Harnets.AI rebuilds AI cluster networks with the original ZCube architecture and supporting technologies, using less network hardware, more balanced communication, and more efficient tuning and operations to raise effective cluster compute power.
On a thousand-GPU LLM serving cluster, GPUs, the software stack, and applications stayed the same; only the network architecture was upgraded from traditional ROFT to ZCube
15%
Average GPU inference throughput increase
More effective compute power unlocked by upgrading the network architecture alone.
40.6%
Time to first token (TTFT) P99 reduction
High-concurrency inference tail latency improved significantly.
33%
Network hardware cost reduction
Required switch and optical-module counts dropped sharply.
Focused on customer needs
For new clusters, existing-network upgrades, performance tuning, and intelligent NetOps, we combine traffic patterns, hardware, and runtime goals to deliver products and services across the full lifecycle of AI cluster networking.
Starting from business traffic, we jointly design the network architecture, devices, cabling, and deployment tooling.
Customer value
Build a high-throughput, low-latency AI computing foundation with less network hardware, speed up cluster delivery, and raise effective cluster compute.
Identify hotspots, PFC backpressure, and uneven link load, then upgrade the network architecture without changing business applications.
Customer value
Reduce tail latency and GPU idle time, and raise effective compute on existing clusters.
Profile the full path across training/serving frameworks, communication libraries, network, hosts, and storage to locate throughput and latency bottlenecks.
Customer value
Raise effective large-model training and serving output through cross-layer coordinated tuning.
Unify compute, network, job metrics, and logs to address scattered monitoring, slow fault localization, and heavy manual troubleshooting.
Customer value
Find anomalies and root causes faster, shorten MTTR, and form a reusable intelligent NetOps loop.
Products and services
Harnets.AI is not a single-device vendor. It is a solution provider built around ZCube and covering the full lifecycle of AI cluster networking.
AI cluster networking
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.
Learn about the product
AI infra tuning
For large-scale AI workloads, systematically tune training/inference frameworks, communication libraries, network configuration, and host configuration.
Learn about the product
AI infra operations
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 product
Professional services
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 productCore technology
ZCube removes the Spine layer used only for transit, splits Leaf switches into two groups with a complete bipartite interconnect, cutting hop count while improving load balance.
At the same cluster scale, the number of switches and optical modules is 2/3 of a Clos architecture.
Removing dedicated Spine devices simplifies forwarding nodes.
GPU-to-GPU forwarding across the fabric is only 2 hops, versus 3 to 5 hops in Clos.
Structural global load balancing avoids path-hash collisions and local hotspots.
No Spine-switch failure propagation; a single device failure affects only local nodes.
Link-level traffic isolation reduces interference when multiple jobs run in parallel.
Deployments
Open a case study to see the environment, validation or delivery approach, and full results.
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
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
Industry connections
ZCube-related work was published at ACM SIGCOMM 2025 and selected for the 2026 WAIC SAIL Award TOP30 and as an ODCC outstanding project.
Trade hierarchy for efficiency, and determinism for performance
SIGCOMM 2025 and SAIL TOP30
Multiple GPU types and multiple thousand-GPU clusters
Controller, toolchain, and full-lifecycle delivery
Whether you are building a new cluster, upgrading an existing network, tuning performance, or deploying intelligent NetOps, we can assess how the network can unlock more effective compute power.