KAYTUS has upgraded KSManage, adding what it describes as full-stack, four-level operations and maintenance (O&M) visibility aimed at AI data centers. The update is built around a monitoring framework that spans components, servers and cabinets, clusters, and AI jobs, with the goal of improving fault localization and incident response in environments where heterogeneous infrastructure and workload dependencies can make troubleshooting slow.
KAYTUS ties the KSManage upgrade to common failure and response problems seen as AI data centers scale: more complex infrastructure, higher component failure rates under sustained high load, difficulty correlating hardware and network issues to specific AI jobs, and manual processes that can lengthen mean time to repair (MTTR). The company also claims a single outage can exceed USD 1 million in losses, and cites “industry data” that GPU power consumption has risen more than fivefold over the past decade, with cabinet power density rising to 20–50 kW and “gradually approaching 200 kW.”
On the telemetry side, KSManage collects real-time metrics including GPU and CPU utilization, video memory usage, power consumption, network bandwidth, and storage health, and aggregates operational events and network logs. KAYTUS says the platform uses automated topology discovery to track cross-node workloads and build a “measurement–log–trace” foundation, correlating device health with port-level telemetry through the job lifecycle. KSManage also includes real-time 3D modeling intended to visualize resource allocation, and KAYTUS claims troubleshooting efficiency improvements of up to 90%.
For predictive monitoring, KAYTUS says KSManage applies algorithms to analyze performance trends for critical components including GPUs and storage devices, flagging early indicators of abnormal wear and predicting hardware failure risk up to seven days in advance. It also monitors operating parameters such as load and temperature to mitigate issues under sustained high-load conditions.
At the workload correlation layer, KAYTUS says KSManage monitors bandwidth, latency, and packet loss, and “reserves a 20% bandwidth margin” to stabilize data transmission, targeting millisecond-level internal latency and packet loss below 0.01%. The company says this helps map hardware anomalies to specific training jobs and trace network faults, including optical module or fiber issues, through to training interruptions.
But operators should read the big percentage claims as aspirational until they see them proven in their own environments—instrumentation is only half the battle, and the other half is whether the correlation and automation logic holds up under real failure modes at scale. KAYTUS says KSManage can reach nearly 99.8% automated backup success rates, automatically identify up to 90% of root causes within five minutes using knowledge graphs and time-series anomaly detection, increase O&M efficiency by up to four times, and reduce TCO by up to 40%.
KAYTUS says KSManage is available as a trial that can be launched “in just a few clicks” via this form.
Source: KAYTUS
