BlackwellLive Marketplace Asks Available

NVIDIA B200

B200 is the Blackwell datacenter GPU most teams mean when they say Blackwell: 192 GB HBM3e, 8 TB/s, NVLink 5, and a Transformer Engine with FP4. It is the step up from H200 when you need more than 141 GB or you want FP4. A common node shape is about 30 vCPUs and 184 GB of host memory beside the 192 GB of VRAM.

Browse Live Marketplace →All Hardware Fleet

Flexible Compute Modes

Deploy NVIDIA B200 Your Way: Docker Pods or Root KVM VPS

NovaPod lets you decide how you want to harness GPU power. Whether you need an instant pre-cached container stack, a full root virtual machine with dedicated public IP, or a multi-node NVLink cluster, pay only for what you run with second-by-second crypto billing.

Docker GPU PodCold Boot < 15s

Fast, Lightweight AI Containers

Pure GPU compute without hypervisor overhead. Boot straight into PyTorch, vLLM, ComfyUI, or JupyterLab.

1-second precision metering
37+ Pre-cached ML framework templates
Zero idle fee when instance is stopped

Deploy as Docker Pod

Root KVM VPSPersistent NVMe

Full Root Virtual Machine

Dedicated operating system (Ubuntu 24.04 LTS / Windows GPU Server) with complete sudo root SSH access from minute one.

Dedicated public IPv4 (all ports open)
Pause anytime; pay only $0.00014/GB/hr for disk
Run Docker-in-Docker, k3s, custom drivers

Deploy as Root VPS

Multi-GPU Cluster900 GB/s NVLink

Supercomputing & Distributed Training

Pool up to 8× NVIDIA B200 on a unified node fabric with Quantum-2 InfiniBand RDMA for multi-node scale.

Zero-loss all-reduce gradient exchange
PyTorch FSDP, DeepSpeed Zero-3 ready
Native GPUDirect Storage & RDMA fabric

Configure Cluster

Hardware Specifications

Architecture: NVIDIA Blackwell
VRAM: 192 GB HBM3e
Bandwidth: 8.0 TB/s
Tensor Cores: 5th generation
CUDA cores: 20,480
FP64: 40 TFLOPS
FP32: 80 TFLOPS
TF32 dense: 1,125 TFLOPS
FP8 dense: 4,500 TFLOPS
FP4 dense: 9,000 TFLOPS
NVLink: 1.8 TB/s
Reference host: 30 vCPUs, 184 GB RAM
Local disk: Often a few hundred GB of NVMe, listed on the offer
8-GPU pool: About 1.5 TB
TDP: About 1,000 W
Form factor: SXM6

Blackwell in one GPU

B200 is the Blackwell datacenter GPU: 192 GB of HBM3e at 8 TB/s, NVLink 5 at 1.8 TB/s, and a second-generation Transformer Engine with FP4. Against H100 that is about 2.4 times the memory and bandwidth. Against H200 it is more memory, more bandwidth, and a numerics format Hopper does not have.

The menu host is about 30 vCPUs and 184 GB of RAM beside the 192 GB of VRAM. An 8-GPU HGX node pools about 1.5 TB, which is the box people use to serve a 405B model in FP4 with a long context.

FP4, without the brochure

FP4 is 4-bit floating point on Blackwell. On layers that tolerate it, throughput is roughly double FP8 and the weights take half the memory. The engine mixes FP4, FP8, and FP16. It is an inference tool first. Do not flip a pre-training run to FP4 because a banner said the chip supports it.

Published inference comparisons often show large jumps versus H100 at FP4. Treat those as vendor or MLPerf conditions, then measure your own model. A 13B chat model will not show a 15× gain. A 100B+ serve that was sharding across H100s might.

Where it sits in the stack

Under about 70B parameters, H200 or even H100 is usually the better dollar. Between 70B and the point where 192 GB runs out, B200 is the default Blackwell card. Above 192 GB, move to B300. The rack version of this silicon is GB200 NVL72, which is a different page from a single SXM6.

Software expectation is current PyTorch, TensorFlow, or JAX, CUDA 12.4-class libraries, and NCCL that knows Blackwell. Bring the image you already tested. The deploy flow on this site is the same as any other GPU: pick the offer, pick a template, pay per second.

What the Numbers Mean in Production

VRAM: 192 GB
Versus H100 memory: 2.4×
Versus H200 bandwidth: ~1.7×
FP4: Native

Storage & Tooling Infrastructure

High-Speed Storage, Native Ports & Crypto Rails

Every NVIDIA B200 instance on NovaPod connects directly to high-throughput infrastructure designed to eliminate I/O bottlenecks and onboarding delays.

25 Gb/s Shared Network Storage

Mount large model checkpoints and training sets across one or multiple instances simultaneously. Data outlives single pods and stays accessible when instances stop.

JupyterLab, VS Code & SSH

Connect from your local terminal with standard SSH keys or open web-based IDEs directly in browser tabs without managing complex reverse proxy configs.

Non-Custodial Multi-Chain Rails

Deposit USDT, USDC, SOL, ETH, or BNB to your dedicated personal deposit address. Confirmed on-chain in seconds with zero mandatory identity verification.

Target Workloads for NVIDIA B200

Large-model training

192 GB and 8 TB/s are for runs that stall on H100 memory or bandwidth, including mixture-of-experts jobs that want Blackwell numerics.

100B and larger fine-tunesMixture-of-experts trainingFP8 training, FP4 only with a checked recipeMultimodal towers that do not fit Hopper

405B-class serving

An 8-GPU node pools about 1.5 TB, which is the box people use for a 405B model in FP4 with a long context.

Tensor parallel across NVLinkFP4 servingHigh batch inference100B and larger, single-node

Generative models

Video, 3D, and audio models that need both the framebuffer and the Tensor Cores.

High-resolution videoLong video contextCode and agent models above 70BBatch generation

Research past Hopper

Use B200 when the experiment needs FP4 or more than 141 GB. A 13B chat model will not show a dramatic gain.

New precision recipesModels that shard across H100s only for memoryLong-context ablationsStop the pod between runs

When to Choose NVIDIA B200

Pick B200 for FP4 and 192 GB

This is the default Blackwell card when H200 memory is no longer enough, or when you want 4-bit inference on a current stack.

Pick H100 under about 70B if 80 GB fits

H100 is mature and usually a lower ask. B200 is wasted capacity if the profile never leaves 80 GB.

Pick H200 for 141 GB without Blackwell

Same Hopper software as H100, more memory, no FP4. A common middle step for 70B–180B serving.

Pick B300 or a GB200 rack above this

B300 is 288 GB on one GPU. GB200 is the NVL72 built from B200s and Grace. Both are different pages from a single SXM6.

What fits

A 405B model can sit in FP4 on one B200. A 200B-class model in FP8 still has room for cache. An 8-GPU node pools about 1.5 TB.

FP4 in practice

Four-bit math is an inference format. It cuts memory and raises throughput when the stack supports it. Training stays on FP8 or BF16 unless you have a recipe that has been checked.

Frequently Asked Questions

How much does a B200 cost here?

Whatever the live ask is when you open the marketplace. We bill that ask per second. An 8-GPU node is eight of those asks. There is no standing rate on this page, and no 20-minute minimum.

What is the cheapest way to run one?

Take the lowest ask you trust, use FP4 only if the model was exported for it, keep the batch high enough to fill 192 GB, and stop the machine between runs. A reclaimable offer is cheaper only when you can lose the machine.

Can I rent it by the hour?

You can run it for an hour. The meter is seconds. A 20-minute benchmark costs about 20 minutes.

Is B200 better than H200?

On memory and on FP4, yes: 192 GB at 8 TB/s versus 141 GB at 4.8 TB/s, and a numerics format Hopper does not have. Under about 70B, H200 is usually the better buy.

What makes it different from H100?

About 2.4× the memory and bandwidth, fifth-generation Tensor Cores, and FP4. Real speedups depend on the model. A small chat model will not look 15× faster.

What is FP4?

4-bit floating point on Blackwell. On layers that tolerate it, throughput is about double FP8 and the weights take half the memory. The engine mixes FP4, FP8, and FP16. It is an inference tool first.

Can I train a trillion-parameter model?

An 8-GPU node has about 1.5 TB of HBM. That is a serious training box, and it is not automatically a 2T-parameter run. Past one node you need a fabric and a framework that already scales: FSDP, DeepSpeed, or Megatron.

Which frameworks?

PyTorch, TensorFlow, and JAX on a CUDA 12.4-class image, plus NCCL that knows Blackwell. Bring the container you already tested. vLLM and TensorRT-LLM are the usual serving paths.

What does NVLink change?

1.8 TB/s between GPUs, so gradient sync and tensor parallel do not sit on PCIe. Eight B200s on NVLink are a different machine from eight B200s on a motherboard.

Is buying smarter?

Street prices for the GPU are in the tens of thousands of dollars, and an 8-GPU server with power and 400G networking is several times that. Rent unless utilization stays high for a year and you already run the datacenter.

Inference only?

Yes, especially at 100B and up. Under 70B, H100 or A100 usually wins on dollars per token.

Dedicated or reclaimable?

Read the offer. A reclaimable card can disappear. Use it for batch work and sweeps with checkpoints. A long training run that cannot resume belongs on a card the host will not take back.

Alternative Hardware Options

Ready to Deploy NVIDIA B200?

Launch an instance on the live marketplace in under 15 seconds. Fund your balance with cryptocurrency and pay per second with zero commitments.

Launch on Marketplace →