://lium.io/gpu/b200) 22 B200s available from $5.94 per GPU-hour, billed per second · read 16 Oct 2026
5. 05 [GPUBank pricing](https://gpubank.io/pricing) B200 SXM rates, inclusions, billing, spot and reserved terms, top-up bonuses · read 16 Oct 2026
6. 06 [Deep Infra, GPU instances](https://deepinfra.com/gpu-instances) 1x B200 at $3.69/hour with 16 vCPU, 250 GB RAM, 2 TB SSD; per-minute billing, no egress fees · read 16 Oct 2026
7. 07 [Verda pricing](https://verda.com/pricing) 1x B200 SXM6 $7.87 on demand, $3.94 spot, serverless $8.66, long-term discounts · read 16 Oct 2026
8. 08 [fal pricing](https://fal.ai/pricing) B200 $7.99/h list, as low as $5.49/h · read 16 Oct 2026
9. 09 [NVIDIA DGX B200 product page](https://www.nvidia.com/en-us/data-center/dgx-b200/) 1,440 GB HBM3e, 64 TB/s, 14.4 TB/s NVLink, FP4 and FP8 figures, about 14.3 kW maximum · read 16 Oct 2026
10. 10 [NVIDIA technical blog, Blackwell in MLPerf Training v4.1](https://developer.nvidia.com/blog/nvidia-blackwell-doubles-llm-training-performance-in-mlperf-training-v4-1) 2x GPT-3 and 2.2x Llama 2 70B LoRA per GPU against Hopper; 1,000 W TDP per GPU · read 16 Oct 2026
11. 11 [Lambda pricing](https://lambda.ai/pricing) B200 instances $6.69 (8x) to $6.99 (1x); 1-Click Clusters $9.86 to $8.87 · read 16 Oct 2026
12. 12 [Oracle Cloud price list (cetools API), OCI Compute GPU B200](https://www.oracle.com/cloud/compute/gpu/pricing/) B200 GPU per hour $14.00 pay as you go (part B110978) · read 16 Oct 2026
13. 13 [Northflank blog, How much does an NVIDIA B200 GPU cost](https://northflank.com/blog/how-much-does-an-nvidia-b200-gpu-cost) $5.87 per GPU-hour; Exxact HGX B200 at $390,703.50 · read 16 Oct 2026
14. 14 [Beam pricing](https://www.beam.cloud/pricing) B200 machine from $4.11/hr with 28 vCPU, 256 GB RAM, 2 TB NVMe; no egress fees · read 16 Oct 2026
15. 15 [GetDeploying, Nvidia B200 rental prices](https://getdeploying.com/gpus/nvidia-b200) Stock flags, Sesterce, Lyceum, Daytona, Hostinger, Cirrascale, Vultr spot (aggregator, 15 October) · read 16 Oct 2026
16. 16 [Civo pricing](https://www.civo.com/pricing) 8x B200 node on 12, 24 and 36-month commitments; no on-demand price · read 16 Oct 2026
17. 17 [Together AI pricing](https://www.together.ai/pricing) HGX B200 $8.19 on demand, $4.09 preemptible, reserved $7.99 to $6.79 · read 16 Oct 2026
18. 18 [Koyeb pricing](https://www.koyeb.com/pricing) B200 $5.50/hr with 15 vCPU, 240 GB RAM, 320 GB disk · read 16 Oct 2026
19. 19 [Baseten pricing](https://www.baseten.co/pricing/) B200 $0.16633 per minute · read 16 Oct 2026
20. 20 [packet.ai pricing](https://packet.ai/pricing) B200 Dynamic (shared) $3.75 and Dedicated $6.99, US East · read 16 Oct 2026
21. 21 [GMI Cloud pricing](https://www.gmicloud.ai/pricing) B200 from $5.00, limited availability, contact sales · read 16 Oct 2026
22. 22 [Massed Compute pricing](https://massedcompute.com/home/pricing/) 8x B200 SXM6 $43.46/hr with 280 vCPU, 2,800 GB RAM · read 16 Oct 2026
23. 23 [Hyperstack GPU pricing](https://www.hyperstack.cloud/gpu-pricing) B200 $6.00 on demand, reservation from $5.10 · read 16 Oct 2026
24. 24 [Runpod pricing](https://www.runpod.io/pricing) B200 Secure Cloud $7.99, Community $5.98, serverless $8.64, storage rates · read 16 Oct 2026
25. 25 [E2E Networks GPU pricing](https://www.e2enetworks.com/pricing) B200 $7.99/hr with 32 vCPU, 400 GB RAM, per-minute billing · read 16 Oct 2026
26. 26 [Crusoe Cloud pricing](https://www.crusoe.ai/cloud/pricing) B200 HGX $8.50 per GPU-hour on demand · read 16 Oct 2026
27. 27 [CoreWeave pricing](https://www.coreweave.com/pricing) HGX B200 node $68.80 on demand, $34.11 spot; GB200 NVL72 $42.00 an hour for 4 GPUs · read 16 Oct 2026
28. 28 [Shadeform, H200 vs B200 directory page](https://shadeform.com/directory/gpus/compare/H200/B200) B200 from $6.52/hr; no live instance listed · read 16 Oct 2026
29. 29 [UpCloud pricing](https://upcloud.com/pricing/) New B200 prices from 20 October in EUR, billed by the started hour · read 16 Oct 2026
30. 30 [ComputePrices, B200](https://computeprices.com/gpus/b200) Akash $5.00, Hyperbolic $5.99, Vast.ai $6.25 (aggregator, 13 to 15 October) · read 16 Oct 2026
31. 31 [Vast.ai, B200 pricing page](https://vast.ai/pricing/gpu/B200) No current offers on 16 October · read 16 Oct 2026
32. 32 [Vast.ai documentation, Billing](https://docs.vast.ai/documentation/reference/billing) Storage billed while stopped; bandwidth billed per byte · read 16 Oct 2026
33. 33 [Cerebrium pricing](https://www.cerebrium.ai/pricing) B200 $0.00167 per second · read 16 Oct 2026
34. 34 [Modal pricing](https://modal.com/pricing) B200 $0.001736 per second (GPU only) · read 16 Oct 2026
35. 35 [Amazon EC2 On-Demand pricing (AWS price list feed, US East N. Virginia, Linux)](https://aws.amazon.com/ec2/pricing/on-demand/) p6-b200.48xlarge $113.9328/hr, 192 vCPU, 2,048 GiB, 8 x 3,840 GB NVMe · read 16 Oct 2026
36. 36 [Amazon EC2 Capacity Blocks for ML pricing](https://aws.amazon.com/ec2/capacityblocks/pricing/) p6-b200.48xlarge $113.666 per instance-hour; u-p6e-gb200x72 $761.904 · read 16 Oct 2026
37. 37 [Google Cloud documentation, accelerator-optimized machines](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines) A4 requires a reservation, Spot, Flex-start or a MIG resize request · read 16 Oct 2026
38. 38 [Google Cloud, accelerator-optimized VM pricing](https://cloud.google.com/products/compute/pricing/accelerator-optimized) a4-highgpu-8g: on demand N/A, Spot $39.6336, Flex-start $64.44, calendar $90.22, 1-year $88.9272, 3-year $56.7072 · read 16 Oct 2026
39. 39 [Azure Retail Prices API, ND GB200 v6](https://azure.microsoft.com/en-us/pricing/details/virtual-machines/linux/) Standard_ND128isr_NDR_GB200_v6 $108.16/hr in East US; no ND B200 v6 SKU listed · read 16 Oct 2026
40. 40 [packet.ai, packet.ai vs Verda comparison](https://packet.ai/articles/comparison/packet-ai-vs-verda) Dynamic tier described as scheduler-enforced multi-tenant, billed hourly · read 15 Oct 2026
41. 41 [MLCommons, MLPerf Inference v5.0 results (summary_results.json)](https://github.com/mlcommons/inference_results_v5.0) Closed datacenter results 5.0-0056 (DGX B200), 5.0-0057 (DGX H100), 5.0-0060 (H200), 5.0-0010, 5.0-0039, 5.0-0065 · read 16 Oct 2026
42. 42 [NVIDIA technical blog, Blackwell in MLPerf Inference v5.0](https://developer.nvidia.com/blog/nvidia-blackwell-delivers-massive-performance-leaps-in-mlperf-inference-v5-0) Llama 2 70B B200 against H200, FP4 on Blackwell · read 16 Oct 2026
43. 43 [Nebius blog, bare-metal-class performance in MLPerf Inference v5.1](https://nebius.com/blog/posts/bare-metal-class-performance-mlperf-inference) HGX B200 Llama 2 70B 101,611 tokens/s server, 101,246 offline (entry 5.1-0077) · read 16 Oct 2026
44. 44 [Lambda blog, MLPerf Inference v5.1 on HGX B200](https://lambda.ai/blog/lambda-mlperf-inference-v5.1) Llama 2 70B 99,993.9 tokens/s server, 102,725 offline, FP4, TensorRT 10.11, CUDA 12.9 · read 16 Oct 2026
45. 45 [NVIDIA data center inference performance page](https://developer.nvidia.com/deep-learning-performance-training-inference/ai-inference) Latest-round 8x B200 Llama 2 70B result (102,398 tokens/s server) · read 16 Oct 2026
46. 46 [NVIDIA technical blog, Blackwell in MLPerf Training v5.0](https://developer.nvidia.com/blog/nvidia-blackwell-delivers-up-to-2-6x-higher-performance-in-mlperf-training-v5-0/) Up to 2.6x per GPU over Hopper on GB200 NVL72; Llama 2 70B LoRA 2.5x · read 15 Oct 2026
47. 47 [Mercatus, NVIDIA B200 server price 2026](https://www.mercatus-ai.com/blog/b200-server-price) $400,000 to $500,000, typical $450,000; DGX B200 about $515,000; no secondary market (16 July 2026) · read 16 Oct 2026
48. 48 [Thunder Compute, NVIDIA B200 pricing](https://www.thundercompute.com/blog/nvidia-b200-pricing) 8-GPU HGX B200 system $450,000 to $550,000 (updated 9 October 2026) · read 16 Oct 2026
49. 49 [EIA, Electric Power Monthly, table 5.3](https://www.eia.gov/electricity/monthly/epm_table_grapher.php?t=epmt_5_3) US commercial average 13.97 cents per kWh, January to July 2026 · read 16 Oct 2026
50. 50 [Vultr, NVIDIA B200 page (read by the research pass on 15 October; blocked on 16 October)](https://www.vultr.com/products/cloud-gpu/nvidia-b200/) From $2.99 per GPU-hour on a 48-month contract · read 15 Oct 2026
51. 51 [Vultr public API, bare-metal plans](https://api.vultr.com/v2/plans-metal) vbm-256c-3072gb-8-b200-gpu at $68.00/hr, no locations listed · read 16 Oct 2026

What we could not confirm

Vultr's B200 page, which the research pass read on 15 October as "from $2.99 per GPU-hour on a 48-month contract", was behind a bot check on 16 October, and Vultr's public API listed its 8-GPU bare-metal plan at $68 an hour with no location, so the $2.99 is kept as a dated reading and the $3.20 spot node comes from GetDeploying. Akash's pricing page loaded without its table and Vast.ai's page showed no offers, so their $5.00 and $6.25 come from ComputePrices; Sesterce's $4.11 and Vultr's spot price come from GetDeploying. The $3.74 Boost Run node seen on Shadeform in September, Yotta Shakti's $2.91 and E2E Networks' $3.02 reservations could not be found again: Yotta's page now shows rupee prices per term and E2E's shows $7.99 on demand. Deep Infra and Beam do not show live stock and Lambda's on-demand billing granularity was not stated. packet.ai's pricing page gives no billing granularity, and the hourly billing in our 61-minute example comes from its comparison pages. Runpod's Secure Cloud B200 is $7.99 on its pricing page but $6.79 in several of its own articles. The 1,000 W per GPU is the thermal design power NVIDIA gives for its MLPerf Training v4.1 systems, not a datasheet figure. NVIDIA's current performance page lists no H100 or H200 rows, so the generation table uses MLPerf Inference v5.0, the latest round we could verify with all three cards on the same benchmarks; the MI300X has no single-node v5.0 entry on these benchmarks. The resale value, colocation rate, cost of capital, support cost and the 10 to 30 percent marketplace slowdown are our assumptions.

B200 SXM on GPUBank

## One B200, not a node of eight, billed by the second.

- 28 vCPU, 283 GB RAM, 2 TB NVMe, IPv4 and traffic in $4.15/hr
- Billed by the second after the health check
- Interruptible $2.49, reserved $3.32

[Launch a B200 SXM · $4.15/hr](https://console.gpubank.io/launch?gpu=b200-sxm)

[Compare 63 GPU clouds](https://gpubank.io/compare)

More GPU price guides

## Where to rent the other cards at the cheapest price.

[H200 SXM 141 GB HBM3e From $1.49 spot · $2.49 all-in](https://gpubank.io/rent-h200)

[H100 SXM 80 GB HBM3 From $1.11 spot · $1.85 all-in](https://gpubank.io/rent-h100)

[H100 PCIe 80 GB HBM2e From $0.81 spot · $1.35 all-in](https://gpubank.io/rent-h100-pcie)

[MI300X 192 GB HBM3 From $0.99 spot · $1.65 all-in](https://gpubank.io/rent-mi300x)

[RTX PRO 6000 96 GB GDDR7 ECC From $0.69 spot · $1.15 all-in](https://gpubank.io/rent-rtx-pro-6000)

[A100 SXM 80 GB HBM2e From $0.57 spot · $0.95 all-in](https://gpubank.io/rent-a100)

[L40S 48 GB GDDR6 ECC From $0.33 spot · $0.55 all-in](https://gpubank.io/rent-l40s)

[RTX 5090 32 GB GDDR7 From $0.29 spot · $0.48 all-in](https://gpubank.io/rent-rtx-5090)

[RTX 4090 24 GB GDDR6X From $0.14 spot · $0.23 all-in](https://gpubank.io/rent-rtx-4090)