Scaling AI Clusters
The Infrastructure Challenge

As GPUs advance, new infrastructure bottlenecks emerge.
Data queues at the switches: the network sets the ceiling of the cluster.

Explore solutions

Move your mouse to see the network flow freely

As compute scales.
networks must keep pace.

As GPU compute leaps ahead, clusters need high-bandwidth, low-latency, high-throughput networks to carry the data.

8 × 50G lanesToday’s mainstream high-speed port

As AI demand grows. deploying
AI data centers becomes more complex.

As clusters keep growing, compute and networking become the key constraints on user experience.

10K

A proven scale for LLM trainingOne AI data hall

SUPRANETT

Network Infrastructure
Provider for Hyperscale

Supranett delivers end-to-end infrastructure for large-scale AI clusters. We combine high-performance Ethernet switches, SONiC-based networking software, and heterogeneous AI servers to support generative AI and distributed model training.

Click any row of GPU racks: its traffic balances across every spine

From Architecture to Operations
One Team. End-to-End Delivery.

  • Switching products from 400G and 800G up to 1.6T, covering switch silicon, optical modules and high-speed cables, for a high-bandwidth, low-latency interconnect foundation.

  • Full-stack system software R&D across OS architecture, protocol stacks, SAI and driver adaptation, including SupraNOS, our in-house platform built on SONiC.

  • Suprasophis combines Telemetry-based monitoring with automated operations, unifying monitoring, performance analysis, alerting and intelligent O&M across network, compute and facilities.

  • GPU and heterogeneous compute servers, storage, racks, micro-module data halls, power distribution and liquid cooling: the full infrastructure behind an AI data center.

  • We plan compute topology, four-plane isolation and rail-optimized networking alongside equipment selection, then refine the architecture through PoC validation and performance tuning.

  • Site survey, deployment, system integration and EVT/DVT/PVT/MP validation, tuned under real NCCL workloads for fully traceable delivery.

  • Maintenance, repair, spare parts, expansion and upgrades, and equipment reuse, keeping systems running reliably and preserving asset value.

  • Deep integration with NVIDIA Spectrum-X, ConnectX and BlueField DPU, advancing hardware-software adaptation, host-network coordination and heterogeneous compute compatibility.

Unlock the Full Potential
of Every GPU

SUPRANETT focuses on building hyperscale AI clusters. We bring together compute, network systems and cluster architecture into full-stack solutions, from design and integration to performance validation, so AI data centers deploy efficiently and run reliably.

About SUPRANETT
  • Senior AI cluster experts

  • Years of industry experience on average

  • Large-scale AI projects delivered

  • GPUs supported in production clusters

Hover a number to see it in the data hall

Products

Data center switches
for AI clusters

Built for high-speed Ethernet and lossless RoCE interconnects, our switches deliver high bandwidth and low latency to keep large clusters running reliably.

AI switch

CE11-128N

Learn more

Move your mouse to look around; hover the front panel to light the ports

Software

Software-Defined
Networking for AI

Network operating system

Open SONiC architecture, multi-silicon support, high performance and scalability. It runs on every switch.

  • One of the few companies in China recognized by the SONiC Foundation as a Premier Member, building on the open-source networking ecosystem to drive system innovation and hardware-software co-design.
  • Together with the Shanghai Institute for Advanced Study of Zhejiang University, we co-founded the Compute Optimization Joint Innovation Lab to explore SONiC in depth and build high-performance open AI networks.
About SupraNOS

Intelligent O&M platform

End-to-end network management for 10K-GPU AI clusters, removing every invisible cost across the vastness of compute.

About the platform
leaf-27 · Risk prediction

Solutions

Integrated AI Infrastructure
Solutions

  • For large-model training: a network isolated into security, management, training and storage domains. The training zone runs Leaf-Spine-Core with lossless RDMA Ethernet, delivered end to end from power and cooling to networking.View solution

  • A flat design with servers connected directly to aggregation switches, cutting hops and forwarding latency. Storage runs hybrid flash on Spine-Leaf, with four separate networks and no parameter network needed.View solution

  • One network split into training and inference zones that share core switching, security and storage. Four separate networks isolate the traffic while allowing resource reuse and elastic scheduling.View solution

  • Dual A/B planes of 400G Ethernet in full mesh with 1:1 non-blocking uplinks. Each server’s 8 NICs span both planes for load balancing and mutual redundancy, with failover in seconds.View solution

Four-domain isolation—training, storage, management, security; side rays show power and cooling.

Partners

Building Connections Through Trust

Building an open AI networking ecosystem with industry partners and research institutions.

Let’s talk about
Your AI Infrastructure

Tell us your cluster size and business goals, and our architecture team will get in touch.