Google Cloud unveils AI framework for telecoms networks

Google Cloud has outlined an Autonomous Network Operations framework for telecommunications providers that combines a network digital twin, graph-based machine learning and AI agents.

The approach is intended to help operators manage increasingly complex networks that generate vast amounts of operational data and are becoming harder to run through manual processes and conventional machine learning models alone.

The framework has three main layers: a digital twin held in Spanner Graph, a machine learning layer built on Distributed Graph Flow, and an AI layer designed to interpret outputs and take action. It is intended to support network operations tasks including anomaly detection, root cause analysis, capacity planning, traffic forecasting and scenario testing.

At the centre of the design is a digital representation of a live telecoms network. Google Cloud describes this as a temporal graph that reflects changes in the network over time, allowing operators to analyse both current and historical conditions.

The structure is intended to map how routers, interfaces, VPNs and traffic flows relate to one another. It also records links between physical equipment, control-plane relationships, service membership and traffic anchoring, giving machine learning models and AI agents a shared operational picture of the network.

Graph approach

Google Cloud is positioning graph neural networks as a way to process telecoms data that is relational, time-based and spread across very large infrastructures. It argues that standard machine learning techniques can identify patterns in structured data but are less effective at reasoning across the wider system in a way that mirrors human operational decision-making.

By pairing graph neural networks with AI agents, the framework aims to bridge that gap. The graph models are meant to produce predictions and narrow down possible causes of network issues, while the AI layer is intended to interpret those findings and decide on a response.

The machine learning component relies on Distributed Graph Flow, an open-source Python library developed by Google CoreML and Google Research. According to Google Cloud, the software is intended to manage the full lifecycle of graph neural network modelling, from data access to training, evaluation and prediction.

The library can pull data directly from Spanner Graph and can be used through both lower-level tools and a simplified programming interface. In Google Cloud's example, engineers can train a node prediction model, evaluate it, generate predictions and save the model for later use.

Operational uses

Google Cloud identified several telecoms uses for the framework. In anomaly detection, graph models generate representations of nodes and edges based on historic behaviour and current health, with unusual patterns flagged for review before they affect service quality.

For root cause analysis, the system can isolate a smaller section of the network linked to a fault rather than examine the full network. Google Cloud used the example of attach failures in a specific ZIP code, where a troubleshooting agent would analyse only the relevant subgraph.

Predictive maintenance is another target area. The framework can estimate the likelihood of equipment failures or connection breaks, including handover failures involving fast-moving equipment, which could allow operators to rebalance loads or reroute traffic earlier.

The system is also intended to support what-if analysis. Engineers can model events such as fibre cuts, traffic surges or configuration changes and examine how those local disruptions might spread across the wider network topology.

Autonomy target

Google Cloud linked the framework to Level 5 autonomy as defined by TM Forum, at which point network operations would run autonomously. It said the digital twin is meant to provide a trusted environment for testing changes before they are applied to the live network.

Google Cloud also outlined a path for using trained models in production by exporting them to Gemini Enterprise Agent Platform for inference. In that setup, an anomaly event would be sent to an endpoint, which would return the network entities most likely to be behind the issue.

The broader pitch reflects a shift in the telecoms sector as operators look for ways to automate more of their network management. Rising network scale, more heterogeneous infrastructure and the need to act on real-time operational signals have pushed vendors and carriers towards graph-based data models and AI-driven workflows.

Google Cloud said graph neural networks and agentic AI can be combined with human oversight, including human-on-the-loop or human-in-the-loop models, when automated remedial action is taken.

It added that the digital twin can also be used to revisit earlier states of the network so models can be trained and tested on historical conditions, while still giving AI systems a live operational view for current decision-making.

Overall, the framework is intended to help operators move from reactive maintenance to more proactive network operations, with Distributed Graph Flow providing what Google Cloud called the essential primitives for scalable graph neural network training and inference.