As the telecommunications and data centre sectors race to support exploding AI workloads, network operators are being forced to fundamentally rethink both their physical infrastructure and the software layers used to manage increasingly complex operational environments. At APNIC 62, during Technical Session 2 focusing on AI in networks, industry experts and engineering leaders gathered to explore how recent advances in machine learning, agentic artificial intelligence, data centre design, and operational automation are actively reshaping network engineering from the ground up.
The discussions covered a diverse spectrum of critical topics, ranging from next-generation optical data centre fabrics and autonomous network management to predictive fault detection and AI-assisted IP geofeed authoring. Together, these presentations highlighted both the massive opportunities and the profound operational challenges associated with building, scaling, and operating resilient networks in the modern artificial intelligence era.
From Fat Tree to Jellyfish: Rewiring India’s Networks for the AI Era
Opening the technical session, Lalit Singh Chowdhary, Chief Technology and Innovation Officer at Lightstorm Telecom Connectivity Pvt Ltd, examined the often-overlooked physical and architectural infrastructure that underpins modern data centre design. Chowdhary pointed out the staggering growth of hardware deployment over recent years, noting that while approximately 200 million Ethernet ports were sold for internal switching fabrics back in 2019, that figure has skyrocketed closer to one billion today. Crucially, 40 percent of these active ports now operate at speeds of 400Gbps or faster, representing a scale where there are now more active network ports being deployed globally than human babies being born.
Chowdhary traced the lineage of standard data centre topologies back to 1985, referencing the concept of the "fat tree" originally pioneered by Charles Leiserson at the Massachusetts Institute of Technology. In a traditional fat tree layout—which underpins the familiar spine-and-leaf architecture—available bandwidth systematically increases as traffic moves up the switching hierarchy to prevent bottlenecks. At the core layers of these networks, oversubscription is strictly unacceptable, requiring a rigorous 1:1 capacity ratio, whereas minor oversubscription can occasionally be tolerated at the edge.
According to Chowdhary, traditional data centre architecture has historically relied on three critical variables: the number of equal-cost paths available, the precise time required to traverse those paths, and the tolerable level of oversubscription while maintaining an availability standard of five nines. These historical networks were intentionally designed to handle standard front-end web workloads, reflecting the capacity limits and metro-area latency parameters of their respective eras.
To meet the demands of modern cloud environments, Lightstorm currently utilizes a robust three-spine, fully meshed interconnect architecture where physical network growth primarily occurs at the leaf layer. Under this design, any connected data centre can reach another within a maximum of two hops. Interconnects are configured as dedicated wavelengths, and traffic is intelligently distributed across the three-spine layout to achieve five nines availability—representing roughly 26 seconds of allowable downtime per month. If a specific path fails, traffic can be rapidly redistributed without necessitating large-scale, disruptive workload migrations between disparate data centres.
However, escalating infrastructure costs continue to present a persistent challenge for operators, who are compelled to transition legacy 10Gbps and 40Gbps line cards to high-capacity 100Gbps and 400Gbps platforms. This hardware shift can often prove difficult for enterprise customers accustomed to older rack and interconnect cost models. Concurrently, growing application complexity, demanding sharding requirements, and intricate fan-out traffic patterns make accurate workload placement decisions increasingly difficult.
To overcome these physical bottlenecks, Chowdhary introduced an alternative networking model known as the "jellyfish" architecture. Originally developed by researchers at the University of Illinois Urbana-Champaign and HP Labs, the jellyfish design entirely discards structured physical tiers in favor of a conceptually random topology that lacks a visible, centralized core. This innovative approach relies heavily on colourless, directionless, contentionless reconfigurable optical add-drop multiplexer—or CDC-ROADM—technology, significantly reducing path costs through dynamic wavelength switching.
Because the underlying wavelength-switching plane is entirely transparent, any wavelength is capable of travelling in any direction without suffering from resource contention. When a required wavelength is already actively engaged, the system can dynamically and seamlessly recolour the traffic on the fly. This architectural transition is being driven by industry expectations of a tenfold increase in overall capacity demand, with future growth expected to be increasingly constrained by physical power availability rather than raw silicon limits. Consequently, workloads will need to migrate dynamically to geographic regions where power is abundantly available, with network capacity forced to follow.
AI in Network Operations and Autonomous Management
Shifting the focus from physical optics to operational software, Akhilesh Thakur of HPE Networking outlined a comprehensive vision for highly autonomous networks designed to require minimal manual human intervention, thereby reducing the heavy operational burden traditionally associated with reactive network troubleshooting. Thakur noted that historical attempts to apply artificial intelligence to network management frequently faltered due to a fundamental scarcity of clean, accessible operational data. Today, by contrast, modern network operators possess access to exceptionally rich datasets spanning active network elements, control and data planes, underlying infrastructure platforms, enterprise business systems, and external intelligence feeds.
Common AI use cases within this modern paradigm focus heavily on streamlining administrative overhead and accelerating root-cause analysis. Thakur paid particular attention to the emergence of agentic AI frameworks, which possess the capability to systematically break down complex, multi-layered operational problems into smaller, manageable actions while carefully preserving conversational context through a retrieval-augmented generation knowledge base.
To bridge the gap between artificial intelligence models and real-world network management tools, Thakur highlighted the Model Context Protocol. MCP acts as a standardized, secure interface layer sitting directly between AI agents and operational tooling. Instead of requiring developers or system administrators to code direct, brittle integrations with disparate APIs, proprietary databases, or NetConf implementations, MCP functions as a clean intermediary. This architectural abstraction grants agentic systems a consistent operational view across diverse hardware platforms and vendor ecosystems while substantially mitigating integration complexity.
Demonstrating practical applications, Thakur showcased a Junos MCP implementation. Following Hewlett Packard Enterprise’s acquisition of Juniper Networks, traditional Junos routing and switching platforms can now be deeply integrated into large-scale artificial intelligence and data centre environments. Large language models, such as Anthropic’s Claude, are capable of directly interacting with these complex network environments via Junos MCP. Furthermore, MCP natively supports feature introspection, enabling systems to dynamically identify available hardware and software capabilities on the fly. This capability effectively bridges the gap between high-level natural language requests and fully executable operational workflows.
In one demonstrated use case, an automated network audit was successfully generated through a conversational interface with Claude utilizing MCP, complete with detailed recommendations concerning reporting structures and traffic analysis. Thakur strongly emphasized the operational importance of deploying rigorous negative prompts to effectively suppress potential model hallucinations. A secondary demonstration successfully integrated MCP with external data sources, including Cloudflare Radar and RIPEstat, to assist engineers in the rapid detection and investigation of complex BGP route hijacking incidents.
Despite the immense promise of agentic automation, Thakur acknowledged that fully autonomous network operation remains a long-term future objective, and current enterprise deployments continue to strictly maintain a cautious, human-in-the-loop operational approach. Introducing these advanced technologies introduces fresh operational risks, including potential model hallucinations, credential exposure, an expanded operational blast radius, and gradual model drift. Effective governance, therefore, requires pristine RAG training data, robust internal auditing practices, and vigilant human oversight.
Live Machine Learning Solutions for Broadband ISP Operations
Addressing the practical realities faced by smaller regional operators, Dr Girish Saraph of Vegayan Systems Pvt. Ltd. discussed the widening gap between exponential network traffic growth and the limited engineering resources available to small and medium-sized enterprises and traditional Network Operations Centres. As consumer expectations for broadband responsiveness continue to rise, operators are under intense pressure to prioritize issues more effectively and identify potential hardware or software failures before they result in disruptive customer outages.
Dr Saraph presented a real-world case study involving a massive live network encompassing 10 million active customer endpoints distributed across optical network terminal and gigabit passive optical network infrastructure. This sprawling infrastructure routinely generated upwards of 10 million raw system alarms, translating to roughly 200 active, critical issues every single day. Historically, the network required an average of four to six hours stretching from initial diagnosis to final remediation—a staggering operational scale that vastly exceeded the capacity of available NOC personnel.
Because missing a major regional outage can trigger severe financial penalties tied to strict service-level agreements, intelligent prioritization is paramount. To solve this, Vegayan’s machine learning operations platform was deployed to act as an automated network expert, ingesting historical telemetry from centralized data lakes and operational support systems. Years of historical outage records and fault logs were utilized to train multiple specialist machine learning models, each dedicated to a distinct problem domain. These specialized models continuously process live telemetry to generate forward-looking predictive insights and granular incident prioritization recommendations.
When deployed within ONT and PON environments, the platform successfully identified subscriber access ports carrying a historical pattern of recurring faults. In the network studied, roughly 7.5 percent of customer access ports fell into this high-risk category, and the machine learning system achieved an impressive 74 percent hit rate when predicting problematic ports, directly improving customer-facing reliability and cutting down on subscriber churn.
Similar predictive techniques were subsequently applied to GPON aggregation nodes, where 5.8 percent of nodes were classified as high-risk, yielding a reported predictive accuracy of 65 percent. This capability afforded operations teams a crucial three- to four-day prediction window, enabling engineers to accurately distinguish between simple physical fibre cuts and complex hardware-related anomalies such as thermal overheating or underlying voltage instability. Overall, the deployment reportedly improved overall service performance by 20 percent.
Authoring IP Geofeeds Using AI Tools and MCP
Shifting to external routing metadata, Sid Mathur of Fastah Inc. delivered a detailed presentation covering modern IP geofeed publication best practices and the innovative use of MCP-enabled artificial intelligence tools to dramatically improve geofeed data quality. Mathur opened his session by illustrating a common frustration experienced by internet users: a scenario where a customer’s IP address is erroneously geolocated to an entirely different economy or region due to outdated or inaccurate geolocation databases. In the vast majority of cases, content and service providers direct these frustrated geolocation complaints straight back to the customer’s originating internet service provider.
To combat this, Internet Service Providers publish accurate geolocation metadata for their allocated IP address blocks using the standardized comma-separated values format defined under RFC 8805. Emerging from decades of industry and regulatory requirements, the format allows network operators to associate specific IP prefixes with an ISO 3166 economy code, a regional identifier, and a specific city name. Crucially, all geographical fields within the format are strictly optional, granting operators the flexibility to publish geolocation data at the economy, regional, or city level depending on their operational preferences.
Furthermore, operators retain the explicit ability to publish a "NONE" descriptor, indicating to the broader internet ecosystem that specific address blocks should not be actively geolocated. This mechanism is particularly vital for core network infrastructure addresses and widespread anycast deployments, where assigning a rigid geographic location to a globally distributed service can introduce significant routing and mapping anomalies. Additional standards, such as RFC 9632, define automated discovery mechanisms within Regional Internet Registry frameworks, allowing geofeed consumers to easily locate and interpret published network data, while emerging proposals like RFC 9877 explore representing identical information via a JSON-based Registration Data Access Protocol model.
Mathur emphasized that geofeed management must be treated as a continuous, dynamic operational process rather than a static, one-time administrative chore. While smaller ISPs may update their metadata infrequently, larger transit networks frequently require continuous, automated updates as their underlying IP allocations evolve. However, publishing geofeeds via RIR-linked URLs places the absolute responsibility for data accuracy squarely on the address holder. While public platforms like GitHub are frequently used to host these CSV files, external geolocation providers exercise no direct control over the underlying content. Because publishing precise geolocation data carries inherent privacy implications, city-level accuracy is generally recommended as a practical ceiling, and data should always be treated as an approximation rather than an absolute physical coordinate.
To streamline this administrative burden, Mathur demonstrated how modern large language models can be integrated directly into geofeed management workflows. When provided with explicit RFC guidance and internal IP address management documentation, an LLM can easily generate fully compliant geofeeds and execute automated audits. Mathur showcased a specialized AI skill he developed specifically for geofeed lifecycle management, which successfully identified common, persistent errors in public geofeeds—such as missing regional identifiers for ambiguously named cities like Frankfurt, the excessive overuse of do-not-geolocate entries, and instances where numeric values were mistakenly inserted in place of standardized ISO 3166 economy codes. These automated checks function similarly to a software linter, ensuring strict compliance before publication.
Ultimately, a common thread united all the presentations delivered during APNIC 62 Technical Session 2: the undeniable, growing industry-wide requirement for enhanced intelligence, operational flexibility, and comprehensive automation across every layer of the network stack. Whether implemented through novel data centre topologies engineered specifically to handle massive AI workloads, autonomous software agents assisting human network engineers, machine learning models predicting physical faults before they occur, or validation tools ensuring the integrity of routing metadata, the overarching goal remains consistent. As networks continue to scale in complexity, these advanced technologies are rapidly transitioning from experimental concepts into indispensable, practical tools for securing long-term network performance, resilience, and operational efficiency.
Leave a Reply