Compute / Networking

Network Infrastructure

A network designed from first principles for distributed AI training, low-latency storage access and strict multi-tenant isolation — with logical segmentation, full observability and DDoS protection at every boundary.

In planning

The network is the nervous system of an AI compute platform. In distributed training, gradient synchronisation across hundreds of GPUs demands non-blocking, low-latency east-west traffic measured in microseconds. During inference serving, every millisecond of network latency adds to the user-facing response time. Storage access requires dedicated high-bandwidth paths that never contend with customer workload traffic. And multi-tenant isolation demands logical segmentation that is as robust as physical separation. NOVACORE network architecture addresses all of these requirements through a layered approach: an InfiniBand NDR fabric for GPU-to-GPU and GPU-to-storage communication, logically segmented Ethernet networks for customer workloads and management, diverse internet transit and peering for ingress and egress, and comprehensive observability across every segment. Each network function is separated, monitored and protected — no flat networks, no shared broadcast domains, no implicit trust between segments.

Network Objectives

High east-west traffic — non-blocking fabric for GPU-to-GPU and GPU-to-storage flows
Distributed training support — InfiniBand NDR with RDMA, GPUDirect and SHARP in-network reductions
Low-latency storage access — dedicated storage network with no contention from customer traffic
Multi-tenant isolation — VPCs, private subnets, firewalls and policy-based segmentation per customer
Internet transit — redundant upstream providers with BGP and DDoS mitigation at the edge
Private connections — VPN, dedicated interconnect and private link options for enterprise customers
DDoS protection — volumetric and application-layer mitigation at every network boundary
Management separation — out-of-band network independent of production data paths
Out-of-band access — IPMI, serial console and remote management on isolated physical paths
Full observability — sFlow, NetFlow, packet capture and latency monitoring on every segment

SEG

Logical segmentation

Public services, customer workloads, storage, cluster communication, management, monitoring and backup — each on its own logically isolated network with explicit routing policies, access control lists and ingress/egress filtering. No traffic crosses segment boundaries without explicit authorisation and logging.

Multi-tenant controls

Virtual networks per customer with private IP ranges, stateful firewalls, network policies and secure gateways. Each tenant operates in an isolated VPC with no default route to other tenants. Cross-tenant communication requires explicit peering, and all inter-tenant traffic is logged and auditable.

OBS

Observability

Traffic logging, egress controls, DDoS mitigation and latency monitoring at every boundary. Flow data exported to centralised analytics with automated anomaly detection. Packet-level capture available on demand for forensic investigation. Network performance dashboards provide real-time visibility into throughput, latency and error rates per segment.

Segmented Networks

Every network function operates in its own logically isolated domain. This segmentation is not a design aspiration — it is a deployment requirement enforced through routing policy, access controls and continuous monitoring from day one.

Public services — API gateways, portals, DNS, CDN origins
Customer workloads — isolated VPCs, private subnets, customer-defined IP space
Storage — dedicated network for NVMe fabric, object and block storage access
Cluster communication — InfiniBand fabric for GPU-to-GPU and GPU-to-storage
Management — orchestration, provisioning, configuration and identity services

Monitoring — telemetry collection, metrics aggregation, alerting infrastructure
Backup — isolated network for snapshot replication and backup data transfer
Facility systems — BMS, power monitoring, cooling control, physical security
Security systems — cameras, access control, intrusion detection, SIEM collectors
Out-of-band administration — IPMI, iDRAC, iLO, serial console, PXE boot

Design your network segmentation

Secure AI and high-performance computing for enterprises, governments and research.

Start a Technical Discussion Explore Partnerships

Network Objectives

High east-west traffic — non-blocking fabric for GPU-to-GPU and GPU-to-storage flows
Distributed training support — InfiniBand NDR with RDMA, GPUDirect and SHARP in-network reductions
Low-latency storage access — dedicated storage network with no contention from customer traffic
Multi-tenant isolation — VPCs, private subnets, firewalls and policy-based segmentation per customer
Internet transit — redundant upstream providers with BGP and DDoS mitigation at the edge
Private connections — VPN, dedicated interconnect and private link options for enterprise customers
DDoS protection — volumetric and application-layer mitigation at every network boundary
Management separation — out-of-band network independent of production data paths
Out-of-band access — IPMI, serial console and remote management on isolated physical paths
Full observability — sFlow, NetFlow, packet capture and latency monitoring on every segment

Segmented Networks

Public services — API gateways, portals, DNS, CDN origins
Customer workloads — isolated VPCs, private subnets, customer-defined IP space
Storage — dedicated network for NVMe fabric, object and block storage access
Cluster communication — InfiniBand fabric for GPU-to-GPU and GPU-to-storage
Management — orchestration, provisioning, configuration and identity services

Monitoring — telemetry collection, metrics aggregation, alerting infrastructure
Backup — isolated network for snapshot replication and backup data transfer
Facility systems — BMS, power monitoring, cooling control, physical security
Security systems — cameras, access control, intrusion detection, SIEM collectors
Out-of-band administration — IPMI, iDRAC, iLO, serial console, PXE boot