First fully flat fabric
Break the traditional multi-tier tree of data-center networks, remove Spine-layer switches, and flatten the entire interconnect.
Products and services
ZCube｜Next-generation AI cluster networking architecture
Replace the traditional multi-tier tree with a fully flat interconnect, remove Spine-layer switches, and combine single-rail plus multi-rail hybrid access with dedicated routing so GPU communication inside the cluster takes shorter, more balanced paths.
Inference latency reduced 20%–40%Inference throughput increased 10%–25%Network hardware spend reduced by 1/3Any two GPUs reachable in 2 hops
Product positioning
Trade hierarchy for efficiency, and determinism for performance
ZCube rebuilds the AI cluster compute network from the bottom up, using less network hardware, shorter forwarding paths, and more balanced traffic to fully unlock cluster compute efficiency.
Core capabilities
- First fully flat fabric, with Spine-layer switches removed
- Single-rail plus multi-rail hybrid access to balance AI traffic across switches
- ZCube dedicated routing and supporting network control software
- Structural global load balancing and physical multi-tenant isolation
Core capability architecture
From multi-tier stacks to a flat interconnect
Physical topology and routing policy are designed together to build shorter, more balanced communication paths for AI workloads.
Single-rail plus multi-rail hybrid access
A distinctive NIC access design plans exactly one optimal communication path between any two GPUs in the cluster, improving load balancing structurally.
Complete ZCube software stack
For the new underlying topology, we built a full set of network control software, including ZCube routing, so hardware topology and software routing work together.
Core hardware and software
ZCube networking architecture ecosystem
The ZCube ecosystem is built from ZCube switches, the agent-native operating system YZOS, the ZCube controller, and supporting deployment tools.
Efficient networking
Switch and optical-module counts drop by 1/3, and any two GPUs are 2 switch hops apart; multi-tenant traffic can be physically isolated, and the PFC broadcast domain is optimized.
Intelligent load balancing
Built-in AI Agents support congestion-aware adaptive routing and active path switching, plus load-balancing policies based on source and destination IP.
Fast fault recovery
Device state, traffic, microbursts, and port faults are monitored in real time; switches and the controller work together for millisecond-level self-healing.
Agent-native operating system YZOS
Supports intent-as-code, automatic construction of domain Skills, validation in real physical environments, and continuous iteration from production feedback.
ZCube switch
A core device for high-density AI cluster networking. Specifications follow official product materials.
- Switching capacity
- 102.4 Tbps
- Packet forwarding rate
- 21,000 Mpps
- Service ports
- 128 × QSFP112
- Typical power
- 2,432 W
- Chassis size
- 446 × 800 × 173.6 mm
- Port speeds
- 100G / 200G / 400G
ZCube controller: automate and verify large-scale networking
The ZCube controller covers architecture design, cabling error correction, configuration push, fault monitoring, and diagnosis. It supports automatic recovery from service faults, intelligent RoCE tuning, cluster expansion and hitless replacement, and tenant-level GPU orchestration, forming automated operations, service resilience, and long-term operations capability.
Value and results
ZCube versus traditional Clos
Using 51.2T switches (128×400G ports) to build a 16K H200 cluster (NICs at 2×200G) as an example, ZCube improves networking efficiency across device spend, communication path, load balancing, failure domain, and construction time.
33% fewer network devices
Switch count drops from 384 to 256, reducing both network devices and failure sources.
33% fewer 400G modules
400G module count drops from 49,152 to 32,768.
25% fewer fiber cables
Fiber cable count drops from 32,768 to 24,576.
Any two GPUs in 2 hops
Any two GPUs in the fabric pass through only 2 switches, for a shorter communication path.
Structural global load balancing
Architecture and path design together reduce hash collisions and uneven load.
Smaller failure domains
No Spine-switch failures; a switch interconnect link failure affects only some GPUs.
Physical multi-tenant isolation
Multi-tenant traffic isolation is raised from logical isolation to physical-link isolation.
Construction shortened to 5–7 days
Compared with a typical 7–12 days for traditional designs, workloads can go live sooner.
Validation
From deployment validation to production delivery
AI cluster deployments that show ZCube in real environments.
Domestic large-model companyThousand-GPU NVIDIA inference clusterProduction delivery
Large-model company NVIDIA thousand-GPU inference production cluster
Using a coding inference service and comparing against traditional ROFT, this case validates that ZCube unlocks more effective inference compute through a flat fabric and more balanced network communication.
15%
Average GPU inference throughput increase
40.6%
Time to first token (TTFT) P99 reduction
33%
Network hardware cost reduction
Training groundDomestic GPU inference clusterIndependent validation
Training-ground domestic GPU inference cluster
In a domestic AI computing ecosystem, compare ZCube with traditional Clos across collective communication, training, and inference.
14.5%
All-to-All throughput increase
22%–30%
TTFT P99 reduction
7%–10%
Inference throughput increase
A telecom operatorDomestic GPU clusterIndependent validation
Operator domestic GPU cluster
Compare ZCube with Clos and validate performance gains in collective communication, model training, and multi-tenant end-to-end inference.
119%
Collective communication performance increase
20.1%
Model-training throughput increase
11.5%
Inference latency reduction
Domestic accelerator companyDomestic thousand-GPU clusterProduction delivery
Domestic accelerator company thousand-GPU production cluster
A domestic thousand-GPU cluster has been live and stable for months, continuously carrying real traffic and serving real users.
Thousand-GPU scale
Domestic compute cluster deployed
Months
Stable production operation
Real users
Continuous service
Products and services
Other products and services
Networking, tuning, operations, and professional services work together to cover the full lifecycle of AI cluster networking.
AI computing tuning
YCOSAI computing tuning system
For large-scale AI workloads, systematically tune training/inference frameworks, communication libraries, network configuration, and host configuration.
Learn about the productAI computing operations
XOPSAI computing operations system
For large-scale training and inference on AI clusters, build an integrated operations platform across compute, network, and storage for faster problem localization and efficient troubleshooting.
Learn about the productProfessional services
Full Lifecycle ServicesComplete delivery from planning to go-live
One-stop delivery of network devices, compute-network tuning, and operations services to keep AI cluster networks stable, communication efficient, and SLAs met, helping clusters go live faster.
Learn about the productStart with the network
Make the next AI data center more efficient, starting with the network
Whether you are building a new cluster, upgrading an existing network, tuning performance, or deploying intelligent operations, we can assess how the network can unlock more effective compute.