How NTT DATA and NVIDIA help networks spot trouble before the alarm

Say your broadband connection has become slower or less reliable. It hasn’t failed completely and the network hasn’t flagged it yet, so everything still looks broadly normal to the operation team. However, your employees are already noticing the difference.

We call that silent network degradation — a gradual change in performance that can affect service before conventional alarms make the problem obvious. For communications service providers, it presents a challenge: How do you find a problem that hasn’t yet announced itself?

NTT DATA and NVIDIA are exploring one answer through a gigabit passive optical network (GPON) use case with a leading communications service provider in Brazil. GPON is the fiber-optic technology that connects a service provider’s network to homes and businesses.

The approach brings our network operations support capabilities together with NVIDIA agentic AI technologies to help operations teams move from spotting unusual behavior to understanding possible causes and deciding what to do next.

We want to make network assurance less about reacting to visible failures and more about identifying and investigating emerging issues before they escalate.

The limits of waiting for a network alarm

Traditional network operations are designed to respond to known conditions. When a device fails or a measurement crosses a defined threshold, an alarm gives engineers a clear starting point.

But not every service problem begins with a dramatic event. Performance can deteriorate gradually, with early signs spread across different measurements or parts of the network. Each signal may look unremarkable on its own, but taken together, they point to a developing issue.

In the case of GPON networks, understanding those signals may require teams to compare performance trends, network relationships and previous incident reports. When that information is distributed across systems and experts, emerging issues remain difficult to see.

Proactive assurance starts somewhere else: Look for patterns that depart from normal behavior, even when no conventional alarm has fired, and flag them for investigation.

Fault detection is only the beginning

Finding something unusual is useful, but it doesn’t resolve the operational problem. An engineer still needs to answer a series of practical questions: Where is the issue occurring, and how important is it? What might be causing it? Has the team seen something similar before? What’s the most appropriate next action?

This is where agentic AI could help. Rather than treating every detected anomaly as an isolated alert, AI can support an end-to-end investigation by bringing together relevant network information, examining plausible explanations, consulting operational knowledge and preparing a recommendation for the team.

NVIDIA technologies, including NemoClaw and OpenShell, provide the foundation for bringing these agentic capabilities into controlled operational environments.

In the Brazil GPON use case, we’re exploring how that process can be made more structured and repeatable — from the early signs of degradation to a probable cause and a recommended action.

Process only the data that’s needed

In network operations, feeding an AI model as much data as possible can be wasteful. Networks generate huge volumes of information, and asking an AI system to examine everything for every potential issue can add cost and complexity without necessarily improving the answer.

It’s more practical to process only what’s relevant to the issue being investigated.

When a monitoring process notices a GPON traffic pattern drifting from its usual behavior, AI may at first need only a high-level view of performance over time to decide whether the pattern deserves attention. If the signal looks significant, the investigation can then bring in more detailed measurements and the relevant network context. If the evidence points toward a particular area, the AI can then consult information about connected equipment, service relationships or similar past incidents.

Use the right intelligence for the right task

Different stages of an investigation call for different kinds of intelligence. Spotting a change in a performance trend is not the same task as working through several possible root causes. A simple classification or prioritization task may not need the same depth of reasoning as a complex diagnosis.

You need to match the level of AI capability to the job. Efficient methods can help screen and organize potential cases, while more capable reasoning can be reserved for investigations that need a richer understanding of network context and operational history.

The NVIDIA Nemotron family of AI models is being considered for tasks where AI needs to interpret operational information and support deeper diagnostic reasoning. The broader collaboration combines NTT DATA’s telco and network operations experience with NVIDIA’s AI capabilities to explore how these approaches can work in a real operational setting.

But choosing a capable general-purpose model is not the whole answer. Telco troubleshooting has its own terminology, alarm patterns, equipment relationships, vendor documentation and established procedures. The clues that matter to an experienced network engineer may be highly specific to the domain.

For AI to be useful in this environment, it needs relevant telco knowledge and operational context.

Why telco-specific knowledge is key

We have separately evaluated a customized NVIDIA Nemotron 3.5 Lightning model for telco root-cause analysis and incident troubleshooting. In that benchmark, the customized model achieved 86% accuracy, compared with 72% for the base version, and recorded the highest accuracy among the models we evaluated. It also required significantly less input context.

These results are encouraging, but they do not prove that the same performance will be achieved in every network or operational environment. They do suggest that a model adapted to telco troubleshooting can make more focused use of the information it receives.

That distinction is important. Providing more information is not always the best way to improve an investigation. Organizing relevant knowledge and adapting AI to the domain can help it interpret evidence more effectively, while keeping the investigation focused on the case being handled.

What could this mean for telco operators?

Finding problems earlier is one of the clearest potential benefits. A degradation pattern that would otherwise remain unnoticed until its impact becomes more apparent can instead become a case for proactive investigation.

But the potential extends beyond detection. Once a possible issue is identified, AI can bring together the information needed to understand what’s happening and help the team move from abnormal network behavior to a probable cause and recommended response.

Applied at scale, this type of approach could:

• Identify silent degradation earlier, including network conditions that may remain below conventional alarm thresholds.
• Reduce investigation and diagnosis time by up to 50%, accelerating the journey from detection to probable root cause and recommended action.
• Reduce manual investigation effort by up to 40% by bringing together relevant network and operational knowledge during the investigation.
• Create more consistent diagnostic processes that can be extended across a broader range of network scenarios.

One use case, a much bigger opportunity

GPON silent degradation is a specific starting point, but the same principles could apply to other network performance issues, fault management and service assurance. In telco operations, teams face a similar challenge: making sense of growing volumes of signals, identifying what’s important and responding before the service impact becomes more serious.

The goal is not to introduce AI for its own sake, but to help operations teams identify an emerging condition, understand its significance, investigate the likely cause and determine an appropriate response.

Operators must retain oversight, apply their security and governance requirements, and build confidence in recommendations before extending automation to actions with greater operational impact. Human expertise remains particularly important when the evidence is ambiguous or a decision could affect customers.

Telco networks are increasingly capable of reporting what has gone wrong. The opportunity now is to get better at recognizing what may be going wrong, understanding why and deciding what to do while there’s still time to make a difference.

What to do next

Read more about NTT DATA’s partnership with NVIDIA to see how we can help you accelerate enterprise transformation with AI infrastructure.