The network edge to the GPU, end to end.
Move large datasets in and out of the cluster fast, encrypted between sites and resilient at every layer.
Always-on Connectivity
Diverse carriers
Independent physical routes for internet and inter-site traffic.
Bring-your-own addressing
Private space preserved, with tenant-scoped routing and NAT.
Edge protection
L3/L4 DDoS mitigation at the internet edge, extensible to L7.
Secure, reliable connections between sites.
Secure by Design
MACsec on private links; IPsec to enterprise environments over public networks.
Built for Bursts
Committed capacity with plenty of headroom for bursty traffic.
Always Redundant
Redundant data paths between every pair of connected sites.
Dedicated planes for contention-free performance.
Quantum-X800 InfiniBand at 800 Gb/s per port, non-blocking across the cluster.
Dedicated throughput for parallel loads and checkpoints, separated from the compute path.
Customer workloads isolated, connected to platform services.
A dedicated fabric to keep operational activity from competing with production traffic.
Rapid data movement for scale-up and scale-out.
NVIDIA Quantum-X800 InfiniBand Switches
NVIDIA ConnectX-9 SuperNICs
NVIDIA BlueField-4 DPUs
NVIDIA NVLink 6
800 Gb/s per port, adaptive routing, telemetry congestion control
Up to 1.6 Tb/s per GPU, for InfiniBand or Ethernet scale-out
Networking, storage, and security offload at 800 Gb/s per port
3.6 TB/s per GPU across the rack, double the previous generation
Failover paths at every layer.
Redundant fabric
Multiple spine-and-leaf paths absorb the loss of a link or switch.
Dual-homed nodes
Compute, storage, and control connect through separate switches.
Diverse carriers
Independent providers and physical routes to your infrastructure.
Proven for reliable, scalable operations before handover.
Pre-production Validation
Routing, control-plane behavior, and traffic flows tested before handover.
Expansion without Redesign
One design scales cleanly across racks, pods, and sites.
Full-lifecycle Operations
Monitored, maintained, and expanded as part of the running fleet.
