NVIDIA B200 cloud prices

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 12 14. The cheapest fixed spot price with published terms is GPUBank's $2.49, with five minutes of notice 1. 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 7 8 9. On multi-year contracts Vultr quoted $2.99 for 48 months 28. Which of these is cheapest for you depends on how much interruption, sharing or commitment the job can accept.

A B200 costs between about $2.50 and $14.25 per GPU-hour in October 2026 on fixed terms, plus a dynamic spot floor of $0.99 at Nebius, and the spread is set by terms more than by the GPU. Single-GPU on-demand instances with CPU and RAM included run from $3.69 at Deep Infra and $4.15 at GPUBank to $7.87 at Verda, $7.99 at Runpod and $8.50 at Nebius 7 1 11 15 12. Neocloud 8-GPU nodes sit at $5.43 to $8.60 per GPU, multi-year reservations at $2.99 to $4.49, and the hyperscalers at $14.00 to $14.24 on demand, with Google offering no on-demand price at all 20 17 25 35 39 37. Multiply by eight for any offer sold only by the node.

You can rent a single B200 from a growing number of providers, though the large clouds still sell only 8-GPU instances. Deep Infra, GPUBank, Beam, Verda, Nebius, Runpod, Lambda, Koyeb, Modal and packet.ai all sell one GPU, and on GPUBank a machine of 2, 4 or 8 GPUs costs that multiple of the single-GPU price with the same multiple of vCPU, RAM and NVMe 7 1 16. AWS's p6-b200.48xlarge, Google's a4-highgpu-8g, Oracle's BM.GPU.B200.8 and CoreWeave's HGX B200 come only as eight GPUs, and Lambda's 1-Click Clusters start at sixteen 35 37 39 17. One B200 is enough to serve a model of up to about 170 GB in its deployed precision, which in FP4 covers models of roughly 300 billion parameters with room for the KV cache, and to fine-tune anything that fits in that memory.

Both figures come from NVIDIA, but 180 GB is what a rented B200 gives you. Early HGX B200 specifications listed up to 192 GB of HBM3e per GPU, and several providers still print that number, while NVIDIA's DGX B200 page lists 1,440 GB for eight GPUs, or 180 GB each 2. The clouds that state usable memory, including Lambda, Verda, Deep Infra and GPUBank, expose 180 GB, and so does Oracle's BM.GPU.B200.8 shape 16 11 7 39. The difference is capacity on the HBM stacks that is not exposed to software. When you size a model, a KV cache or a batch, plan on 180 GB per GPU and leave a few gigabytes for the CUDA context and the framework.

A B200 spot instance is worth it for any job that checkpoints regularly and can lose a few minutes of work: batch inference, evaluation sweeps and checkpointed training all qualify, and user-facing endpoints do not. Notice periods differ more than prices. GPUBank's interruptible tier at $2.49 gives five minutes of notice, runs a checkpoint hook, stops rather than deletes the machine and credits 30 minutes if a notice is missed 1. Nebius gives 60 seconds through SIGTERM and prices its spot dynamically between $0.99 and $8.49 12 14. Verda's fixed spot is $3.94, Together AI's preemptible $4.09 and CoreWeave's $4.26 per GPU on a node 11 18 17. Measure the cost of one interruption, the work since the last checkpoint plus the restart, before counting the discount as real.

On a marketplace an independent host sets the price and supplies the machine around the GPU, while on a managed cloud the provider fixes the instance shape and the billing rules. On Lium, Vast.ai, Akash or Runpod Community Cloud the CPU, RAM, disk speed and network vary from one listing to the next; Akash prices the GPU alone and adds CPU, memory, storage and the IP lease in each provider's bid, and Vast bills disk and bandwidth on top and keeps billing storage until the instance is destroyed 42 44. Reliability is the host's, and so is a machine slower than its listing suggests. A managed offer such as Deep Infra, Beam, Verda, Nebius or GPUBank gives a fixed vCPU and RAM allocation per GPU and published billing terms. In mid-October 2026 no marketplace B200 we could see was cheaper than the cheapest managed single GPU, the reverse of the usual pattern on consumer cards, so on this card a marketplace is where you look for a whole node when the managed clouds are sold out rather than for a lower rate 10.

The B200 is worth it over both cards for high-throughput inference of large models and over the H200 for training, while against the H100 for training it buys time rather than savings. In MLPerf Inference v5.0, eight B200s served Llama 2 70B at 98,443 tokens per second against 33,072 for eight H200s and 31,106 for eight H100s, which at GPUBank's prices is $0.094 per million tokens on the B200, $0.167 on the H200 and $0.132 on the H100 5. The B200 ran in FP4 and Hopper in FP8, and on SDXL at equal FP8 precision the gain shrinks to 1.6 times, which makes the H100 cheaper per image. For fine-tuning, NVIDIA's measured 2.2x over the H100 makes the cost per job about equal while the B200 finishes in less than half the time 3.

No, Google Cloud does not sell its B200 machine type on ordinary on-demand terms. Its documentation says that to provision A4 machine types you must reserve capacity, use Spot VMs, use Flex-start VMs or create a resize request in a managed instance group, and its pricing page shows N/A in the on-demand column 38 37. In us-central1 the a4-highgpu-8g, with eight B200s, 224 vCPU, 3,968 GB of RAM and 12,000 GiB of local SSD, costs $39.6336 an hour on Spot, $64.44 on Dynamic Workload Scheduler Flex-start, $90.22 in calendar mode, $88.9272 on a one-year commitment and $56.7072 on three years 37. Flex-start is the closest thing to on demand: you request capacity for up to a set duration and the scheduler starts the VM when it is available. AWS does sell the B200 on demand, but at $14.24 per GPU inside an 8-GPU instance, two to four times the neocloud rate, and what that premium buys is the surrounding platform rather than a faster GPU 35.

Azure's public price list does not include an ND B200 v6 SKU, so we found no B200 price on Microsoft's own pages. What the Azure Retail Prices API does list is the ND GB200 v6, a virtual machine with four Grace Blackwell GPUs, at $108.16 an hour on Linux in East US, East US 2 and West US 3, which is $27.04 per GPU and the highest per-GPU figure in this survey 40. The same API lists one-year and three-year reservations for that VM. Aggregator figures for an ND B200 v6 disagree with each other and could not be traced to Microsoft, so we leave the B200 row for Azure empty rather than guess. Teams committed to Azure for platform reasons should ask their account team for B200 capacity, and teams that are not will find the B200 far cheaper elsewhere.

A single B200 running all month, 730 hours, costs $3,029.50 at GPUBank on demand, $2,423.60 reserved and about $1,893 effective with $1,000 top-ups, and $2,693.70 at Deep Infra's $3.69 1 7. At other providers the same month costs about $4,365 at Runpod Community's $5.98, $5,745 at Verda's $7.87 and $6,205 at Nebius's $8.50, before storage 15 11 12. An 8-GPU node costs $24,236 at GPUBank on demand, $19,388.80 reserved, $21,549.60 at Deep Infra, $22,133.60 on Civo's 36-month contract and $82,976.18 on an AWS Capacity Block 25 36. Add persistent storage, $0.03 per GB-month at GPUBank and $0.071 per GiB-month at Nebius, and egress on clouds that bill it. For part-time use, multiply your real hours by the rate, rounding each session up first on providers that bill by the hour.

Buy only if you can keep an 8-GPU server busy for more than about 77 percent of every month for two years and already have colocation that can take about 14.3 kW. With a $450,000 server, 40 percent resale after two years, colocation at $150 per kW-month, 8 percent cost of capital, $1,000 a month for support and US commercial power at 13.97 cents per kWh, owning costs about $17,500 a month plus $2.00 per server-hour of power 45 47. That breaks even with GPUBank on demand at $33.20 per node-hour at about 561 hours a month and with the reserved rate at about 713 hours, while against the effective rate after $1,000 top-ups, $20.75 per node-hour, owning never breaks even within a month. Buying makes sense for data that must stay on your premises, for fleets of dozens of nodes, or for teams with cheap power and space.

An 8-GPU HGX B200 server, the only practical way to buy a B200, costs about $390,000 to $550,000 in 2026 depending on vendor and configuration. Northflank cites Exxact's TensorEX HGX B200 Essential configuration at $390,703.50, Mercatus puts a typical integrated quote at $450,000 within $400,000 to $500,000 and the DGX B200 near $515,000, and Thunder Compute gave $450,000 to $550,000 in October 22 45 46. Per GPU, Mercatus says the B200 lists at $30,000 to $40,000 in 8-GPU volumes with street quotes of $45,000 to $50,000, but an SXM module is useless without a baseboard. NVIDIA publishes no DGX price list, and Mercatus notes that no functioning secondary market for B200 systems exists yet, which makes resale value the most uncertain input in any rent-or-buy calculation.

A DGX B200 draws up to about 14.3 kW according to NVIDIA, and NVIDIA's MLPerf write-up gives each Blackwell GPU in those systems a thermal design power of 1,000 W 2 3. At the US commercial average of 13.97 cents per kWh for January to July 2026, a fully loaded server uses about $2.00 of electricity an hour, or about $1,460 over a 730-hour month, before the cooling overhead of the facility 47. Most traditional colocation racks are provisioned at 5 to 10 kW, so a single B200 server can need a high-density rack or a dedicated power feed, and some HGX B200 chassis are liquid-cooled. When you rent, power and cooling are inside the hourly price, which is part of why a B200 hour costs more than the server price divided by its lifetime hours.

On a provider with stock, a B200 can be running in about a minute, while reserved and hyperscaler capacity can take days or weeks to arrange. GPUBank gives a median of 48 seconds from launch to a machine that has passed its health check, in Frankfurt or Ashburn, and its meter starts only then 1. Nebius requires a quota increase above its default of zero, Google requires a reservation, Spot or Flex-start, AWS Capacity Blocks are booked for future dates, and Lambda's clusters start at two weeks and 16 GPUs 13 38 36 16. GetDeploying marked Lambda, Massed Compute, Hyperstack's on-demand flavour, Sesterce and Runpod Community out of stock on 15 October, so check live availability before planning around a price 41.

GPUBank is paid from a prepaid US dollar balance topped up in cryptocurrency, and its B200 hour includes the vCPU, RAM, local NVMe, an IPv4 address and all traffic, so the only possible extra line is a persistent disk at $0.03 per GB-month 1. Accepted coins include Bitcoin, Ethereum, USDT, USDC, Monero and more than a dozen others, top-ups run from $25 to $5,000, and the bonus of 20 percent from $25, 35 percent from $100 and 60 percent from $1,000 brings the effective on-demand hour to about $3.46, $3.07 and $2.59. Cards are not accepted for now, which rules GPUBank out for teams that can only pay by card. The balance never expires, every account and every team member can have a monthly spend cap, billing is per second with no minimum, and if a host fails during a job the machine restarts on a healthy host with the same disk and the interrupted session is credited in full.