Just Launched Gruve PulseAI Platform, your private AI infrastructure, production-ready in under 2 weeks.PulseAI is live — private AI, ready in 2 weeks.

See PulseAI
Back to platform

PulseAI by Gruve Cisco UCS + Supermicro · Red Hat · NVIDIA

One AI platform.
Scale from 16 to 512 GPUs.

Private AI infrastructure with Cisco UCS or Supermicro servers, NVIDIA GPU compute, Red Hat OpenShift, and Cisco networking. Explore the architecture behind a platform managed by Gruve.

Explore topology and sizing ↓
Architecture at a glance Two network fabrics
4Scale
profiles
2Dedicated
network fabrics
512GPU target
configurable sizing

Platform building blocks

Built around the technologies customers ask for.

Choose Cisco UCS or Supermicro compute, with Red Hat OpenShift, NVIDIA GPUs, and a dedicated Cisco network fabric.

Red Hat platform RHOCP

Red Hat OpenShift provides the enterprise Kubernetes control plane for AI operations, policy, scheduling, and lifecycle management.

NVIDIA compute RTX PRO 6000
& B300 GPUs

Accelerated compute options for inference and training. GPU choice and density are matched to the supported server configuration.

Supermicro compute Supermicro
GPU servers

GPU worker and CPU server options for PulseAI, integrated with OpenShift and the Cisco frontend and backend fabrics.

Managed operations

Operations off your plate

Gruve monitors, patches and upgrades PulseAI around the clock. Your team keeps building AI, not running the platform.

24×7
Managed monitoring and operations
0
Added headcount for operations
14 days
To first served endpoint, from Ready for Install

One point of contact

Gruve detects, triages and escalates faults in your GPU servers, switches and storage to the responsible vendor.

Committed service levels

Essential, Standard or Premium, with a 15-minute severity 1 acknowledgment on Premium.

COMPUTE

The right server for each role.

Compare GPU counts and power ratings. Open any server for full specifications.

PSU ratings describe power-supply capacity, including redundancy. Actual server draw depends on the GPU, CPU, memory and workload; rack operating power is labeled separately.

Also available: 6 Supermicro systems

Roles reflect the PulseAI design. Specifications describe vendor options and maximums, not a preconfigured system. CPU, memory, drive and adapter combinations vary by SKU. GPU counts and power ratings reviewed 27 Sep 2026.

THE PULSEAI PLATFORM

From enterprise workloads to GPU infrastructure.

Compute is the foundation. Software, security and operations make it usable.

REFERENCE TOPOLOGIES

Frontend and backend topology by GPU scale.

Select 16, 32, 64 or 512 GPUs, then click any switch, server or software layer for details.

GPU workers

Control-plane servers are Cisco UCS CPU servers in every topology. GPU workers may be Cisco UCS or Supermicro. These are reference topologies; final ports, optics, power and server SKUs are confirmed during design.

16 GPUs reference topology1 rack · 2 GPU workers · 8 GPUs/worker
Select a component
Every leaf pair is dual-homed; every leaf has an uplink to every spine.
400G FRONTEND CLOSNX-9332HR
North–south inference, OpenShift services and storage RDMA
Dual 400G attachments to both frontend spines
SPINE TIER2 × NX-9332HR frontend spines
Full-mesh Clos · 2 spines × 2 frontend leaves
LEAF TIER2 × NX-9332HR frontend leaves · 1 redundant leaf pair
Dual frontend attachment from every leaf pair
800G BACKEND CLOSCISCO N9164E SPECTRUM-4
Dedicated GPU east–west traffic · Rail A / Rail B · BGP + ECMP only here
SPINE TIER2 × N9164E backend spines
Full-mesh Clos · 2 spines × 2 Rail A/B leaves
LEAF TIER2 × N9164E rail leaves · Rail A / Rail B pairs
Rail A and Rail B attachments into every GPU worker group

WHY THIS ARCHITECTURE

Scale the fabric.
Keep the traffic separated.

Explore 16 to 512 GPU targets. Physical rack and switch counts follow the selected server, NIC layout and sizing assumptions.

01

400G frontend on Nexus 9332

North–south services and VAST or NetApp storage RDMA use the frontend fabric as distinct QoS classes.

Result Fast application access and fast data delivery.
02

N9164E Spectrum-4 800G backend

GPU collectives stay on dedicated rail A and rail B paths, away from user and storage traffic.

Result Predictable east–west AI performance.
03

Backend routed Clos with BGP

The backend uses eBGP and ECMP for routed leaf-spine connectivity. Frontend routing and multihoming are selected during detailed design.

Result Clear protocol boundaries and smaller failure domains.

MANAGED BY GRUVE

Support across the platform lifecycle.

24/7 operations across compute, storage, networking and the AI platform.

Design

Size workloads, choose hardware and plan the fabric.

Deploy

Integrate infrastructure, OpenShift and platform services.

Operate

Monitor availability, investigate incidents and support workloads.

Evolve

Plan patching, upgrades and capacity growth.

Plan your PulseAI configuration.

Share your workload, server preference and capacity requirements with Gruve.

Discuss this configuration

The time is now

Cost of delay in compounding value is the real risk.

    Frequently asked questions

    Customers can purchase PulseAI Platform either through Cisco as a lead partner, through a channel partner, or direct from Gruve. Hardware components of the PulseAI Platform are offered through channel partners or through NeoCloud Providers (NCPs). Contact us for more details at info@gruve.ai

    PulseAI Platform includes the stack below, which brings AI Architecture from Hardware Layer to Agent Framework. PulseAI Platform brings this optionality to enterprise customers.

    • Agent Framework
    • Enterprise Security & Quota Management
    • Model Repository, Management
    • AI Security Layer
    • Operating Systems | Kubernetes
    • Hardware Infrastructure
      • GPU & CPU Compute
      • Storage
      • Network

    through channel partners

    PulseAI Platform is designed to give customers multiple options for consumption based on your business needs. Here are some ways to purchase and consume PulseAI Platform:

    • OPEX Model
      • Charged on a per GPU basis (GPU per hour)
      • Reserved model (fixed term length)
      • Custom designed to your specifications
    • CAPEX Model
      • Purchasing a pre-validated design through your channel partner

    Both models are then hosted based on your business requirements (on-premise, managed, or via an NCP)

    Gruve provides complete managed services. Customers only need to think about the business outcome that they want to build on top of the PulseAI Platform.

    Yes, PulseAI Platform is offered in small to large-scale GPU cluster environments for customers looking to pilot or take it to a full-scale production.