Start Building →
GPU Clusters

Multi-node GPU.
Wholesale pricing.

InfiniBand-connected GPU clusters for distributed training and frontier models. Up to 1,024+ GPUs in a single fabric, at wholesale rates around 30% below retail. Flexible terms, dedicated TAM, custom storage.

Around 30% below retail · flexible terms · named technical account manager
CLUSTER · InfiniBand fabric 1,024 GPUs
128 nodes · 400 Gb/s per node · 1.8 TB/s NVLink
~30%
below retail pricing
1,024+
GPUs per fabric
400 Gb/s
InfiniBand per node
1.8 TB/s
NVLink GPU-to-GPU
Live inventory

Available clusters, ready to deploy.

Real capacity across our vetted provider network. Filter by GPU and request a wholesale quote. New regions and silicon land every week.

8 clusters
Featured
Available 29 JunIndia & Indonesia

B300 Hyperscale

HGX B300 · SXM6 270GB

Dell PowerEdge XE9680L with direct liquid cooling across Tier III data centers. 16 MW facility live, 8,000+ GPUs in stock. No lead-time risk.

GPU
HGX B300 8-GPU SXM6 270GB
CPU
Dual Xeon Platinum 8570
Memory
3 TB RDIMM DDR5-5600
Storage
8× 3.2 TB NVMe U.2
Network
ConnectX-8 800GbE OSFP
Cooling
Direct Liquid Cooling
8,000+ GPUsTier III DCNo lead-time risk
2-3 yr commit · 6-7 weeks to deploy
Available Jun-JulUnited States

B200 SXM — USA

HGX B200 SXM · NVLink + NVSwitch

HGX B200 SXM nodes with 180GB HBM3e per GPU and 3.2 Tbps InfiniBand fabric via 8× ConnectX-7 400G NICs. Supermicro 10U platform, redundant power.

GPU
8× NVIDIA B200 SXM
GPU memory
180GB HBM3e · 1,440GB/node
CPU
Dual Intel Xeon 6972P
Memory
1,536 GB DDR5-6400
Storage
30.72 TB Gen5 NVMe
Network
3.2 Tbps InfiniBand
NVLink + NVSwitchRedundant powerSupermicro 10U
Flexible commitment · Jun-Jul 2026
Available mid-MayUnited States

B200 Hyperscale

Blackwell NVL72 · 1,024 GPUs

128-node Blackwell NVL72 cluster with 1,024 B200 GPUs total. AMD CPU architecture, 800G Ethernet backend, 6 PB solid-state storage, dedicated Kubernetes control plane.

GPU
1,024× NVIDIA B200
Architecture
Blackwell NVL72
Storage
6 PB solid-state
Backend
800G Ethernet
Throughput
2× 400 Gb/s
Compliance
SOC 2
Vast Data optionalKubernetes readySOC 2 compliant
Min 12-month commitment
Available nowIndia

H200 NVL — 15+ nodes

Enterprise H200 NVL bare metal

Enterprise H200 NVL bare-metal nodes with Ubuntu 24.04 LTS pre-installed. RoCE v2 networking with NVLink bridge. Optional PFS and NFS storage tiers.

GPU
8× H200 NVL
CPU
192 vCPUs
Memory
1,024 GB
Network
2× 200 Gbps
Storage
PFS / NFS
OS
Ubuntu 24.04
RoCE v2NVLink bridgeIndia sovereign
Min 6 mo · 20% prepay
Available nowDallas, TX

H100 SXM5 — 35+ nodes

H100 SXM5 bare metal

8× NVIDIA H100 SXM5 bare-metal nodes in Dallas. Dual AMD EPYC 9454 (48 cores each), 1.5 TB DDR5-5600, and 8× ConnectX-7 400G single-port VPI networking.

GPU
8× NVIDIA H100 SXM5
CPU
2× AMD EPYC 9454 @ 2.75GHz
Memory
1.5 TB DDR5-5600
Network
8× ConnectX-7 400G VPI
Storage
960 GB M.2
Region
Dallas, TX
8× ConnectX-7 400GDDR5-560035+ nodes
Contact for commitment terms
Available nowIndia

H100 SXM — 10+ nodes

H100 SXM bare metal

8× H100 SXM nodes with dual AMD EPYC 9554 and 1.5 TB DDR5. 8× 400G + 2× 100G QSFPs, 100 Gbps node interconnect with up to 5 Gbps upstream.

GPU
8× H100 SXM
CPU
2× AMD EPYC 9554
Memory
1.5 TB DDR5
Network
8× 400G + 2× 100G QSFP
Storage
2× 960GB SSD · 4× 7.68TB NVMe
Interconnect
100 Gbps
100 Gbps interconnect8× 400G QSFPIndia sovereign
Contact for commitment terms
Available nowPhiladelphia, US

B200 SXM6 — 12+ nodes

B200 SXM6 180GB bare metal

8× NVIDIA B200 SXM6 180GB across 12+ nodes in Philadelphia. Dual Intel Xeon 6960P (72 cores each), 3 TB DDR5 RAM, RoCE fabric, and 15.36 TB local storage.

GPU
8× B200 SXM6 180GB
CPU
2× Intel Xeon 6960P
Memory
3 TB DDR5
Network
RoCE fabric
Storage
15.36 TB local
Nodes
12+ nodes
RoCE fabric3 TB DDR5Philadelphia US
Contact for commitment terms
Available nowUnited States

RTX PRO 6000 — US

Workstation-class AI & rendering

NVIDIA RTX PRO 6000 cluster on dual Intel Xeon 6960P Granite Rapids with 1.5 TB DDR5 and 8× 100 GbE networking. Built for rendering and workstation AI at wholesale pricing.

GPU
NVIDIA RTX PRO 6000
CPU
2× Intel Xeon 6960P (Granite Rapids)
Memory
1.5 TB DDR5-6400 ECC
Network
8× 100 GbE QSFP56
Storage
4× 7.6 TB NVMe (PCIe 5.0)
Region
United States
PCIe 5.0 NVMe8× 100 GbEUS region
Contact for commitment terms
Don't see your shape? We source across 20+ providers globally.
Available silicon

Frontier silicon, wired for scale.

Clusters are sold by the node at wholesale rates. The per-GPU figures below are retail starting points. Your committed cluster price lands around 30% lower.

A100 80GB

80GB HBM2eAmpere
from$1.43/GPU-hr
Deploy
Retail reference rates · wholesale quotes ~30% below · See full pricing →
How it works

From quote to running fabric.

01

Tell us your shape

GPU type, node count, region, and term. A wholesale quote comes back within one business day, typically ~30% below retail.

02

We build the fabric

InfiniBand cabling, NVLink topology, and custom storage tiering, validated end-to-end over 2-6 weeks.

03

Train at scale

A dedicated TAM, hands-on bring-up, and SLAs tuned to your run. Reserve capacity for as long as you need it.

Why Clusters

Built for distributed training.

NVIDIA Quantum-2 InfiniBand

400 Gb/s per node, non-blocking fabric. The interconnect frontier training actually needs.

5th-gen NVLink / NVSwitch

1.8 TB/s GPU-to-GPU inside the node. Aggregate HBM3e up to ~1.5 TB per node.

Up to 1,024+ GPUs

Scale a single job across hundreds of B200s in one coherent fabric.

Wholesale pricing

Committed multi-node deployments land around 30% below retail per-GPU rates.

Custom storage tiering

Up to 2 PB hot storage and S3-compatible cold tiers, sized to your dataset and checkpoint cadence.

Named TAM & SLAs

A dedicated technical account manager, hands-on cluster bring-up, and SLAs tuned to your workload.

Built for

Made for the largest runs.

Frontier pre-training

Multi-week runs across hundreds of GPUs.

  • 1,024+ GPU fabric
  • Non-blocking InfiniBand
  • Checkpoint-grade storage

Distributed fine-tuning

64-256 GPU jobs with tight interconnect needs.

  • NVLink + InfiniBand
  • Topology-aware scheduling
  • Wholesale node pricing

Reserved capacity

Guaranteed GPUs for a known program or season.

  • Flexible 6-24 mo terms
  • ~30% below retail
  • Named TAM
FAQ

GPU Clusters, answered.

For anything not here, reach help@packet.ai.

How is cluster pricing set?
Clusters are quoted per deployment based on GPU type, node count, fabric topology, storage, and term. Committed multi-node deals land around 30% below retail per-GPU rates. A typical 64-node B200 cluster lands at $2.80-$3.20/GPU-hr on 12-month terms.
What interconnect do you use?
NVIDIA Quantum-2 InfiniBand at 400 Gb/s per node (non-blocking), plus 5th-gen NVLink/NVSwitch at 1.8 TB/s GPU-to-GPU inside each node.
How large can a cluster get?
Up to 1,024+ GPUs in a single coherent fabric. Larger deployments are possible. Talk to us about your run.
How long does setup take?
Multi-node clusters take 2-6 weeks for InfiniBand cabling, topology validation, and storage provisioning.
What support is included?
A named technical account manager, hands-on cluster bring-up, and SLAs tuned to your workload.

Need many nodes?
Let's talk.

Tell us your shape and our team comes back within one business day, typically ~30% below retail.

Wholesale pricing · flexible terms · named TAM