// AI CLOUD

Private AI Cloud
Engineered as One System

QumulusAI combines dedicated compute, including supported bare-metal deployments, with cluster fabric, storage, orchestration, access control, security, and infrastructure support in environments designed around your workload.

DESIGN YOUR AI CLOUD

Every Layer Engineered for AI Performance

Start with compute, storage, and networking, then add platform services according to the workload, data path, isolation, and operating requirements.

// GPU+CPU INFRASTRUCTURE

Compute

Dedicated GPU and CPU capacity for customer-controlled workloads.

// AI DATA PATH

Storage

Object, file, block, local NVMe, and dedicated clusters designed around the AI data path.

// FABRIC + CONNECTIVITY

Networking

Cluster fabric, service traffic, management access, and private connectivity.

// PLATFORM SERVICES

Take Dedicated AI into Production

Managed Kubernetes, Lifecycle Management, Secure AI, and AI Factory add orchestration, lifecycle automation, security controls, and production platform capabilities around your compute, storage, and networking foundation.

Managed
Kubernetes

A supported orchestration foundation for dedicated AI infrastructure.

Lifecycle Management

Automated provisioning, validation, telemetry, and lifecycle control for bare-metal infrastructure.

Secure
AI

Design layered controls across firewalls, isolation, access, data protection, operations, and deployment-specific confidential-computing requirements.

AI
Factory

A complete production AI platform spanning software, data, infrastructure, and operations.

PLAN DEPLOYMENT

NVIDIA GPUs for Production AI Systems

Choose the NVIDIA platform around what matters most: model development, inference economics, or enterprise operating requirements.

Rack Scale Systems

Rack-scale systems deliver the highest performance for large-scale training, reasoning, and inference. VR200 NVL72 and GB300 NVL72 connect 72 GPUs in one NVLink domain with Arm-architecture NVIDIA Vera or Grace CPUs. VR200 specifications are preliminary; configuration, availability, region, and timing are confirmed for each deployment.

Vera Rubin NVL72

Next-generation agentic AI, long-context reasoning, and large-scale training

GPU COUNT

72 per rack

VRAM PER GPU

288 GB HBM4

VRAM PER RACK

20.7 TB

CPU CORES

3,168 cores

GB300

Rack-scale reasoning, training, and high-throughput inference

GPU COUNT

72 per rack

VRAM PER GPU

288 GB HBM3e

VRAM PER RACK

20.7 TB

CPU CORES

2,592 cores

Blackwell Systems

Blackwell systems are built for large-model training, reasoning, high-throughput inference, multimodal AI, and enterprise visual computing.

B300

Memory-intensive training, reasoning, and high-throughput inference

GPU COUNT

8 per node

VRAM PER GPU

288 GB HBM3e

VRAM PER RACK

2,304 GB

CPU CORES

256 vCPUs

B200

Large-model training and production inference at scale

GPU COUNT

8 per node

VRAM PER GPU

180 GB HBM3e

VRAM PER RACK

1,440 GB

CPU CORES

224 vCPUs

RTX Pro 6000

Enterprise AI, video, graphics, simulation, and visualization

GPU COUNT

8 per node

VRAM PER GPU

96 GB HBM3e

VRAM PER RACK

768 GB

CPU CORES

384 vCPUs

Hopper Systems

Hopper systems support large-model inference, training, fine-tuning, and HPC, with H200 providing more memory for larger models and longer contexts.

H200

Larger models, longer contexts, and memory-intensive HPC

GPU COUNT

8 per node

VRAM PER GPU

141 GB HBM3e

VRAM PER RACK

1,128 GB

CPU CORES

128 vCPUs

H100

Established Hopper infrastructure for training, inference, and HPC

GPU COUNT

8 per node

VRAM PER GPU

80 GB HBM3

VRAM PER RACK

640 GB

CPU CORES

224 vCPUs

Reserve Cluster Capacity with Room to Grow

Reserved clusters provide a predictable pool of dedicated infrastructure for steady workloads. Expansion planning can be incorporated into the design so future nodes, storage, or network capacity do not require a new architecture.

B300 Clusters

Memory-intensive model training, reasoning, and high-throughput inference.

CLUSTER SIZE

64 – 1,024 nodes | 512 – 8,192 GPUS

FABRIC

6.4 Tbps aggregate RoCEv2

ORCHESTRATION

Kubernetes & Slurm

GPUS PER NODE

8× B300 288 GB HBM3e SXM6

GPU MEMORY PER NODE

2,304 GB

LOCAL NVME PER NODE

15.36 TB

B200 Clusters

Large-model training, fine-tuning, and production inference.

CLUSTER SIZE

64 – 1,024 nodes | 512 – 8,192 GPUS

FABRIC

3.2 Tbps aggregate RoCEv2

ORCHESTRATION

Kubernetes & Slurm

GPUS PER NODE

8× B200 180 GB HBM3e SXM6

GPU MEMORY PER NODE

1,440 GB

LOCAL NVME PER NODE

15.36 TB

H200 Clusters

Memory-heavy Hopper training, inference, and HPC workloads.

CLUSTER SIZE

64 – 1,024 nodes | 512 – 8,192 GPUS

FABRIC

3.2 Tbps aggregate RoCEv2

ORCHESTRATION

Kubernetes & Slurm

GPUS PER NODE

8× H200 141 GB HBM3e SXM5

GPU MEMORY PER NODE

1,128 GB

LOCAL NVME PER NODE

15.36 TB

Vera Rubin NVL72 Clusters

Rack-scale reasoning, training, and high-throughput inference

CLUSTER SIZE

2-80 racks

GPU COUNT

72 per rack

VRAM PER GPU

288 GB HBM4

VRAM PER RACK

20.7 TB

CPU CORES

3,168 cores

GB300 NVL72 Clusters

Rack-scale reasoning, training, and high-throughput inference

CLUSTER SIZE

4-112 racks

GPU COUNT

72 per rack

VRAM PER GPU

288 GB HBM3e

VRAM PER RACK

20.7 TB

CPU CORES

2,592 cores