NETWORKING & STORAGE FOR AI

Keep every GPU in sync with your data.

Distributed training depends on more than fast accelerators. Plan dedicated InfiniBand, NVMe, and shared storage as one data path for your GPU cluster.

Explore the data path Plan my cluster

Open network rack with blue and red fiber cables, switches, and a clear data-center aisle. — AI DATA CENTER · NETWORK FABRIC

NETWORK, STORAGE, AND TRANSFER, PLANNED TOGETHER

01Up to 800 Gb/s

InfiniBand XDR link on selected Blackwell systems.

02Up to 720 GB/s

Aggregate read throughput per pod with dedicated WEKA.

0320 TB

Egress included per node, per month.

04Up to 30 TB

Local NVMe on selected HGX nodes.

FOLLOW THE DATA

A clear path from workload to checkpoint.

Keep compute, cluster fabric, and storage under one configuration plan. The right path depends on the shape of the job, its parallelism, and how it saves data.

01
COMPUTE
Dedicated GPU nodes
Choose node or rack-scale capacity for your model, memory needs, and training or inference workload.
02
INTERCONNECT
InfiniBand for the cluster
Connect GPUs with dedicated high-speed fabric designed around communication between nodes.
03
DATA
Storage for each stage
Match local NVMe, parallel file storage, and S3-compatible object storage to datasets and checkpoints.

TENSORBAY AIFACTORY01 / 03
COMPUTE
**Dedicated GPU nodes**

04Networking and storage

Your GPUs. Working as one.

Bring compute, data, and storage together in infrastructure dedicated to your model, from the first training batch to the final checkpoint.

Plan my cluster

Dedicated networking

A network built for your GPUs.

Dedicated InfiniBand connects your cluster with bandwidth for distributed training and data exchange between GPUs.

800Gb/s

per InfiniBand XDR link, Quantum-X800 for Blackwell

Network specifications

Training fabric — Hopper: 8×400 Gb/s InfiniBand NDR per node (3.2 Tbps), rail-optimized non-blocking fat-tree, SHARP in-network reduction
Training fabric — Blackwell: Quantum-X800 InfiniBand XDR, 800G per link; GB200 racks add a 72-GPU coherent NVLink 5 domain at 1.8 TB/s per GPU
Oversubscription: 1:1, all tiers. No blended east-west fabric, no shared spine with other tenants
Front-end network: Dual 100 GbE per node, DDoS-protected, BYO-IP supported

Storage

Keep your data in stride.

Local NVMe, WEKA, and S3-compatible storage to load datasets and save checkpoints at the scale your project needs.

Up to

720GB/s

aggregate read throughput per pod, with dedicated WEKA

Storage specifications

Local scratch: 30 TB NVMe Gen5 per node, ~55 GB/s read — checkpoint staging without touching the network
Parallel filesystem: Managed WEKA, dedicated per cluster: up to 720 GB/s aggregate read per pod, POSIX + GPUDirect Storage
Object storage: S3-compatible, NVMe-cached, co-located with compute — $0.055/GB-month hot tier

Data transfer

Take your data further.

Free ingress and traffic between nodes, with an included egress allowance to help you plan your operation.

20TB

egress included per node, per month

Data transfer terms

Data transfer: Ingress $0 · 20 TB/node-month egress included, then $1.25/TB · inter-node $0. Free 100G Direct Connect on yearly reservations

Networking and storage are configured for your cluster architecture.View rates and terms

TESTED BEFORE HANDOVER

Start with a cluster you can trust.

Every cluster goes through 72 hours of load, memory, network, and thermal tests. Your team reviews the results before approving delivery.

72h

ACCEPTANCE TESTING

Load · memory · network · thermal checks

Talk with our team

AI WORKLOADS

TENSORBAY AIFACTORY
Model trainingDedicated GPUs and high-speed networking for long training runs.
TENSORBAY AIFACTORY
Fine-tuning & research8-GPU nodes for adapting models and iterating on experiments.
TENSORBAY AIFACTORY
Production inferenceMemory-rich systems for serving models on dedicated infrastructure.

NEVER BUILD ALONE

People who know the stack, within reach.

Bring your model, capacity goals, and timeline. TensorBay’s team can help shape the configuration and next steps.

Talk to sales

Engineer inspecting server racks inside a data center

01A configuration built around your project
02A proposal with a delivery timeline
03Acceptance testing before you start

05Invest in your next move

Compute to grow. Clarity to invest.

Choose your reservation term, compare capacity, and plan your investment. Here, the conversation starts with open pricing.

USD per hour, complete 8-GPU node

GPU nodes, USD per hour, complete 8-GPU node
GPU and memory	Monthly termCapacity reserved on a monthly term	Yearly termLower hourly rate
NVIDIAHGX B300288 GB HBM3e per GPU	Monthly from US$ 83.20/h Quote monthly	Yearly from US$ 58.50/h Quote yearly
NVIDIAHGX B200180 GB HBM3e per GPU	Monthly from US$ 72.80/h Quote monthly	Yearly from US$ 50.70/h Quote yearly
NVIDIAHGX H200141 GB HBM3e per GPU	Monthly from US$ 44.20/h Quote monthly	Yearly from US$ 31.20/h Quote yearly
NVIDIAHGX H10080 GB HBM3 per GPU	Monthly from US$ 28.60/h Quote monthly	Yearly from US$ 19.50/h Quote yearly

The hourly rate covers the complete 8-GPU node. Choose a term to request your quote.

SLA: 99.9% monthly + 15-min node replacement + goodput credits
Scale: 128–10,000+ GPUs · dedicated fabric per tenant
Billing: Per-minute metering inside the reservation · $0 ingress

Every detail accounted for

Your infrastructure. No blind spots.

See what is included and explore rates for storage, data transfer, and networking.

Talk to a specialist

Storage

Local NVMe scratch — 30 TB per nodeIncluded
Parallel filesystem — managed WEKAdedicated per cluster$0.09/GB-mo
Object storage — hot, S3-compatibleNVMe-cached, co-located$0.055/GB-mo
Object storage — archive$0.015/GB-mo
Snapshots$0.021/GB-mo

Data transfer

Ingressalways$0
Egress included20 TB/node-mo
Egress overageall sites$1.25/TB
Inter-node / east-westInfiniBand fabric$0

Networking

East-west InfiniBand fabricdedicated per clusterIncluded
Private VLAN / VPCIncluded
Additional public IPv4IPv6 free$4/IP-mo
BYO-IP — /24 v4 or /48 v6$200 setupFree
Managed firewall$4/node-mo
DDoS protectionIncluded
Direct Connect 10G / 100G100G free on yearly terms$450 / $1,800-mo

USD prices per 8-GPU HGX node-hour; rack-scale systems are quoted per rack. Taxes excluded. Managed Slurm and Kubernetes control planes are included. Storage and Direct Connect follow the rates above. Ask about long-term contract options.

Get my proposal

Frequently asked questions

Your next cluster starts with clarity.

Talk to a specialist

Is GPU pricing per GPU or per node?

Published HGX prices are in USD per hour for a complete node with 8 GPUs, within a monthly or yearly reservation. NVL72 systems are quoted per rack. Taxes are not included.

Which NVIDIA GPU should I choose for my AI project?

Compare GPU memory, network requirements and your training or inference workload. The hardware section lists H100, H200, B200, B300 and full rack systems. Our team can help size a configuration based on your model, dataset and capacity requirements.

Can I use a dedicated cluster for training, fine-tuning and inference?

Yes. TensorBay offers dedicated NVIDIA GPU infrastructure for model training, fine-tuning and inference. You can work with Slurm, Kubernetes or SSH and configure the software environment for your workload.

What happens after I request a quote?

Share the GPU, capacity and reservation term you need, or describe your project. Our team reviews the requirements and prepares a proposal within 48 hours, including configuration and delivery timing. Capacity and final terms are confirmed in the proposal.

08Give your idea room to grow

Design a data path that keeps your AI work moving.

Tell us about your GPUs, dataset, and training or inference pattern. We will help map the network and storage configuration around the workload.

A configuration built around your project
A proposal with a delivery timeline
Acceptance testing before you start

tensorbay@tensorbay.io