How much does an NVIDIA B200 GPU cost?

B200 GPU costs can look very different even when two providers advertise the same accelerator. One price may cover a GPU attached to a cloud instance, another a managed inference deployment, and another your share of an eight-GPU server that you must rent in full.

The useful comparison is what you will pay to run your model: the hardware allocation, time provisioned, supporting resources, and deployment model. Here is how B200 purchase prices and cloud rental costs compare in 2026.

A configured eight-GPU B200 server can cost hundreds of thousands of dollars. One concrete example is the Exxact TensorEX HGX B200 server: its Essential configuration costs $390,703.50, including eight B200 GPUs, dual AMD EPYC processors, and 1,536 GB of system memory.

That is a complete system price from a server supplier, not an NVIDIA retail price for one B200. Dividing it by eight gives roughly $48,838 per GPU, but that calculation also allocates the cost of the host hardware to each accelerator. It does not tell you what a standalone module would cost.

For a purchase decision, request a quote for the system you can actually install and operate. Include networking, rack space, power, cooling, support, and any additional storage. For scale, the NVIDIA DGX B200 specifications list eight GPUs, 1,440 GB of total GPU memory, and approximately 14.3 kW maximum system power. That power figure is a system maximum, not an estimate of your workload's average consumption.

An HGX-based server from an integrator and an NVIDIA DGX system are different products. Compare the complete configuration and support package before treating their prices as interchangeable.

B200 cloud rental examples include Northflank at $5.87 per GPU-hour for the GPU component, a $6.99/hour single-GPU Lambda instance, a $7.99/hour Runpod Secure Cloud Pod, and managed serving options with different resource and billing models. An eight-GPU allocation can have a lower rate per GPU while still producing a much larger hourly bill.

Prices are in USD, checked on 12 October 2026. These are different deployment and purchasing models, not equivalent configurations or a cheapest-to-most-expensive ranking. Capacity, storage, networking, and contract terms can affect the final bill.

The AWS and Google Cloud per-GPU figures divide the full instance price by eight. They do not mean you can rent one B200 at that rate. AWS EC2 Capacity Blocks charge for the reserved period; Google Cloud DWS Flex-start schedules capacity for workloads that can tolerate a flexible start time.

For Runpod, the rate above is specifically Secure Cloud Pods, not Community Cloud or Serverless. Likewise, Fireworks and Baseten's serving deployments should be evaluated against your model-serving needs rather than treated as interchangeable general-purpose VMs.

A single B200 at $6.99/hour costs $167.76 for 24 hours or $5,102.70 for 730 hours, before any additional charges. Monthly cost follows provisioned hours and the allocation you rent, even if your workload leaves some of that capacity idle.

Using Lambda's single- and eight-GPU instance rates illustrates the difference:

These calculations assume the allocation remains billable for the entire period. They exclude additional charges and do not predict training duration or inference throughput. A 730-hour month is a budgeting convention; actual calendar months vary.

For a fixed training or evaluation run, calculate the cost per completed job. For inference, measure cost per successful request or million output tokens at the latency and quality your application needs. Those measurements tell you more than the GPU's hourly rate alone. For published inference comparisons, MLPerf Inference Datacenter results provide a useful reference: check the model, scenario, and complete system configuration before applying a result to your workload.

The largest cost differences come from GPU count, provisioned time, host resources, and how effectively the workload uses the allocation.

• Minimum allocation: Check whether the deployment requires one, two, four, or eight GPUs. A model that fits on one GPU does not automatically benefit from an eight-GPU rental.
• CPU, RAM, and storage: Model loading, preprocessing, checkpoints, and temporary files need resources outside GPU memory. Check which are included and which are separate charges.
• Warm capacity and idle time: An inference service may keep GPUs available between requests. Reducing that capacity can save money, but model loading affects how quickly it can serve again.
• Location and commitments: Region restrictions, reserved periods, and contract terms can change the price and when you can access capacity.
• Software and utilisation: Model precision, batching, context length, and parallelism affect memory use and useful throughput. Test with your own requests or training configuration.

For a practical deployment and cost-control walkthrough, see how to rent GPUs for AI. It covers serving a model, testing its endpoint, and stopping unused compute.

Northflank runs GPU containers for model inference, training, and batch processing, with managed deployment, persistent storage, and service autoscaling. You can use a continuously running service for an inference API or a job for work that should finish after producing its results.

The managed-cloud GPU plan catalogue includes an eight-B200 configuration with 180 GB per GPU in selected regions. At the published B200 rate of $5.87 per GPU-hour, eight GPUs total $46.96/hour for the GPU component. Check the complete deployment estimate, including supporting resources, and confirm capacity in your intended region before launch.

For a B200 deployment, choose a compatible container image, configure the model's multi-GPU execution, and preserve weights or checkpoints on suitable storage. Assigning eight GPUs makes them available to the container; your serving or training framework still needs to use them. Northflank's GPU workload configuration guide covers runtime and storage preparation.

If your workloads need to run inside your own infrastructure, bring your own cloud (BYOC) lets Northflank act as the control plane, managing Kubernetes and GPU deployments inside your cloud account and VPC. GPU models and capacity then depend on your connected provider and cluster configuration.

Rent B200 capacity when you need to validate a workload, run a temporary project, or change the allocation as demand develops. Buying becomes more attractive when you have sustained utilisation and the facilities and team to operate the server.

Compare ownership over the hardware's useful life with the rental configuration you would actually use. Include installation, electricity, cooling, maintenance, financing, and idle capacity. A purchase-price-to-hourly-rate calculation alone leaves out much of the ownership cost.

For an initial evaluation, rent the smallest allocation that supports your model and measure useful throughput. That gives you a stronger basis for a longer commitment or hardware purchase.

These answers distinguish the accelerator price, the server configuration, and the cost of running a workload.

For deployment planning, obtain a supplier quote for a compatible B200 server or module configuration. B200 SXM hardware requires a suitable server platform; dividing an eight-GPU system price by eight does not create a standalone retail offer.

Yes. Lambda offers a single-B200 instance, and Runpod Secure Cloud Pods list a single-B200 configuration. Other offers require a larger allocation, so confirm the GPU count attached to the price before renting.

B200 can be cheaper per completed inference if its throughput improvement outweighs the higher deployment cost. The result depends on the model, precision, batching, context length, and latency target. Benchmark both configurations with the same workload; our B200 vs H200 comparison explains the hardware differences.

Not always. Some instances include local SSD, while persistent storage and data transfer can have separate charges. Modal separates GPU, CPU, and memory costs; other providers bundle host resources into an instance. Check the complete configuration rather than assuming the GPU rate covers every resource.

Compare other hardware options and the steps for deploying a GPU workload:

• Cheapest cloud GPU providers
• Best GPU for AI
• Rent H100 GPUs: pricing, performance, and providers