Modernizing Network Infrastructure for the Age of Artificial Intelligence

This article discusses the critical need for modernizing network infrastructure to support the demands of Artificial Intelligence (AI) workloads. It highlights how legacy networks are inadequate for AI, outlines the key requirements of AI-ready networks, explores enabling technologies like SDN and 5G, and addresses the challenges and benefits of network modernization for AI.

Artificial Intelligence (AI) is rapidly transforming industries, driving innovation, and creating new efficiencies across diverse business functions. However, the successful adoption and deployment of AI technologies are intrinsically linked to the capabilities of the underlying network infrastructure. This report examines the critical requirements for modernizing enterprise networks to support the unique demands of AI workloads. It establishes that legacy network architectures, designed for traditional IT traffic, represent a significant bottleneck, hindering AI performance and potentially impeding competitive advantage. A truly AI-ready network requires more than increased bandwidth; it demands a holistic approach encompassing enhanced throughput, ultra-low and predictable latency, robust and adaptive security frameworks, exceptional scalability and flexibility, intelligent management potentially powered by AI itself, and seamless integration of cloud and edge computing environments. Key enabling technologies such as Software-Defined Networking (SDN), Network Functions Virtualization (NFV), 5G wireless, and advanced routing protocols are instrumental in achieving this modernization. While significant challenges exist, including upfront costs, integration complexity, the need for specialized skills, data governance concerns, and the risk of vendor lock-in, the benefits—improved AI performance, accelerated innovation cycles, enhanced security posture, greater operational agility, reduced costs, and a crucial competitive edge—make network modernization a strategic imperative. Organizations must proactively align their network infrastructure evolution with their AI ambitions through careful assessment, strategic planning, and phased implementation to unlock the full transformative potential of AI in an increasingly intelligent world.

1. Introduction: The Strategic Imperative of AI-Ready Networks

1.1 The Unstoppable Convergence: AI and Network Infrastructure

Artificial Intelligence (AI), defined as a set of technologies enabling computers to simulate human intelligence functions like reasoning, learning, language understanding, data analysis, and decision-making 1, has transitioned from a futuristic concept to a present-day operational reality. Its integration into core business processes is reshaping industries globally.3 AI applications are diverse, ranging from automating repetitive tasks 1, enhancing customer experiences through personalization and prediction 4, improving decision-making via advanced analytics 2, and driving the development of innovative products and services.9

This escalating reliance on AI introduces a unique and demanding set of requirements for the network infrastructure that underpins these applications. AI workloads, encompassing tasks from model training on massive datasets to real-time inference for immediate action, often differ significantly from traditional IT traffic patterns.10 They frequently involve the transfer and processing of enormous data volumes, necessitate near-instantaneous responsiveness, and demand stringent security measures to protect sensitive data and valuable AI models.10

The sheer pervasiveness of AI adoption underscores its strategic significance. A remarkable 85% of organizations globally report using generative AI (GenAI) in at least one department, with forecasts suggesting near-universal (99%) adoption across various use cases by 2027.3 This widespread integration elevates network modernization beyond a mere technical upgrade; it becomes a fundamental strategic imperative.3 A significant majority (75%) of business leaders personally view AI as critical to their organization's success 6, directly linking network capabilities to the ability to compete and thrive in the evolving digital landscape.

The profound impact of AI extends beyond operational enhancements; it acts as a catalyst for fundamental business transformation. AI is not merely another application to be supported by the network; it is increasingly recognized as a core driver of value creation, enabling new business models, enhancing market intelligence, and redefining competitive dynamics.3 This strategic role implies that network infrastructure must evolve not just to accommodate AI traffic, but to actively enable the strategic goals powered by AI.

Furthermore, a fascinating feedback loop is emerging: as AI applications drive the need for more complex, high-performance networks, AI itself is becoming essential for managing these sophisticated environments. AI for Networking, often termed AIOps for networking, leverages machine learning (ML) and other AI techniques to automate network management tasks, optimize traffic flow, predict potential failures, enhance security, and improve resource allocation 7-5-.19 This creates a symbiotic relationship where advanced network infrastructure is necessary to support demanding AI workloads, and AI-driven automation becomes indispensable for operating these advanced networks effectively and efficiently. This suggests a future trajectory towards increasingly autonomous, self-optimizing network operations.

1.2 Why Legacy Networks Hinder AI Ambitions

Legacy network infrastructures, architected for the predictable, often centralized traffic patterns of traditional enterprise applications, are fundamentally ill-suited to meet the rigorous demands of modern AI workloads.20 They frequently act as critical bottlenecks, constraining the performance and potential of AI initiatives.13

The limitations manifest across several key dimensions:

• Insufficient Bandwidth and Throughput: Legacy networks often lack the capacity to handle the massive dataset transfers required for training large AI models, leading to prolonged training times.10

• High Latency: Unpredictable and often high latency in older networks impedes the performance of real-time AI applications, such as autonomous systems or instant fraud detection, where millisecond delays can have critical consequences.10

• Inadequate Security: Outdated security protocols and architectures may not provide sufficient protection for valuable AI models (intellectual property) and the sensitive datasets used for training and inference.11

• Lack of Scalability and Flexibility: Legacy networks are often rigid and difficult to scale, unable to cope with the dynamic and fluctuating resource demands of AI training cycles and inference traffic.10

Specific bottlenecks can arise from legacy storage systems ill-equipped for AI's I/O patterns (e.g., centralized metadata servers causing congestion, kernel overhead introducing delays 22) and network designs based on outdated assumptions of slow network speeds, which fail to leverage modern high-speed interconnects effectively.13

Reliance on such inadequate infrastructure directly hinders AI innovation, slows down the deployment of AI applications, increases operational costs, and places organizations at a significant competitive disadvantage.20 The Cisco AI Readiness Index starkly illustrates this preparedness gap. Despite nearly universal recognition of AI's urgency (98% report increased urgency), infrastructure readiness remains alarmingly low and has even declined. Only 13% of global organizations were deemed fully AI-ready in 2024, down from 14% the previous year.24 Critically, only 21% possess the necessary GPU compute power, and a mere 30% have adequate data protection capabilities for AI models.24 Furthermore, a significant 79% of companies report experiencing network latency issues when managing AI workloads.27

These network limitations impose a tangible "performance tax" on substantial AI investments. Organizations investing heavily in expensive AI compute resources like GPUs may find their utilization severely hampered if the network cannot supply data efficiently.22 Network bottlenecks directly translate into wasted compute cycles and prolonged Job Completion Times (JCT) for AI tasks, diminishing the return on investment for AI initiatives.12

Addressing this requires recognizing that the AI readiness gap extends beyond technology. The Cisco AI Readiness Index reveals low readiness scores not only in infrastructure but also in data management (80% report data inconsistencies), talent availability (24% lack necessary skills globally), and even organizational culture (board receptiveness to AI declined).24 This indicates that network modernization is not merely a hardware refresh but part of a complex, systemic challenge. Successful transformation demands a holistic strategy encompassing infrastructure upgrades, robust data governance, workforce upskilling, and supportive organizational change management 20-.35

1.3 Foundational Concepts for AI Networking

Understanding the modernization journey requires clarity on key technological concepts:

• Artificial Intelligence (AI): A broad field encompassing technologies that enable computers to simulate human cognitive functions like learning, problem-solving, perception, and decision-making, often by identifying patterns in vast datasets.1 Key subsets include Machine Learning (ML) and Deep Learning (DL).1 In the networking context, AI Networking refers to integrating AI techniques to enhance network operations through intelligent decision-making, automation, and adaptation.8

• Network Infrastructure: The composite system of hardware (e.g., servers, switches, routers, cables, access points), software (e.g., operating systems, protocols), and connectivity mediums (e.g., fiber optics, wireless spectrum) that enables data communication and resource sharing between devices.8 Modern infrastructure increasingly incorporates virtualized components and cloud services alongside physical elements.37

• Bandwidth: The maximum theoretical data transfer capacity of a network connection, typically measured in bits per second (e.g., Mbps, Gbps, Tbps).10 High bandwidth is crucial for transferring the large datasets common in AI.10

• Throughput: The actual rate at which data is successfully transferred over a network, accounting for factors like overhead, latency, and packet loss. It represents the effective data transfer speed achieved in practice.39

• Latency: The time delay experienced for data to travel from its source to its destination, usually measured in milliseconds (ms).10 Low latency is critical for real-time AI applications requiring immediate responses.10 Total latency includes both network transmission time and compute processing time at the endpoints.41

• Cloud Computing: A model for delivering IT resources—including compute power, storage, databases, networking, and software—on demand over the internet, typically with pay-as-you-go pricing.39 It offers significant scalability and flexibility but can introduce latency due to the physical distance between users and centralized data centers.39

• Edge Computing: An architectural approach that involves processing data closer to its point of generation (e.g., on devices, sensors, local servers) rather than exclusively in a centralized cloud.39 Its primary goals are to reduce latency, minimize bandwidth consumption, and enable real-time analysis and decision-making.39

• 5G: The fifth generation of cellular network technology, designed to deliver significantly higher data speeds (multi-Gbps), ultra-low latency (potentially sub-10ms or even 1ms), massive device connectivity, and enhanced reliability compared to 4G/LTE 48-.49 It is a key enabler for demanding applications like edge AI, autonomous systems, and the massive Internet of Things (IoT).51

• Software-Defined Networking (SDN): An architecture that decouples the network's control plane (which makes decisions about where traffic is sent) from the data plane (which forwards the traffic). This allows network behavior to be controlled and programmed centrally via software (the SDN controller), often using protocols like OpenFlow.38 SDN enhances network agility, automation, visibility, and traffic management capabilities.55

• Network Functions Virtualization (NFV): An architectural concept that decouples network functions (e.g., firewalls, load balancers, routers, 5G core elements) from dedicated hardware appliances. These functions run as software, known as Virtual Network Functions (VNFs) or Cloud-Native Network Functions (CNFs), on standard IT infrastructure (servers, storage, switches).37 NFV increases agility, reduces capital and operational expenditures (CapEx/OpEx), and simplifies the deployment and scaling of network services.37 Management and Orchestration (MANO) frameworks, standardized by ETSI, oversee the NFV environment 37-.74

Collectively, these foundational concepts signal a fundamental architectural evolution in networking. The trend is decisively away from static, hardware-centric, and predominantly centralized models towards architectures that are software-defined, virtualized, distributed, and increasingly intelligent. This evolution is not merely incremental; it is a necessary transformation to build networks capable of supporting the diverse, dynamic, and demanding nature of AI workloads and the future applications they will enable.

2. Decoding the Network Demands of Diverse AI Workloads

2.1 Beyond Bandwidth: Critical Performance Metrics

The successful deployment and operation of AI applications are contingent upon the network infrastructure's ability to meet a confluence of critical performance requirements simultaneously.11 While high bandwidth is often the most cited need, it represents only one facet of the complex demands AI places on the network. A holistic view reveals that achieving optimal AI performance necessitates a delicate balance across several key metrics:

• High Bandwidth and Throughput: Essential for moving the massive datasets involved in AI model training and handling high-volume data streams for inference.10

• Low Latency: Critical for applications requiring real-time or near-real-time responsiveness, where delays can compromise functionality, safety, or user experience.10

• Robust Security: Necessary to protect valuable AI models, sensitive training and inference data, and the integrity of AI-driven decisions from cyber threats.11

• High Reliability and Availability: Fundamental for ensuring the continuous operation of AI systems, especially those integrated into mission-critical processes where downtime is unacceptable.11

Crucially, the specific profile of these requirements varies significantly depending on the type of AI workload being supported.12 For instance, the network demands for training a large language model (LLM) differ substantially from those for deploying a real-time computer vision system for autonomous navigation. This diversity necessitates network architectures that are not only high-performing but also flexible and adaptable, capable of catering to the unique needs of different AI applications, often requiring specialized configurations or network segments.13

2.2 High Bandwidth & Throughput: Fueling Data-Intensive AI

One of the most prominent network requirements driven by AI is the need for substantial bandwidth and high throughput.10 This demand stems primarily from the data-intensive nature of many AI processes, particularly the training phase of machine learning (ML) and deep learning (DL) models.10 Training these models often involves processing colossal datasets that can range from terabytes to petabytes in size. The efficient and rapid transfer of this data between storage systems, compute nodes (especially Graphics Processing Units - GPUs), and memory is paramount.10 Insufficient bandwidth acts as a major bottleneck, significantly prolonging training times, delaying the development cycle, and hindering the pace of AI innovation.10

Beyond training, certain real-time inference applications also impose significant bandwidth demands. Use cases like high-resolution video analytics, complex scientific simulations driven by AI, or processing continuous streams of sensor data require the network to sustain high throughput for ingesting input data and delivering inference results without delay.10

Modernized networks, equipped with high-bandwidth capabilities, demonstrably accelerate AI initiatives. While specific quantitative studies vary, the principle is clear: alleviating network bottlenecks speeds up AI processes. Infrastructure limitations are a known cause of delays in AI deployment and training.22 Conversely, optimized infrastructure featuring high bandwidth and low latency demonstrably reduces the Job Completion Time (JCT)—a key metric for AI tasks like model training or inference operations.12 The documented infrastructure readiness gaps highlighted by the Cisco AI Readiness Index further imply that overcoming these limitations through modernization will lead to faster deployment cycles.24

It is critical to recognize that these bandwidth requirements are not static; they are continuously escalating. This upward trend is fueled by the relentless progress in AI research, leading to increasingly complex models with billions or even trillions of parameters and the utilization of ever-larger datasets to achieve higher accuracy.10 Consequently, network planning for AI cannot merely address current needs but must anticipate this future growth, incorporating scalability as a fundamental design principle.10

The immense bandwidth demands associated with AI training, particularly the intense inter-GPU communication required for distributed training across large clusters, are driving a significant shift in data center network design. This involves the adoption of specialized, high-performance network fabrics often referred to as "back-end" networks, distinct from the traditional "front-end" networks used for general data center traffic.12 These AI fabrics prioritize lossless data transmission, ultra-low latency, and massive throughput, often employing technologies like high-speed Ethernet (400G, 800G, and emerging 1.6T) enhanced with protocols like RDMA over Converged Ethernet (RoCE), or alternative interconnects like InfiniBand.13 This trend towards specialized AI networking domains adds architectural complexity but is deemed necessary to optimize the performance and utilization of costly GPU resources dedicated to AI workloads.

2.3 Ultra-Low Latency: Enabling Real-Time AI

While bandwidth addresses data volume, latency addresses the speed of response, a critical factor for a growing class of AI applications that operate in real-time.10 Network latency, the delay in data transmission, must be minimized for AI systems that need to perceive, decide, and act within fractions of a second.

Use cases demanding ultra-low latency are numerous and span critical sectors:

• Autonomous Systems: Self-driving cars, drones, and industrial robots require immediate processing of sensor data (LiDAR, cameras, radar) and rapid control responses for safe and effective operation.10

• Remote Surgery and Telemedicine: Enabling surgeons to operate remotely with haptic feedback or facilitating real-time remote diagnostics necessitates minimal delay to ensure precision and patient safety.42

• Real-time Financial Trading: Algorithmic trading systems rely on sub-millisecond latency to execute trades based on rapidly changing market data.10

• Industrial Automation and Control: Real-time control loops in smart factories require extremely low latency for precise coordination and safety.42

• Augmented/Virtual Reality (AR/VR) and Tactile Internet: Immersive experiences and applications involving real-time haptic feedback demand minimal lag to maintain realism and user comfort.42

For these applications, latency requirements are often specified in the single-digit milliseconds (sub-10ms) or even sub-millisecond range.42 Even minor delays can have severe consequences, ranging from poor user experience in interactive applications 97 to critical safety failures in autonomous systems.11 Human reaction time benchmarks (e.g., under 100ms) are sometimes used as a reference point for the responsiveness needed in autonomous safety systems.93 Key performance indicators for latency-sensitive AI inference include Time-to-First-Token (TTFT) and Output Tokens Per Second (OTPS).22

Achieving such stringent low-latency targets necessitates fundamental shifts in network architecture and technology. Placing computation closer to the data source through Edge Computing is a primary strategy to minimize the physical distance data must travel, thereby reducing transmission delays.10 This approach is often essential for meeting sub-10ms goals.44 Advanced wireless technologies like 5G and future 6G provide the underlying high-speed, low-latency radio access network required to connect edge devices and enable mobile real-time AI 48-.42 Consequently, distributed network architectures, where compute, storage, and network resources are strategically placed closer to end-users or devices generating data, become paramount 31-.90

Furthermore, low latency is not solely about achieving the minimum possible delay; it also encompasses predictability and minimizing jitter (the variation in latency).28 For many real-time control systems or interactive AI experiences, inconsistent latency can be as detrimental as high average latency.83 This underscores the need for network designs that prioritize consistent and reliable delay characteristics, often leveraging techniques like Quality of Service (QoS) prioritization or dedicated network slices (e.g., 5G URLLC) to guarantee performance for critical AI traffic.8

The fundamental physics of data transmission dictates that achieving ultra-low latency (sub-10ms or sub-1ms) is often impossible if data must traverse long distances to a centralized cloud data center and back.22 This physical constraint fundamentally challenges the traditional cloud-centric computing model for a significant and growing class of AI applications. It mandates the adoption of edge computing and distributed architectures where processing occurs much closer to the data source.39 Therefore, network modernization for low-latency AI extends beyond simply upgrading network links; it requires a strategic embrace of distributed computing paradigms (edge, Multi-access Edge Computing - MEC) and the enabling connectivity technologies (like 5G) that support them, leading inevitably towards hybrid edge-cloud infrastructure models.101

2.4 Robust Security: Protecting AI Data and Models

Security is a non-negotiable cornerstone of any network modernization strategy, particularly when supporting AI workloads.11 The reasons are twofold: the inherent sensitivity of the data often used to train AI models (which can include proprietary business information, personal data, or regulated data like health records) and the immense value of the trained AI models themselves, which represent significant intellectual property and competitive assets.11

AI systems also introduce novel security challenges and expand the traditional attack surface.81 Key AI-specific threats include:

• Adversarial Attacks: Carefully crafted, often subtle modifications to input data (images, text, audio) designed to deceive an AI model and cause it to make incorrect predictions or classifications, potentially leading to system failures or manipulation.82

• Data Poisoning: The malicious injection of corrupted or biased data into the training dataset of an AI model. This can compromise the model's integrity, skew its outputs, introduce vulnerabilities, or embed hidden biases.81

• Model Theft or Extraction: Unauthorized attempts to steal, copy, or reverse-engineer proprietary AI models, often by exploiting APIs or other access points. This compromises intellectual property and competitive advantage.81

• Prompt Injection: A threat specific to large language models (LLMs), where malicious instructions are embedded within user prompts to bypass safety controls, extract sensitive information, or generate harmful content.113

Consequently, a modernized network supporting AI must implement a comprehensive, multi-layered security framework, often referred to as "Defense-in-Depth".118 This includes robust network security controls like firewalls, Intrusion Detection and Prevention Systems (IDS/IPS), and network segmentation (including micro-segmentation facilitated by SDN/NFV).34 It also requires strong data security measures such as advanced encryption for data both in transit and at rest, and Data Loss Prevention (DLP) tools.11 Furthermore, stringent Identity and Access Management (IAM) policies, incorporating Multi-Factor Authentication (MFA) and Privileged Access Management (PAM), are critical to control access to AI systems and data.34 Adherence to Zero Trust principles, which assume no implicit trust regardless of location, is increasingly relevant.13 Specific defenses against AI threats, such as adversarial training, input validation, AI model watermarking, and secure execution environments (enclaves), are also necessary.81

Given the sophistication and novelty of AI-related threats, traditional static security measures are often insufficient. Security frameworks must become dynamic and adaptive, capable of identifying and responding to previously unseen attacks and the unique vulnerabilities inherent in AI systems.5 This is where AI itself plays a crucial role in bolstering security. AI-powered cybersecurity tools leverage machine learning to analyze complex network traffic patterns, detect subtle anomalies indicative of threats (including adversarial inputs or data poisoning attempts), predict potential attacks, and automate incident response actions.5 Technologies like adaptive MFA, which dynamically adjust authentication requirements based on real-time risk assessment, exemplify this trend.127 These AI-driven security capabilities are becoming indispensable for adequately protecting AI assets and the networks they run on.5

It is important to distinguish the dual nature of AI security in this context. While AI can be weaponized by attackers to create more sophisticated threats 113, a primary objective of network modernization for AI is the protection of the AI assets—the models and data—residing on the network.11 This necessitates defenses specifically tailored to counter threats like model theft, data poisoning, and adversarial manipulation, which target the AI components directly and go beyond conventional network perimeter security.

The increasingly distributed nature of modern AI deployments, spanning edge devices, private data centers, and multiple public clouds, significantly complicates the security challenge 11-101-.43 This distribution inherently expands the potential attack surface 81 and makes the consistent enforcement of security policies, unified monitoring, and coordinated threat response across diverse environments considerably more difficult than in a centralized model. Ensuring robust security and data privacy across these interconnected systems demands sophisticated, integrated security architectures and advanced management tools capable of providing visibility and control across the entire hybrid landscape.108

2.5 Unwavering Reliability: Ensuring Continuous AI Operations

As AI applications become more deeply embedded within core business operations and mission-critical systems, the reliability and availability of the supporting network infrastructure become paramount.83 Network downtime or performance degradation is no longer just an inconvenience; it can lead to significant operational disruptions, substantial financial losses, and, in the case of applications like autonomous systems or remote healthcare, severe safety risks.12

Consequently, high reliability and continuous availability are fundamental requirements for any network infrastructure intended to support serious AI deployments.11 This entails ensuring uninterrupted connectivity and minimizing the frequency and duration of service interruptions.12 Network design must therefore incorporate robust redundancy mechanisms, such as multiple network paths, backup power systems, and redundant hardware components (routers, switches), along with fault tolerance capabilities that allow the network to gracefully handle component failures.40

For certain types of AI workloads, particularly large-scale distributed training, the concept of a "lossless" network is critical.12 In these scenarios, dropped packets are not merely delayed but can corrupt complex calculations distributed across multiple GPUs or necessitate costly retransmissions, significantly impacting Job Completion Time (JCT) and overall efficiency. Technologies like Ethernet enhanced with RoCE or specialized interconnects like InfiniBand are often employed specifically because they are designed to minimize or eliminate packet loss under heavy load, ensuring the integrity of these computationally intensive tasks.28

The reliability requirements for AI thus extend beyond simple network uptime. They encompass the need for data integrity and consistency, particularly in distributed AI systems where computations rely on the accurate and timely exchange of information between nodes.28 This places a higher premium on the quality and performance of the network fabric itself, demanding not just connectivity but dependable, error-free data delivery.

As AI transitions from experimental phases to controlling critical infrastructure (e.g., power grids, transportation systems 42) or driving essential business processes (e.g., financial forecasting, supply chain optimization 83), the consequences of network failure become far more severe than traditional IT outages. This reality elevates network reliability from a purely operational concern to a core element of business continuity planning and risk management. It compels organizations to view network modernization investments through the lens of mitigating significant business risks, potentially justifying higher expenditures on redundancy, advanced security measures, and proactive network management—often leveraging AI itself for predictive maintenance and anomaly detection—than might have been considered previously 5-.19

2.6 Network Requirements Profiles for Key AI Applications

The network demands of AI are not monolithic; they vary considerably based on the specific application category. Understanding these diverse profiles is crucial for designing an appropriately modernized network.

• Machine Learning (ML) Training:

• Bandwidth/Throughput: Very High. Requires transfer of massive datasets (TBs/PBs) and model parameters between storage and distributed GPU clusters.10 Essential for minimizing training time (JCT).12 Often necessitates high-speed interconnects (e.g., 400G/800G Ethernet with RoCE, InfiniBand).13

• Latency Sensitivity: Moderate to High. While not typically real-time in the user-facing sense, low latency communication between GPUs during parallel processing is critical for efficient synchronization and minimizing compute idle time.12 Lossless transmission characteristics are often required to maintain computational integrity.28

• Security Concerns: Strong protection needed for large, potentially sensitive training datasets.11 The trained model itself represents valuable intellectual property requiring safeguarding.11 Defenses against data poisoning attacks during the training phase are crucial 82-.116

• Reliability: High. Training jobs can be lengthy and consume significant, expensive compute resources; interruptions due to network issues cause substantial delays and increase costs.12 A reliable, often lossless, network fabric is critical for the integrity of distributed training.28

• Real-time Inference:

• Bandwidth/Throughput: Moderate to High. Dependent on the size and rate of input data (e.g., high-resolution video vs. simple text queries) and the complexity of the AI model being used.10 Must support rapid ingestion of input data and timely delivery of inference results.12

• Latency Sensitivity: Very High. This is often the defining requirement. Many applications demand near-instantaneous responses for effective operation or user experience.10 Ultra-low latency (sub-10ms or even sub-1ms) is frequently necessary.42 Key metrics include Time-to-First-Token (TTFT) and Output Tokens Per Second (OTPS) for generative models.22

• Security Concerns: Ensuring the secure and low-latency transmission of potentially sensitive input data.87 Protecting the inference process and model integrity from manipulation (e.g., adversarial attacks designed to cause misclassification) 11-.116 Securing the API endpoints through which inference services are accessed.115

• Reliability: Very High. Inference often powers critical real-time decision-making or user interactions; failures can have immediate negative consequences.78 Consistent low latency (low jitter) is often as important as the average latency value.42

• Computer Vision:

• Bandwidth/Throughput: High to Very High. Processing and transmitting high-resolution images and video streams inherently requires significant network capacity.10

• Latency Sensitivity: Moderate to Very High. Real-time applications such as object detection for autonomous vehicles, live security video analysis, or interactive AR/VR demand very low latency.10 Batch processing of images or video archives may tolerate higher latency.

• Security Concerns: Secure transmission, processing, and storage of image and video data, which can often be sensitive (e.g., surveillance footage, medical scans, facial recognition data).11 Protecting models against adversarial attacks that manipulate visual inputs to cause misidentification.82

• Reliability: High to Very High. Essential for applications where visual analysis directly informs critical real-time actions (e.g., navigation, threat detection) or diagnostic decisions.

• Natural Language Processing (NLP):

• Bandwidth/Throughput: Moderate to High. Requirements depend heavily on the specific task and model size. Processing simple text queries might be moderate, but training large language models (LLMs) requires very high bandwidth.84 Real-time voice processing or multimodal applications integrating NLP with other data types also increase bandwidth needs.10

• Latency Sensitivity: Moderate to High. Interactive applications like real-time translation, voice assistants, or chatbots require low latency to provide a responsive and natural user experience.10 Batch processing tasks like document summarization or sentiment analysis are less latency-sensitive.

• Security Concerns: Protection of potentially sensitive text or voice data being processed (e.g., personal conversations, confidential business documents, customer service interactions).81 Preventing prompt injection attacks designed to manipulate LLM outputs.113 Ensuring data privacy and compliance during processing.81

• Reliability: Moderate to High. Depends on the criticality of the application. A customer service chatbot failure might be inconvenient, whereas errors in AI analyzing legal or medical text could have serious consequences.

• Autonomous Systems (Vehicles, Robotics, Drones):

• Bandwidth/Throughput: High to Very High. These systems must handle large volumes of continuous sensor data (LiDAR, cameras, radar, GPS, IMU) for perception and localization, as well as transmitting control commands and receiving updates.76

• Latency Sensitivity: Extremely High (Ultra-Low Latency). This is often the most critical requirement. Safety-critical functions like obstacle avoidance, path planning, and control loop adjustments demand near-instantaneous response times, frequently in the sub-10ms or even sub-1ms range.10

• Security Concerns: Paramount importance on ensuring the security and integrity of real-time sensor data and control signals to prevent malicious takeover or manipulation.11 Protection against adversarial attacks targeting perception systems (e.g., tricking object recognition) or control algorithms is vital.82 Preventing unauthorized access, communication interception, or system hijacking is essential.81 Data privacy for location tracking and sensor data must also be addressed.83

• Reliability: Extremely High (Mission-Critical / Safety-Critical). Network failures or performance degradation can have catastrophic consequences, potentially leading to accidents or mission failure.11 Requires highly robust, fault-tolerant, and often redundant communication links.78

This diversity in network requirements across different AI applications underscores a crucial point: a monolithic, one-size-fits-all network infrastructure is unlikely to be efficient or even adequate. The varying demands for bandwidth, latency, security posture, and reliability levels necessitate a network that is inherently flexible, programmable, and potentially segmented. Technologies like SDN, NFV, and network slicing (particularly within 5G frameworks) become essential tools. They allow network administrators to dynamically allocate resources (e.g., provision higher bandwidth for a training job), prioritize traffic based on sensitivity (e.g., guarantee low latency for real-time inference using QoS mechanisms), and enforce specific security policies tailored to the particular AI workload utilizing the network at any given moment.8

3. The Architectural Pillars of AI Network Modernization

Modernizing network infrastructure to effectively support the demanding and diverse requirements of AI workloads requires a strategic focus on several key architectural pillars. These pillars collectively form the foundation of an AI-ready network, moving beyond legacy constraints towards a more capable, agile, and intelligent infrastructure.

3.1 Pillar 1: Enhanced Bandwidth and Throughput

Addressing the voracious appetite of AI applications for data necessitates significant enhancements in network capacity.10 This involves upgrading the physical and wireless infrastructure to support much higher data transfer rates than typically found in legacy environments. Key technologies enabling this pillar include:

• High-Speed Ethernet: The backbone of data center networking is rapidly evolving to accommodate AI. Speeds of 400 Gbps and 800 Gbps are becoming increasingly common for AI fabrics, with 1.6 Tbps and beyond on the horizon to handle the massive data flows between GPUs and storage systems during training and large-scale inference.13

• Fiber Optic Infrastructure: High-capacity fiber optic cabling is essential for both data center interconnects (DCI) and wide-area network (WAN) links supporting distributed AI workloads. Technologies like dense wavelength-division multiplexing (DWDM) and advanced coherent optics (e.g., Ciena's WaveLogic 6 Extreme enabling 1.6 Tbps per wavelength) maximize the data-carrying capacity of fiber infrastructure.10

• Advanced Wireless Technologies: For mobile and edge AI applications, modern wireless standards are critical. Wi-Fi 6 and 6E offer significantly increased throughput and capacity compared to previous generations.146 More significantly, 5G technology provides a transformative leap, delivering multi-gigabit-per-second speeds, alongside low latency, making it a cornerstone for high-bandwidth AI applications deployed outside traditional data centers, particularly at the network edge 51-.49

The synergy between 5G and edge computing is particularly potent for enabling a new class of AI applications. By delivering high bandwidth and low latency connectivity directly to edge locations, 5G allows for the processing of large data volumes generated by sensors, cameras, or devices in near real-time, close to the source. This combination unlocks possibilities for sophisticated, data-intensive AI applications in domains like smart cities, connected vehicles, industrial automation, and immersive experiences, which were previously constrained by network limitations.51

Achieving the necessary bandwidth and throughput for AI requires a holistic view. Enhancements cannot be confined to just the network core or the data center fabric. Bottlenecks can occur anywhere along the data path—from the wireless access network (Wi-Fi, 5G) connecting edge devices, through the edge computing nodes, across the WAN links, and within the high-performance data center interconnects 48-.49 Therefore, a comprehensive modernization strategy must address capacity upgrades across all these segments to ensure end-to-end performance for distributed AI workloads.

The sheer scale of bandwidth demanded by AI, especially for training massive models, is acting as a powerful catalyst for innovation and competition within the networking industry itself. This is evident in the rapid progression of Ethernet speeds towards 800G and 1.6T 88, the ongoing debate and development around specialized interconnects like InfiniBand versus enhanced Ethernet (RoCE) for AI fabrics 28, and advancements in optical transport technologies designed to push more data through existing fiber.144 This technological race presents organizations with both opportunities to leverage cutting-edge performance and challenges related to cost, complexity, and ensuring interoperability as they modernize their networks.

3.2 Pillar 2: Ultra-Low Latency and Real-Time Capabilities

Meeting the stringent responsiveness requirements of many real-time AI applications necessitates a focused effort on minimizing network latency.10 Achieving the ultra-low latency targets (often sub-10ms, sometimes sub-1ms) demanded by applications like autonomous systems, remote surgery, or real-time industrial control requires a multi-pronged architectural and technological approach:

• Edge Computing: This is arguably the most critical strategy. By moving computation and data processing physically closer to the data source or end-user, edge computing drastically reduces the network distance data needs to travel, thereby minimizing transmission delays.10 Multi-access Edge Computing (MEC) integrates cloud capabilities directly into the network edge, often co-located with 5G infrastructure.38 Edge processing is frequently essential for achieving sub-10ms latency goals.44

• Content Delivery Networks (CDNs): Traditionally used for caching static web content, CDNs can potentially be adapted to reduce latency for certain AI inference tasks by caching frequently used models or data closer to users. Their effectiveness is amplified when integrated with edge computing concepts, enabling dynamic updates and potentially sub-10ms responses via edge-triggered mechanisms.46

• Advanced Routing Optimization: Moving beyond simple shortest-path routing, advanced techniques employ more sophisticated algorithms, often incorporating AI/ML, to select network paths dynamically based on real-time conditions like congestion, link quality, and application-specific latency requirements.8 This ensures latency-sensitive AI traffic is prioritized and routed efficiently.

• High-Speed, Low-Latency Fabrics: Within data centers, specialized interconnect technologies like InfiniBand or Ethernet optimized with RDMA/RoCE provide ultra-low latency communication essential for tightly coupled distributed AI tasks, such as parallel model training across multiple GPUs.28

• 5G and Beyond (6G) Networks: These wireless technologies are designed with low latency as a core tenet, providing the essential radio access network (RAN) connectivity for mobile and edge-based real-time AI applications 48-.42

Successfully achieving ultra-low latency often mandates a fundamental shift towards distributed network architectures. Instead of relying solely on centralized cloud data centers, compute, storage, and network functions must be strategically distributed and placed closer to where data is generated and consumed by AI applications 31-.90 Traditional centralized models simply cannot overcome the physical limitations imposed by the speed of light for latency-critical use cases.22

It is also crucial to understand that achieving low latency for AI involves optimizing the entire end-to-end processing pipeline, not just the network transmission component.41 Compute latency—the time taken by the AI model itself to process the input and generate an output—can be a significant factor, especially for complex models. While network latency might be critical for synchronizing distributed training or delivering real-time control signals, optimizing only the network may not yield the desired overall responsiveness if the edge compute resources are underpowered or the AI model itself is computationally intensive. Therefore, a successful low-latency strategy requires careful co-design and optimization of both the network infrastructure and the compute resources (including hardware accelerators like GPUs or NPUs deployed at the edge) along with efficient data transfer protocols like RDMA that minimize CPU overhead.28

The imperative for low latency is driving a significant convergence of networking, computing, and storage capabilities, particularly at the network edge. Edge computing platforms inherently blend these functions.39 Technologies like MEC explicitly integrate cloud compute and storage into the network infrastructure 38, and 5G standards are designed with native support for edge deployments.37 This convergence blurs traditional IT infrastructure silos. Modernizing for low-latency AI is therefore not solely the responsibility of the network team; it necessitates close collaboration between network architects, compute and storage engineers, application developers, and data scientists to design, deploy, and manage these integrated edge platforms effectively.108

3.3 Pillar 3: Robust and Adaptive Security Frameworks

Securing the network infrastructure that supports AI is a critical and complex pillar of modernization.11 The high value of AI assets—including proprietary algorithms, trained models, and often vast quantities of sensitive training data—makes them attractive targets for cyber adversaries.11 Furthermore, AI systems themselves introduce unique vulnerabilities and attack vectors that must be addressed 81-.116

Building a secure AI network requires adopting a multi-layered "Defense-in-Depth" strategy 118, encompassing controls across various domains:

• Network Security: Implementing robust firewalls, Intrusion Detection and Prevention Systems (IDS/IPS), and network segmentation. Technologies like SDN and NFV can enable finer-grained control through micro-segmentation, isolating AI workloads and limiting the potential blast radius of a breach.34

• Data Security: Employing strong encryption for data both at rest (in storage) and in transit across the network. Data Loss Prevention (DLP) tools help monitor and prevent unauthorized exfiltration of sensitive training or inference data.11

• Identity and Access Management (IAM): Implementing strict controls over who can access AI systems, models, and data. This includes strong authentication methods like Multi-Factor Authentication (MFA), potentially adaptive MFA, role-based access control (RBAC), and Privileged Access Management (PAM) to secure administrative accounts.34

• Application and Model Security: Incorporating secure coding practices during AI development, performing adversarial robustness testing to identify and mitigate vulnerabilities to manipulation, potentially using AI model watermarking to detect theft, and running sensitive AI processes in secure enclaves.81

Given the dynamic nature of AI workloads and the evolving landscape of AI-specific threats, security cannot be static. It must be dynamic and adaptive, capable of learning and responding in real-time to new threats and changing network conditions.127 This is where AI-powered cybersecurity tools become essential. These systems use ML algorithms to analyze vast amounts of security telemetry, identify subtle anomalies that might indicate an attack (including sophisticated adversarial attacks or data poisoning attempts), predict potential threats, and automate response actions, significantly reducing detection and containment times.5

A crucial aspect of AI network security is understanding that it encompasses more than just defending the network from external threats, including those potentially augmented by AI.113 A primary focus must be on protecting the AI assets—the models and data—that reside on the network.11 This requires implementing defenses specifically designed to counter AI-centric attacks like model theft, data poisoning, and adversarial manipulation, which target the core components of the AI system itself and necessitate security measures beyond traditional perimeter defenses.

The trend towards distributed AI deployments across hybrid edge-cloud environments further complicates the security posture 11-101-.43 This distribution inherently increases the attack surface 81 and makes it challenging to maintain consistent security policies, achieve unified visibility, and orchestrate effective threat responses across disparate locations and platforms. Securing these complex, interconnected environments demands sophisticated, integrated security architectures, often relying on unified management platforms and Zero Trust principles to enforce consistent security regardless of where AI workloads or data reside.13

This inherent complexity and the unique nature of AI threats create a compelling feedback loop: the need to secure AI infrastructure drives the development and adoption of AI-powered security tools. As organizations deploy more sophisticated AI, they require increasingly advanced AI-based security solutions to protect it. This demand fuels innovation in the field of AI for cybersecurity, as evidenced by investments in startups focused on AI attack simulation and prevention.128 This cycle suggests that AI security will become a critical and rapidly evolving specialization within both the AI and cybersecurity domains, leading towards more autonomous and intelligent security systems designed for and managed by AI.

3.4 Pillar 4: Scalability and Flexibility for Dynamic Needs

AI workloads are characterized by their dynamic and often unpredictable resource requirements. Model training can involve periods of intense, sustained demand for compute, storage, and network bandwidth, while inference workloads might fluctuate based on user activity or real-time events.10 A modernized network infrastructure must therefore possess inherent scalability and flexibility to adapt seamlessly to these varying demands, ensuring efficient resource utilization and the ability to accommodate future growth.10

Several key technologies and architectural approaches enable this crucial pillar:

• Cloud Computing: Public and private cloud platforms offer inherent elasticity, allowing organizations to scale resources (compute instances, storage capacity, network bandwidth) up or down on demand, often with a pay-as-you-go model. This is fundamental for handling the variable needs of AI workloads.10 Hybrid and multi-cloud strategies, combining on-premises infrastructure with services from multiple cloud providers, provide maximum flexibility for optimizing workload placement based on cost, performance, data sovereignty, or access to specialized AI services. However, these distributed environments introduce significant challenges in network management, security consistency, and operational complexity.125

• Virtualization: Decoupling resources from physical hardware through virtualization (for compute, network, and storage) is a core enabler of flexibility and dynamic allocation.10 Network Functions Virtualization (NFV) specifically allows network services to be deployed and scaled as software, providing agility analogous to virtual machines or containers 65-.72

• Software-Defined Networking (SDN): The programmability offered by SDN allows network resources and configurations to be adjusted dynamically via software control, directly responding to the needs of AI applications.38 This enables automated scaling of network capacity or adjustment of QoS policies based on workload demands.

• Modular Architectures: Designing infrastructure using modular building blocks, such as leaf-spine network topologies in data centers 12 or modular hardware platforms 10, facilitates easier and more cost-effective incremental scaling as AI demands grow.

Achieving true scalability for AI requires more than just scaling one infrastructure component in isolation. AI workloads place simultaneous demands on compute (GPUs), storage (I/O performance), and network (bandwidth, latency). Scaling up GPU capacity without commensurate increases in network bandwidth or storage throughput will inevitably lead to performance bottlenecks, where expensive compute resources sit idle waiting for data.10 Therefore, effective scalability for AI necessitates coordinated, elastic scaling across all three domains—compute, storage, and networking. Cloud platforms 158, virtualization technologies 163, and software-defined infrastructure (SDN/NFV) 57 provide the mechanisms, but realizing optimal performance requires intelligent orchestration and automation that manages these resources holistically.158

The inherent complexity of managing scalable and flexible network resources across hybrid and multi-cloud environments poses a significant challenge.139 Each cloud provider has different APIs, management tools, and service capabilities, and integrating these with on-premises and edge infrastructure requires sophisticated solutions. This operational complexity is driving the demand for advanced orchestration and automation platforms. These platforms aim to provide a unified control plane for provisioning resources, enforcing policies, managing security, and automating scaling actions across diverse environments (e.g., NFV MANO for virtualized network services 37-74, cloud management platforms, multi-cloud networking solutions 162). While these orchestration layers are crucial for taming complexity and enabling agility, organizations must carefully evaluate their adoption, as heavy reliance on specific proprietary platforms can introduce risks of vendor lock-in, potentially limiting future flexibility and increasing long-term costs.138 Strategic decisions regarding the use of open standards versus integrated vendor ecosystems become critical in navigating this landscape.

3.5 Pillar 5: AI-Powered Network Management and Automation

Intriguingly, Artificial Intelligence itself is becoming a key pillar in modernizing networks for AI. By applying AI and Machine Learning (ML) techniques to network operations—a field often referred to as AIOps for networking—organizations can significantly enhance the management, optimization, security, and overall efficiency of their increasingly complex network infrastructure 7-5-.19

AI can be leveraged for a variety of network management functions:

• Predictive Analytics: AI algorithms can analyze historical and real-time network telemetry data to forecast future traffic patterns, predict potential congestion points, anticipate hardware failures or performance degradation before they occur, and identify emerging security threats.5 Studies have demonstrated high accuracy in tasks like congestion prediction (e.g., 94% accuracy up to 30 minutes in advance 19) and incident prevention (up to 82% reduction 19).

• Intelligent Traffic Optimization: AI can dynamically analyze network traffic flows and make intelligent decisions to optimize routing paths, balance loads across links, and manage bandwidth allocation to improve performance, reduce latency, and ensure Quality of Service (QoS) for critical applications.8 This can lead to significant improvements in bandwidth utilization (e.g., up to 55% during peak periods 19).

• Automated Resource Allocation: Based on real-time monitoring and predictive insights, AI can automate the allocation and scaling of network resources (bandwidth, virtual functions) as well as related compute and storage resources, ensuring efficient utilization and preventing over- or under-provisioning.5 This contributes to cost savings and improved performance.160

• Automated Troubleshooting and Remediation: AI systems can automatically detect network anomalies and performance issues, diagnose the root cause by correlating events across different domains, and either recommend corrective actions to human operators or automatically implement remediation steps (e.g., rerouting traffic, restarting services, adjusting configurations).5 This significantly reduces Mean Time To Resolution (MTTR) 19, minimizes downtime, and alleviates alert fatigue for IT staff.21

• Enhanced Network Security: AI algorithms excel at identifying subtle patterns and anomalies in network traffic that may indicate sophisticated cyber threats, including zero-day attacks or insider threats. AI can automate threat detection and trigger rapid response actions.5

• Virtual Network Assistants: The integration of Large Language Models (LLMs) and conversational AI creates virtual assistants that allow network operators to interact with management systems using natural language. Operators can ask questions about network status, request troubleshooting assistance, or quickly find relevant documentation, simplifying complex tasks.17

By integrating these AI capabilities into network management platforms and workflows, organizations can achieve more proactive, efficient, automated, and reliable network operations. This reduces the reliance on manual intervention, minimizes human error, lowers operational expenditures (OpEx), and frees up skilled personnel to focus on more strategic initiatives.5

The successful application of AI for network management fundamentally shifts the operational paradigm from a reactive model (fixing problems after they occur) to a proactive and predictive one (anticipating and preventing issues before they impact users or services).131 This proactive stance is essential for maintaining the high levels of performance and reliability demanded by AI workloads running on the network.

However, the effectiveness of AIOps is heavily contingent on the availability of large volumes of high-quality, diverse, and timely network telemetry data.5 AI models require this data for training and real-time analysis. This dependency creates new challenges related to data infrastructure. Implementing successful AIOps requires establishing robust data collection mechanisms across the entire network (including physical, virtualized, edge, and cloud domains), potentially investing in data lakes or specialized platforms for storing and processing this vast amount of telemetry data 17, and ensuring data quality and governance. This adds another layer of complexity and potential cost to the network modernization journey, highlighting the deep interdependencies between network infrastructure, data infrastructure, and AI capabilities.

3.6 Pillar 6: Seamless Cloud and Edge Integration

A comprehensive network modernization strategy for AI must explicitly address the increasingly distributed nature of AI deployments. It is rare for AI workloads to reside solely in one location; more commonly, they span a continuum from edge devices to on-premises data centers and multiple public or private clouds 11-.43 Recognizing and designing for this hybrid reality is a critical pillar of AI readiness.

The different locations offer complementary strengths for AI:

• Cloud: Provides virtually limitless, scalable compute power and vast storage capacity, ideal for training large, complex AI models and aggregating data from diverse sources for analysis.101

• Edge: Offers proximity to data generation and consumption points, enabling the low latency required for real-time inference and decision-making. It also allows for local data filtering and processing, reducing bandwidth consumption back to the cloud, and enables operation in environments with intermittent or no connectivity.39

Given these distinct advantages, hybrid architectures that strategically combine edge, on-premises, private cloud, and public cloud resources are becoming essential for optimally supporting the diverse needs of different AI workloads.43 The optimal placement for a specific AI task depends on factors like its latency sensitivity, data volume and gravity, processing requirements, security needs, and cost considerations.

This distributed model places a premium on seamless, reliable, and high-performance connectivity between these disparate environments. Efficient data transfer is required for moving training datasets to the cloud, deploying trained models to the edge, collecting inference results or filtered data from the edge, and enabling communication between different components of a distributed AI application.11 Technologies facilitating this interconnectivity include Virtual Private Networks (VPNs), dedicated cloud interconnects (like AWS Direct Connect or Azure ExpressRoute), Software-Defined Wide Area Networks (SD-WAN) for managing connectivity across branches and clouds, and high-speed wireless links like 5G.43 Managing data synchronization and maintaining data consistency across the edge-cloud continuum also presents significant technical challenges.107

Effectively integrating edge and cloud for AI requires more than just establishing network connectivity between them. It necessitates intelligent orchestration and sophisticated data management strategies.101 This involves capabilities like:

• Model Partitioning: Deciding which parts of an AI model or workflow should run at the edge versus in the cloud to optimize for latency, cost, or privacy.

• Federated Learning: Training models directly on edge devices using local data without sending raw data to the cloud, enhancing privacy and reducing bandwidth needs.

• Data Orchestration: Managing the flow, synchronization, and consistency of data across the distributed environment.

• Unified Management: Providing a single pane of glass for deploying, monitoring, and managing AI workloads and infrastructure across the edge-to-cloud continuum. These requirements highlight the need for advanced platforms and frameworks that can intelligently manage the entire distributed AI ecosystem, not just provide the network links.

Furthermore, the optimal balance between edge processing and cloud processing is not fixed; it is highly dependent on the specific AI use case and its unique requirements regarding latency, bandwidth, data privacy, computational intensity, and cost.41 Some applications benefit immensely from the real-time responsiveness of the edge 47, while others rely on the massive scale and power of the cloud.158 As AI models evolve, edge hardware capabilities improve, and network technologies advance, the ideal deployment strategy for any given application may shift over time. This dynamic necessitates network architectures that are inherently adaptable and flexible. Leveraging technologies like SDN, NFV, and programmable cloud connectivity options allows organizations to reconfigure data flows and adjust workload placements between edge and cloud environments as needed, ensuring the network can support the evolving landscape of AI deployment patterns.

4. Key Technologies Driving AI-Ready Networks

The modernization of network infrastructure to meet the demands of AI is enabled by a confluence of key technologies. These technologies provide the necessary programmability, virtualization, speed, low latency, and intelligence required for AI-ready networks.

4.1 Software-Defined Networking (SDN): Programmability and Agility

SDN represents a fundamental shift in network architecture by separating the network's control logic (control plane) from the underlying hardware that forwards data packets (data plane).34 This separation allows network behavior to be controlled and managed centrally through software, specifically via an SDN controller, which acts as the "brain" of the network.58 The controller communicates with network devices (switches, routers) using southbound Application Programming Interfaces (APIs), with OpenFlow being a prominent standardized protocol for this purpose.55 Northbound APIs allow applications (including network management tools, orchestration platforms, and even AI applications themselves) to request network services or convey requirements to the controller.57

This programmability and centralized control offer significant advantages for supporting AI workloads:

• Dynamic Resource Allocation: SDN enables the network to dynamically adapt to the specific needs of AI applications. The controller can automatically provision higher bandwidth for data-intensive training tasks, prioritize latency-sensitive inference traffic using QoS mechanisms, or establish optimized network paths based on real-time requirements, moving beyond static configurations.55

• Centralized Management and Automation: Managing the complex network topologies often required for AI (e.g., large data center fabrics, edge deployments) is simplified through the centralized SDN controller. This reduces the need for manual, device-by-device configuration, minimizes errors, and enables automation of routine network tasks like provisioning and policy updates.55

• Enhanced Visibility and Security Control: The centralized controller provides a global view of the network state and traffic flows, facilitating more effective monitoring, troubleshooting, and traffic engineering.57 It also allows for the consistent and dynamic enforcement of security policies, including granular micro-segmentation to isolate AI workloads and contain potential threats.34

• Agility and Flexibility: SDN makes the network infrastructure more responsive to change. New AI applications can be deployed faster, network configurations can be modified rapidly to optimize performance, and the infrastructure can adapt more easily to evolving business needs or technological advancements.34

Real-world examples illustrate SDN's impact. Telecommunications providers utilize SDN for managing virtualized infrastructure and enabling dynamic service chaining.172 Google famously implemented an SDN-based WAN (B4) to optimize traffic flow between its global data centers.58 While specific case studies directly linking SDN to quantifiable AI workload improvements require careful validation, the principles demonstrate its potential for dynamic resource allocation in AI-driven systems like manufacturing quality control or healthcare applications.135

Despite its benefits, SDN implementation faces challenges. Ensuring the scalability and reliability of the centralized controller is crucial, as its failure can impact the entire network segment it manages.164 Security of the controller itself is paramount, as its compromise grants control over the network.35 The complexity of deploying and managing SDN environments, potential interoperability issues between different vendors' solutions, and the need for IT staff with new programming and automation skills are also significant hurdles 34-.35

The core value proposition of SDN for AI lies in its transformation of the network from a collection of statically configured devices into a programmable resource. This programmability allows the network's behavior to be tailored in real-time to the specific, diverse, and often dynamic performance requirements (bandwidth, latency, priority) of different AI workloads.57

However, unlocking this potential fully depends not just on the SDN architecture itself, but on the sophistication of the SDN controller and the applications built atop it (via northbound APIs). For SDN to effectively manage AI traffic, the controller requires intelligence—potentially incorporating AI algorithms itself 64—to understand the nuanced needs of various AI jobs. Furthermore, applications are needed that can translate high-level AI workload requirements (e.g., "start high-bandwidth training job X," "prioritize low-latency inference for application Y") into the specific, low-level flow rules or configuration commands (like OpenFlow instructions) executed by the data plane devices.61 This highlights the critical importance of the software ecosystem surrounding SDN controllers in realizing the vision of an AI-optimized network.

4.2 Network Function Virtualization (NFV): Flexibility and Efficiency

Complementary to SDN, Network Functions Virtualization (NFV) focuses on decoupling network functions—such as routing, firewalling, load balancing, WAN acceleration, intrusion detection, and even core components of 5G networks—from the dedicated physical hardware appliances they traditionally run on.37 In the NFV paradigm, these functions are implemented as software, known as Virtual Network Functions (VNFs) or, increasingly, as containerized Cloud-Native Network Functions (CNFs). These software-based functions run on standard, commodity IT infrastructure—servers, storage, and switches—leveraging virtualization or containerization technologies.37 The entire lifecycle and orchestration of these virtualized functions are managed by a framework known as NFV Management and Orchestration (MANO), standardized by the European Telecommunications Standards Institute (ETSI) 37-.74

NFV offers compelling advantages for building networks capable of supporting AI workloads:

• Agility and Faster Service Deployment: Perhaps the most significant benefit is the ability to rapidly deploy, update, reconfigure, or decommission network services required by AI applications without the lengthy procurement and installation cycles associated with physical hardware.37 This agility supports DevOps practices for network services.122

• Scalability and Elasticity: VNFs and CNFs can be scaled dynamically—instantiating more instances to handle increased load or scaling down during idle periods—based on the fluctuating demands of AI workloads, ensuring resources are available when needed without being permanently overallocated.37

• Cost Reduction: By utilizing standard, lower-cost IT hardware instead of expensive, specialized network appliances, NFV can significantly reduce Capital Expenditures (CapEx). Operational Expenditures (OpEx) can also be lowered through automation, hardware consolidation, reduced physical footprint, and potentially lower power consumption.37

• Resource Efficiency: Consolidating multiple network functions onto shared, standard hardware improves the utilization of compute, storage, and network resources compared to deploying numerous dedicated physical boxes.37

• Flexibility and Vendor Independence: NFV frees organizations from dependency on specific hardware vendors. It allows for the selection of best-of-breed VNF software from different suppliers and enables the placement of network functions strategically where they are most needed, such as at the network edge to support low-latency AI applications.37

NFV and SDN are highly complementary technologies that often work in synergy. While NFV focuses on virtualizing the network functions themselves (the 'what'), SDN provides the programmable control plane needed to dynamically connect, steer traffic through, and manage these virtualized functions (the 'how').55 For example, SDN can automate the creation of "service chains," directing traffic sequentially through multiple VNFs (like a firewall, then a load balancer, then an intrusion detection system) based on policy. Together, SDN and NFV enable the vision of a highly automated, flexible, and software-driven network infrastructure.55

However, NFV adoption is not without challenges. Running network functions in software on general-purpose hardware can sometimes introduce performance overhead compared to highly optimized dedicated hardware, although techniques like the Data Plane Development Kit (DPDK) and Single Root I/O Virtualization (SR-IOV) are used to mitigate this.71 The Management and Orchestration (MANO) layer required to manage the NFV environment adds complexity.177 Integrating VNFs with existing physical networks and legacy systems can be difficult.34 Security in shared virtualized environments requires careful attention, addressing risks like multi-tenancy interference or VNF compromise.66 Interoperability between components from different vendors can still be a concern despite standardization efforts 70, and managing NFV infrastructure requires new skill sets within IT teams.71

A key contribution of NFV to AI network modernization is its ability to make the network infrastructure services supporting AI as agile and scalable as the cloud-based compute and storage resources often used for AI model development and deployment.158 AI workloads benefit greatly from the elasticity of cloud platforms, but if the required network services (security, load balancing, connectivity) remain tied to rigid, physical hardware, this creates an agility mismatch. NFV allows these essential network functions to be provisioned, scaled, and managed programmatically and on-demand, mirroring the flexibility of virtual machines or containers.37 This alignment across compute, storage, and network domains is crucial for achieving true end-to-end agility and efficiency in deploying and operating AI applications.

The realization of NFV's benefits, however, is critically dependent on the capabilities and robustness of the Management and Orchestration (MANO) framework 37-.74 The MANO layer is responsible for the complex tasks of VNF lifecycle management (instantiation, scaling, termination, healing), virtual resource allocation across the NFV Infrastructure (NFVI), and orchestrating end-to-end network services composed of multiple VNFs. The effectiveness, automation level, scalability, and usability of the MANO system directly dictate whether NFV delivers on its promise of agility and efficiency.177 A poorly implemented, overly complex, or unreliable MANO layer can become a significant operational bottleneck, negating the advantages of virtualization. Therefore, the MANO component is a critical element—and potential point of failure or complexity—that must be carefully considered and selected when building an NFV-based network ready for AI workloads.

4.3 5G and Beyond: Unleashing High-Speed, Low-Latency Connectivity

Fifth-generation (5G) cellular technology, along with its anticipated successor 6G, represents a paradigm shift in wireless communication, offering capabilities far beyond previous generations like 4G/LTE, and playing a pivotal role in enabling advanced AI applications, particularly outside the traditional data center 48-.49 Key characteristics that make 5G transformative for AI networking include:

• Significantly Higher Bandwidth and Speeds: 5G is designed to deliver peak download speeds potentially reaching 10-20 Gbps, a dramatic increase over 4G, enabling the rapid transmission of large data volumes required by many AI applications.48 This is achieved through the use of wider spectrum bandwidths, including new mid-band and high-band (millimeter wave - mmWave) frequencies, alongside advanced antenna techniques like Massive MIMO.49

• Ultra-Low Latency: A defining feature of 5G is its potential for drastically reduced latency, targeting end-to-end delays in the sub-10 millisecond range, and potentially as low as 1 millisecond for certain use cases (often associated with Ultra-Reliable Low-Latency Communications - URLLC).42 This near-instantaneous responsiveness is critical for real-time AI applications.

• Massive Machine-Type Communications (mMTC): 5G architecture is designed to efficiently connect a vastly larger number of devices per unit area compared to 4G, supporting the proliferation of IoT sensors and devices that generate data for AI systems.52

• Enhanced Reliability: 5G incorporates features aimed at improving network reliability and availability, particularly through URLLC specifications designed for mission-critical communications.42

• Network Slicing: A key architectural innovation in 5G, network slicing allows operators to create multiple virtual, logically isolated end-to-end networks on top of a common physical infrastructure. Each slice can be customized with specific performance characteristics (e.g., a low-latency slice for autonomous vehicles, a high-bandwidth slice for video streaming, an mMTC slice for IoT sensors), providing tailored connectivity for diverse AI application needs.37

These capabilities position 5G as a crucial enabler for AI network modernization in several ways:

• Enabling Edge AI: 5G provides the essential high-bandwidth, low-latency wireless link needed to connect edge computing nodes and devices, making real-time AI processing at the network edge feasible and effective for a wide range of applications.51

• Supporting Mobile AI Applications: The performance enhancements of 5G unlock the potential for sophisticated AI applications running directly on mobile devices, connected vehicles, drones, and robots, enabling capabilities previously limited to wired environments.42

• Connecting Massive IoT Data Sources: 5G's capacity to handle a massive density of connections is vital for the large-scale deployment of IoT sensors that collect the real-world data fueling many AI analytics and control systems.52

• Providing Flexible and Differentiated Services: Network slicing allows service providers to offer tailored connectivity services optimized for the specific requirements (latency, bandwidth, reliability) of different AI use cases, potentially creating new revenue streams.37

Challenges to widespread 5G adoption include the significant cost of deploying new infrastructure (especially mmWave, which requires denser cell site placement), securing sufficient spectrum licenses, ensuring robust security across the new architecture, and integrating 5G networks smoothly with existing fixed and wireless infrastructure.181

It is important to recognize that 5G represents more than just an incremental speed increase over 4G. Its fundamental architectural design, incorporating native support for ultra-low latency (URLLC), massive device connectivity (mMTC), network slicing, and integration with edge computing, is specifically geared towards enabling new classes of applications dominated by AI and IoT.37 These capabilities directly map to the diverse and demanding network requirements identified for various AI workloads, positioning 5G as a foundational technology for the next wave of AI-driven innovation.

However, realizing the full transformative potential of 5G for AI hinges on its integration within a broader ecosystem. The ultra-low latency benefits of 5G are maximized only when the computational resources are also located nearby, highlighting the critical synergy between 5G and edge computing.50 Furthermore, managing the inherent complexity of 5G networks—with their dynamic traffic patterns, diverse service requirements enabled by slicing, and distributed edge deployments—efficiently and reliably necessitates the use of AI-driven network management and automation (AIOps).53 Thus, 5G, edge computing, and AI for network operations form a powerful, interdependent triad. Each technology enables and enhances the others, collectively paving the way for the deployment and scaling of advanced, real-time, distributed AI applications.

4.4 Advanced Routing Protocols: Intelligent Path Optimization

Routing protocols are the fundamental mechanisms that determine the paths data packets traverse across complex networks. Traditional routing protocols often prioritize finding the shortest path between source and destination, which may not be optimal for the diverse and demanding traffic patterns generated by AI workloads.8 AI applications can produce traffic with specific needs (e.g., extreme low latency for control systems, sustained high bandwidth for data transfers) and challenging characteristics (e.g., bursty inference requests, massive "elephant flows" during model training, many-to-one communication patterns in distributed computing) that can overwhelm networks optimized solely for shortest-path routing.8

To address these challenges, advanced routing protocols and techniques are being developed and deployed, incorporating greater intelligence and adaptability:

• Context-Awareness: Modern routing approaches increasingly consider factors beyond simple hop count. This includes:

• Latency Sensitivity: Explicitly selecting paths with lower delay to meet the requirements of time-critical AI applications.8

• Congestion Awareness: Monitoring network links for congestion in real-time and dynamically

Works cited

1. What Is Artificial Intelligence (AI)? | Google Cloud, accessed April 16, 2025, https://cloud.google.com/learn/what-is-artificial-intelligence

2. What Is Artificial Intelligence (AI)? | IBM, accessed April 16, 2025, https://www.ibm.com/think/topics/artificial-intelligence

3. AI as a Strategic Imperative: Insights from the Latest Industry Report ..., accessed April 16, 2025, https://global.hitachi-solutions.com/blog/ai-as-a-strategic-imperative/

4. The Impact of AI on Business Strategy: What Leaders Need to Know - CMIT Solutions, accessed April 16, 2025, https://cmitsolutions.com/rochester-ny-1109/blog/the-impact-of-ai-on-business-strategy-what-leaders-need-to-know/

5. What is artificial intelligence (AI) in networking? - Neos Networks, accessed April 16, 2025, https://neosnetworks.com/resources/blog/what-is-ai-in-networking/

6. How AI Can Enhance Your Business Strategy and Competitive Edge, accessed April 16, 2025, https://www.redapt.com/blog/how-ai-can-enhance-your-business-strategy-and-competitive-edge

7. www.lenovo.com, accessed April 16, 2025, https://www.lenovo.com/us/en/glossary/ai-networking/#:~:text=AI%20transforms%20network%20decision%2Dmaking,metrics%20to%20make%20informed%20decisions.

8. AI Networking: Understanding AI Networking Concepts | Lenovo US, accessed April 16, 2025, https://www.lenovo.com/us/en/glossary/ai-networking/

9. Generative AI: What Is It, Tools, Models, Applications and Use Cases - Gartner, accessed April 16, 2025, https://www.gartner.com/en/topics/generative-ai

10. Network Optimization for AI: Best Practices and Strategies, accessed April 16, 2025, https://blog.centurylink.com/network-optimization-for-ai-best-practices-and-strategies/

11. Networking for Artificial Intelligence (AI) – Intel, accessed April 16, 2025, https://www.intel.com/content/www/us/en/learn/ai-networking.html

12. Networking for AI workloads - Nokia.com, accessed April 16, 2025, https://www.nokia.com/data-center-networks/networking-for-ai-workloads/

13. Impact of Networking Protocols on AI Data Center Efficiency Strategies - Yotta, accessed April 16, 2025, https://colocation.yotta.com/blog/evaluating-the-impact-of-networking-protocols-on-ai-data-center-efficiency/

14. A new strategic imperative in private equity: The AI operating partner | Heidrick & Struggles, accessed April 16, 2025, https://www.heidrick.com/en/pages/aida/a-new-strategic-imperative-in-private-equity_the-ai-operating-partner

15. The CIO Imperative: Six Priorities for the AI-Fueled Organization | IDC Blog, accessed April 16, 2025, https://blogs.idc.com/2025/03/24/the-cio-imperative-six-priorities-for-the-ai-fueled-organization/

16. Exploring the Impact of AI on Everyday Business Operations - Redapt, accessed April 16, 2025, https://www.redapt.com/blog/exploring-the-impact-of-ai-on-everyday-business-operations

17. What is artificial intelligence (AI) for networking? | Juniper Networks ..., accessed April 16, 2025, https://www.juniper.net/us/en/research-topics/what-is-ai-for-networking.html

18. AI-Powered Traffic Optimization: A Paradigm Shift in Network Management - Corpus Publishers, accessed April 16, 2025, https://www.corpuspublishers.com/assets/articles/ctes-v5-25-1072.pdf

19. (PDF) INTELLIGENT NETWORK OPTIMIZATION: REVOLUTIONIZING NETWORK MANAGEMENT THROUGH AI AND ML - ResearchGate, accessed April 16, 2025, https://www.researchgate.net/publication/387713419_INTELLIGENT_NETWORK_OPTIMIZATION_REVOLUTIONIZING_NETWORK_MANAGEMENT_THROUGH_AI_AND_ML

20. Impact of AI Adoption on IT Infrastructure in 2025 - Zones Blog, accessed April 16, 2025, https://blog.zones.com/impact-of-ai-adoption-on-it-infrastructure-in-2025

21. Legacy Systems Are Holding You Back: How AI and Automation Are Shaping the Future of Enterprise Networking - Presidio, accessed April 16, 2025, https://www.presidio.com/blogs/legacy-systems-are-holding-you-back-how-ai-and-automation-are-shaping-the-future-of-enterprise-networking/

22. Solving Latency Challenges in AI Data Centers - WEKA, accessed April 16, 2025, https://www.weka.io/blog/ai-ml/solving-latency-challenges-in-ai-data-centers/

23. Empowering AI Workloads with High-Performance Storage from GreenNode, accessed April 16, 2025, https://greennode.ai/blog/empowering-ai-workloads-with-high-performance-storage-from-greennode

24. Cisco Study: Only 13% of Companies Ready for AI Despite Urgent Push to Deploy | CSCO Stock News, accessed April 16, 2025, https://www.stocktitan.net/news/CSCO/cisco-s-2024-ai-readiness-index-urgency-rises-readiness-d4gi6ps6cji4.html

25. Cisco's 2024 AI Readiness Index: urgency rises, readiness falls | Digitalisation World, accessed April 16, 2025, https://digitalisationworld.com/news/69035/ciscos-2024-ai-readiness-index-urgency-rises-readiness-falls

26. Cisco Launches New Research, Highlighting Seismic Gap in Companies' Preparedness for AI, accessed April 16, 2025, https://newsroom.cisco.com/c/r/newsroom/en/us/a/y2023/m11/cisco-launches-new-research-highlighting-seismic-gap-in-companies-preparedness-for-ai.html

27. Modern Network Infrastructure: 3 Principles for Connected Data and AI-Ready Architecture, accessed April 16, 2025, https://www.digitalrealty.com/resources/articles/network-infrastructure

28. Optimizing Networking for AI Workloads: A Comprehensive Guide - UfiSpace, accessed April 16, 2025, https://www.ufispace.com/company/blog/networking-for-ai-workloads

29. Optimize your AI Network to Keep Your AI Flowing | Dell USA, accessed April 16, 2025, https://www.dell.com/en-us/blog/optimize-your-ai-network-to-keep-your-ai-flowing/

30. What is Artificial Intelligence (AI) networking? - DriveNets, accessed April 16, 2025, https://drivenets.com/resources/education-center/what-is-ai-networking/

31. Accelerating AI Networks with DDN's Data Intelligence Platform and NVIDIA Spectrum™-X for Storage - DDN, accessed April 16, 2025, https://www.ddn.com/resources/whitepapers/accelerating-ai-networks-with-ddns-data-intelligence-platform-and-nvidia-spectrum-x-for-storage/

32. AI readiness in India declines, with only 18% of organisations fully prepared for deployment, accessed April 16, 2025, https://marketech-apac.com/ai-readiness-in-india-declines-with-only-18-of-organisations-fully-prepared-for-deployment/

33. Prep Your Digital Infrastructure for the Agentic AI Future - Reworked, accessed April 16, 2025, https://www.reworked.co/digital-workplace/prep-your-digital-infrastructure-for-the-agentic-ai-future/

34. Enhanced Network Security by Implementing SDN and NFV and New Routing Algorithm - IIETA, accessed April 16, 2025, https://www.iieta.org/download/file/fid/163615

35. State University of New York Polytechnic Institute SECURITY CHALLENGES IN SDN IMPLEMENTATION, accessed April 16, 2025, https://soar.suny.edu/bitstream/handle/20.500.12648/1081/P%20Patil%20Thesis%20doc%20Final.pdf?sequence=1&isAllowed=y

36. AI Integration Challenges: Insights for Competitive Edge - Aura Intelligence, accessed April 16, 2025, https://blog.getaura.ai/ai-integration-challenges

37. Network Functions Virtualization (NFV) explained - Ericsson, accessed April 16, 2025, https://www.ericsson.com/en/nfv

38. Low Latency 5G Distributed Wireless Network Architecture: A Techno-Economic Comparison - MDPI, accessed April 16, 2025, https://www.mdpi.com/2411-5134/6/1/11

39. What Is the Network Edge? — Intel, accessed April 16, 2025, https://www.intel.com/content/www/us/en/edge-computing/what-is-the-network-edge.html

40. Understanding Network Requirements - NetBox Labs, accessed April 16, 2025, https://netboxlabs.com/blog/network-requirements/

41. Network Latency vs. Compute Latency - Interconnections - The Equinix Blog, accessed April 16, 2025, https://blog.equinix.com/blog/2024/03/27/network-latency-vs-compute-latency/

42. Enabling Tactile Internet via 6G: Application Characteristics, Requirements, and Design Considerations - MDPI, accessed April 16, 2025, https://www.mdpi.com/1999-5903/17/3/122

43. What Is Hybrid Cloud Architecture? - IBM, accessed April 16, 2025, https://www.ibm.com/think/topics/hybrid-cloud-architecture

44. What Is Edge Computing? | Gcore, accessed April 16, 2025, https://gcore.com/learning/what-is-edge-computing

45. Enhancing 5G Networks with Edge Computing: An Overview Study - ITM Web of Conferences, accessed April 16, 2025, https://www.itm-conferences.org/articles/itmconf/pdf/2024/12/itmconf_maih2024_04010.pdf

46. Edge Computing vs. CDN: Optimizing Data Delivery Solutions - FastPix, accessed April 16, 2025, https://www.fastpix.io/blog/edge-computing-vs-cdn-identifying-their-roles-in-data-delivery

47. A beginner's guide to AI Edge computing: How it works and its benefits | Flexential, accessed April 16, 2025, https://www.flexential.com/resources/blog/beginners-guide-ai-edge-computing

48. www.qualcomm.com, accessed April 16, 2025, https://www.qualcomm.com/5g/what-is-5g#:~:text=5G%20wireless%20technology%20is%20meant,experiences%20and%20connects%20new%20industries.

49. What is 5G? - 5G Network Explained - AWS, accessed April 16, 2025, https://aws.amazon.com/what-is/5g/

50. The Open Edge Future – Content delivery and what comes next - Qwilt, accessed April 16, 2025, https://www.qwilt.com/the-open-edge-future-content-delivery-and-what-comes-next/

51. What Is Network Virtualization? | Supermicro, accessed April 16, 2025, https://www.supermicro.com/en/glossary/network-virtualization

52. What is 5G | Everything You Need to Know About 5G | 5G FAQ - Qualcomm, accessed April 16, 2025, https://www.qualcomm.com/5g/what-is-5g

53. What is 5G? How will it transform our world? - Ericsson, accessed April 16, 2025, https://www.ericsson.com/en/5g

54. The Evolution Of Mobile Edge Intelligence: Exploring The Synergy Of AI, Edge Computing, And 5G Networks - ResearchGate, accessed April 16, 2025, https://www.researchgate.net/publication/388186147_The_Evolution_Of_Mobile_Edge_Intelligence_Exploring_The_Synergy_Of_AI_Edge_Computing_And_5G_Networks

55. SDN & NFV: Moving the Network into the Cloud - CableLabs, accessed April 16, 2025, https://www.cablelabs.com/blog/sdn-nfv

56. Mobile Fog Computing by Using SDN/NFV on 5G Edge Nodes - Tech Science Press, accessed April 16, 2025, https://www.techscience.com/csse/v41n2/45191/html

57. Top Benefits of Software-Defined Networking (SDN) - SynchroNet, accessed April 16, 2025, https://synchronet.net/benefits-of-software-defined-networking/

58. What Is Software-Defined Networking? | Built In, accessed April 16, 2025, https://builtin.com/software-defined-networking

59. What Is Software-Defined Networking (SDN)? - Supermicro, accessed April 16, 2025, https://www.supermicro.com/en/glossary/sdn

60. What Is OpenFlow? How Does It Relate to SDN? - FS.com, accessed April 16, 2025, https://www.fs.com/de-en/blog/what-is-openflow-how-does-it-relate-to-sdn-11322.html

61. Software-defined Networking (SDN): Revolutionizing Network Management and Control, accessed April 16, 2025, https://algocademy.com/blog/software-defined-networking-sdn-revolutionizing-network-management-and-control/

62. (PDF) Exploring Traffic Patterns Through Network Programmability: Introducing SDNFLow, a Comprehensive OpenFlow-Based Statistics Dataset for Attack Detection - ResearchGate, accessed April 16, 2025, https://www.researchgate.net/publication/379070720_Exploring_Traffic_Patterns_Through_Network_Programmability_Introducing_SDNFLow_a_Comprehensive_OpenFlow-Based_Statistics_Dataset_for_Attack_Detection

63. 11 Benefits of Software-Defined Networking (SDN) - Trigyn, accessed April 16, 2025, https://www.trigyn.com/insights/11-benefits-software-defined-networking-sdn

64. Deep Dive into Software-Defined Networking (SDN): Transforming the Future of Network Management - Layer8Packet, accessed April 16, 2025, https://www.layer8packet.io/home/av18mamuvk1e7und644swv3xu4g53p

65. What is Network Functions Virtualization (NFV)? - VMware, accessed April 16, 2025, https://www.vmware.com/topics/network-functions-virtualization-nfv

66. Security and Privacy Issues in Network Function Virtualization: A Review from Architectural Perspective - The Science and Information (SAI) Organization, accessed April 16, 2025, https://thesai.org/Downloads/Volume15No6/Paper_49-Security_and_Privacy_Issues_in_Network_Function_Virtualization.pdf

67. Network Function Virtualization: State-of-the-Art and Research Challenges - ResearchGate, accessed April 16, 2025, https://www.researchgate.net/publication/281524200_Network_Function_Virtualization_State-of-the-Art_and_Research_Challenges

68. How SDN and NFV Technologies Are Transforming Network Management, accessed April 16, 2025, https://blog.equinix.com/blog/2019/01/17/how-sdn-and-nfv-technologies-are-transforming-network-management/

69. Software-Defined Networking Challenges and Research Opportunities for Future Interest, accessed April 16, 2025, https://www.ijraset.com/research-paper/software-defined-networking-challenges-and-research-opportunities

70. The Potential of SDN and NFV in Next-Generation Networks - Telecom Review, accessed April 16, 2025, https://www.telecomreview.com/articles/reports-and-coverage/7712-the-potential-of-sdn-and-nfv-in-next-generation-networks

71. Chapter 1. Understanding Red Hat Network Functions Virtualization (NFV), accessed April 16, 2025, https://docs.redhat.com/en/documentation/red_hat_openstack_platform/16.2/html/network_functions_virtualization_planning_and_configuration_guide/understanding-nfv_rhosp-nfv

72. Benefits of Network Function Virtualization (NFV) - Intraway, accessed April 16, 2025, https://www.intraway.com/blog/nfv-benefits/

73. ETSI NFV - Secureframe, accessed April 16, 2025, https://secureframe.com/frameworks-glossary/esti-nfv

74. ETSI GS NFV-IFA 010 V4.5.1 (2023-09) - iTeh Standards, accessed April 16, 2025, https://cdn.standards.iteh.ai/samples/68062/c5fa95e5926f45dba62e8b867a60333d/ETSI-GS-NFV-IFA-010-V4-5-1-2023-09-.pdf

75. What Is AI Networking in Data Centers? - eStruxture, accessed April 16, 2025, https://www.estruxture.com/blog/what-is-ai-networking-in-data-centers

76. News - What are the network requirements for AI? - AIPU, accessed April 16, 2025, https://www.aipuwaton.com/news/networking-for-ai-workloads-what-are-the-network-requirements-for-ai/

77. Network Optimization for AI: Best Practices and Strategies, accessed April 16, 2025, https://blog.centurylink.com/network-optimization-for-ai-best-practices-and-strategies/?utm_source=rss&utm_medium=rss&utm_campaign=network-optimization-for-ai-best-practices-and-strategies

78. AI's Impact on Data Centers and Bandwidth Requirements - LOGIX Fiber Networks, accessed April 16, 2025, https://logix.com/ai-impact-data-centers-bandwidth-fiber-networks/

79. AI-Driven Networking | InterGlobix Magazine, accessed April 16, 2025, https://www.interglobixmagazine.com/ai-driven-networking/

80. How AI Changes Your Network Infrastructure Requirements - The Equinix Blog, accessed April 16, 2025, https://blog.equinix.com/blog/2025/04/16/how-ai-changes-your-network-infrastructure-requirements/

81. Chapter 8 - Networking and Security | AI in Production Guide - Azure documentation, accessed April 16, 2025, https://azure.github.io/AI-in-Production-Guide/chapters/chapter_08_securing_cargo_networking_security

82. AI Model Security Protecting Against Adversarial Attacks and Model Theft - ResearchGate, accessed April 16, 2025, https://www.researchgate.net/publication/390200597_AI_Model_Security_Protecting_Against_Adversarial_Attacks_and_Model_Theft/download

83. Navigating the Challenges of AI Infrastructure Design: Balancing Power, Latency, Reliability, and Data Requirements - F5, accessed April 16, 2025, https://www.f5.com/es_es/resources/white-papers/overcoming-ai-infrastructure-challenges-balancing-power-latency-reliability-and-data-requirements

84. Advanced Networks for Artificial Intelligence and Machine Learning Computing | AFL Hyperscale, accessed April 16, 2025, https://www.aflhyperscale.com/wp-content/uploads/2024/10/Advanced-Networks-for-Artificial-Intelligence-and-Machine-Learning-Computing-White-Paper.pdf

85. Secure And Scalable Networks: Your Key To AI Success - Lumen Blog, accessed April 16, 2025, https://blog.lumen.com/secure-and-scalable-networks-your-key-to-ai-success/

86. 6 Types of AI Workloads, Challenges & Critical Best Practices - Cloudian, accessed April 16, 2025, https://cloudian.com/guides/ai-infrastructure/6-types-of-ai-workloads-challenges-and-critical-best-practices/

87. Networking recommendations for AI workloads on Azure infrastructure (IaaS) - Cloud Adoption Framework | Microsoft Learn, accessed April 16, 2025, https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/scenarios/ai/infrastructure/networking

88. The Rise of AI: Fueling the Shift to 800G Ethernet - FS.com, accessed April 16, 2025, https://www.fs.com/blog/the-rise-of-ai-fueling-the-shift-to-800g-ethernet-16586.html

89. AI Workloads Put Data Center Performance to the Test - Spirent, accessed April 16, 2025, https://www.spirent.com/blogs/ai-workloads-put-data-center-performance-to-the-test

90. AI Networking - Arista Networks White Paper, accessed April 16, 2025, https://www.arista.com/assets/data/pdf/Whitepapers/AI-Network-WP.pdf

91. AI-Assisted Low Information Latency Wireless Networking - Clemson University, accessed April 16, 2025, https://people.computing.clemson.edu/~jmarty/projects/lowLatencyNetworking/papers/Simulators-emulators/CoreEMANEQuagga/AIAssistedLowLatencyWirelessNetworking-2020.pdf

92. MIT Open Access Articles Low-Latency Networking: Where Latency Lurks and How to Tame It - DSpace@MIT, accessed April 16, 2025, https://dspace.mit.edu/bitstream/handle/1721.1/126300/1808.02079.pdf?sequence=2&isAllowed=y

93. An Analysis of Software Latency for a High-Speed Autonomous Race Car—A Case Study in the Indy Autonomous Challenge - ResearchGate, accessed April 16, 2025, https://www.researchgate.net/publication/368369298_An_Analysis_of_Software_Latency_for_a_High-Speed_Autonomous_Race_Car-A_Case_Study_in_the_Indy_Autonomous_Challenge

94. NeuOS: A Latency-Predictable Multi-Dimensional Optimization Framework for DNN-driven Autonomous Systems | USENIX, accessed April 16, 2025, https://www.usenix.org/system/files/atc20-bateni.pdf

95. Real-Time Inference and Low-Latency Models - [x]cube LABS, accessed April 16, 2025, https://www.xcubelabs.com/blog/real-time-inference-and-low-latency-models/

96. Distributed inference with collaborative AI agents for Telco-powered Smart-X - AWS, accessed April 16, 2025, https://aws.amazon.com/blogs/industries/distributed-inference-with-collaborative-ai-agents-for-telco-powered-smart-x/

97. Optimizing AI responsiveness: A practical guide to Amazon Bedrock latency-optimized inference | AWS Machine Learning Blog, accessed April 16, 2025, https://aws.amazon.com/blogs/machine-learning/optimizing-ai-responsiveness-a-practical-guide-to-amazon-bedrock-latency-optimized-inference/

98. Understanding Latency in AI: What It Is and How It Works - Galileo AI, accessed April 16, 2025, https://www.galileo.ai/blog/understanding-latency-in-ai-what-it-is-and-how-it-works

99. Solving AI Foundational Model Latency with Telco Infrastructure - arXiv, accessed April 16, 2025, https://arxiv.org/pdf/2504.03708

100. (PDF) Edge-Cloud Synergy in Real-Time AI Applications : Opportunities, Implementations, and Challenges - ResearchGate, accessed April 16, 2025, https://www.researchgate.net/publication/390274716_Edge-Cloud_Synergy_in_Real-Time_AI_Applications_Opportunities_Implementations_and_Challenges

101. AI's Edge Continuum: A new look at the cloud computing role in edge AI - Latent AI, accessed April 16, 2025, https://latentai.com/white-paper/ai-edge-continuum/

102. Moving AI to the edge: Benefits, challenges and solutions - Red Hat, accessed April 16, 2025, https://www.redhat.com/en/blog/moving-ai-edge-benefits-challenges-and-solutions

103. AI in 5G Networks: Advancements & Real-World Use Cases - HashStudioz Technologies, accessed April 16, 2025, https://www.hashstudioz.com/blog/ai-in-5g-networks-advancements-challenges-and-real-world-use-cases/

104. The Synergy of Private 5G, Edge Computing, and AI: Building the Autonomous Enterprise of the Future - Accelleran, accessed April 16, 2025, https://accelleran.com/the-synergy-of-private-5g-edge-computing-and-ai-building-the-autonomous-enterprise-of-the-future/

105. Exploring the Synergy of 5G and Edge Computing Across Industries - Tech Mahindra, accessed April 16, 2025, https://www.techmahindra.com/insights/views/exploring-synergy-5g-and-edge-computing-across-industries/

106. Low-Latency AI on 5G Edge - Deploying NVIDIA-Powered AI Inference Over 5G Networks, accessed April 16, 2025, https://www.lannerinc.com/news-and-events/latest-news/low-latency-ai-on-5g-edge-deploying-nvidia-powered-ai-inference-over-5g-networks

107. (PDF) EDGE-TO-CLOUD AI INTEGRATION: HYBRID ARCHITECTURES FOR REAL-TIME INFERENCE AND DATA PROCESSING IN IOT APPLICATIONS - ResearchGate, accessed April 16, 2025, https://www.researchgate.net/publication/390349135_EDGE-TO-CLOUD_AI_INTEGRATION_HYBRID_ARCHITECTURES_FOR_REAL-TIME_INFERENCE_AND_DATA_PROCESSING_IN_IOT_APPLICATIONS

108. Architecting Hybrid Edge-Cloud Solutions: Integration Patterns for Public Cloud Platforms, accessed April 16, 2025, https://www.researchgate.net/publication/390174734_Architecting_Hybrid_Edge-Cloud_Solutions_Integration_Patterns_for_Public_Cloud_Platforms/download

109. AI Network Intelligence 2024 Ultimate Guide - Rapid Innovation, accessed April 16, 2025, https://www.rapidinnovation.io/post/ai-agents-for-network-intelligence-use-cases-benefits-challenges

110. Software Defined Networking (SDNs) - Benefits, Challenges & Applications, accessed April 16, 2025, https://deliveredsocial.com/software-defined-networking-sdns-benefits-challenges-applications/

111. The NFV Management and Orchestration (MANO) framework as specified by ETSI (cf. [6]). - ResearchGate, accessed April 16, 2025, https://www.researchgate.net/figure/The-NFV-Management-and-Orchestration-MANO-framework-as-specified-by-ETSI-cf-6_fig5_333838235

112. 9 Effective Network Infrastructure Strategy Best Practices - TierPoint, accessed April 16, 2025, https://www.tierpoint.com/blog/network-infrastructure-strategy/

113. Enterprise AI Security Risks: Are You Truly Protected? - Inclusion Cloud, accessed April 16, 2025, https://inclusioncloud.com/insights/blog/enterprise-ai-security-risks/

114. AI Security Risks Uncovered: What You Must Know in 2025 - TTMS, accessed April 16, 2025, https://ttms.com/my/ai-security-risks-explained-what-you-need-to-know-in-2025/

115. Top 8 AI Security Best Practices - Sysdig, accessed April 16, 2025, https://sysdig.com/learn-cloud-native/top-8-ai-security-best-practices/

116. 6 Key Adversarial Attacks and Their Consequences - Mindgard AI, accessed April 16, 2025, https://mindgard.ai/blog/ai-under-attack-six-key-adversarial-attacks-and-their-consequences

117. AI TRiSM: Tackling Trust, Risk and Security in AI Models - Gartner, accessed April 16, 2025, https://www.gartner.com/en/articles/what-it-takes-to-make-ai-safe-and-effective

118. Layered Security: Your Ultimate Cyber Defense Strategy - PowerDMARC, accessed April 16, 2025, https://powerdmarc.com/layered-security-guide/

119. Network Security in 2025: Threats, Security Models and Technologies - Faddom, accessed April 16, 2025, https://faddom.com/network-security-in-2025-threats-security-models-and-technologies/

120. What is Defense-in-Depth? - Definition - CyberArk, accessed April 16, 2025, https://www.cyberark.com/what-is/defense-in-depth/

121. What Is Defence in Depth? An Introduction to Multi-Layered Security - Creative Networks, accessed April 16, 2025, https://www.creative-n.com/blog/what-is-defence-in-depth-an-introduction-to-multi-layered-security/

122. Improving Infrastructure Security Through NFV and SDN - CableLabs, accessed April 16, 2025, https://www.cablelabs.com/blog/improving-infrastructure-security-through-nfv-sdn

123. AI Data Governance: The Cornerstone of Responsible and Successful AI | Extreme Networks, accessed April 16, 2025, https://www.extremenetworks.com/resources/blogs/ai-data-governance

124. 10 Cyber Security Tools for 2025 - SentinelOne, accessed April 16, 2025, https://www.sentinelone.com/cybersecurity-101/cybersecurity/cyber-security-tools/

125. Top 10 Hybrid Cloud Security Solutions for 2025 - SentinelOne, accessed April 16, 2025, https://www.sentinelone.com/cybersecurity-101/cloud-security/hybrid-cloud-security-solutions/

126. Cyber Signals Issue 9 | AI-powered deception: Emerging fraud threats and countermeasures | Microsoft Security Blog, accessed April 16, 2025, https://www.microsoft.com/en-us/security/blog/2025/04/16/cyber-signals-issue-9-ai-powered-deception-emerging-fraud-threats-and-countermeasures/

127. Adaptive MFA: The Future of Dynamic Identity Security in 2025, accessed April 16, 2025, https://securityboulevard.com/2025/04/adaptive-mfa-the-future-of-dynamic-identity-security-in-2025/

128. Investing in Adaptive Security - Andreessen Horowitz, accessed April 16, 2025, https://a16z.com/announcement/investing-in-adaptive-security/

129. Adaptive Security: Simulates and Prevents AI-powered attacks - AI Cyber Insights, accessed April 16, 2025, https://aicyberinsights.com/adaptive-security-simulates-and-prevents-ai-powered-attacks/

130. What is Adaptive Network Security? Future-Proof Your Network | Nile, accessed April 16, 2025, https://nilesecure.com/ai-networking/adaptive-network-security

131. The Future of Intelligent Connectivity: Understanding AI-Driven Network Infrastructure, accessed April 16, 2025, https://www.nsi1.com/blog/ai-driven-network-infrastructure

132. Gen AI Tops Gartner's 2025 Cybersecurity Trends - Cyber Magazine, accessed April 16, 2025, https://cybermagazine.com/articles/gen-ai-tops-gartners-2025-cybersecurity-trends

133. Top Cybersecurity Trends and Strategies for Securing the Future | Gartner, accessed April 16, 2025, https://www.gartner.com/en/cybersecurity/topics/cybersecurity-trends

134. Enhancing Cybersecurity: AI Innovation in Security - Gartner, accessed April 16, 2025, https://www.gartner.com/en/cybersecurity/topics/cybersecurity-and-ai

135. Software Defined Network and AI for Cybersecurity in Healthcare Industry Innovations in Healthcare - SoftCircles, accessed April 16, 2025, https://softcircles.com/blog/network-and-ai-for-cybersecurity-in-healthcare-industry

136. Harnessing AI in the Cloud: Case Studies of Transformative Success, accessed April 16, 2025, https://www.computer.org/publications/tech-news/trends/harnessing-cloud-ai-case-studies

137. How AI is Transforming Data Privacy and Ensuring Compliance - TestingXperts, accessed April 16, 2025, https://www.testingxperts.com/blog/ai-data-privacy-compliance

138. Multicloud Services & Solutions - Insight, accessed April 16, 2025, https://www.insight.com/en_US/what-we-do/expertise/cloud/multicloud.html

139. What is Multicloud? Benefits, Challenges & Strategy - Nutanix, accessed April 16, 2025, https://www.nutanix.com/info/multi-cloud-environment

140. What Is Multi-Cloud Security? Challenges and Best Practices - TechMagic, accessed April 16, 2025, https://www.techmagic.co/blog/multi-cloud-security

141. How to modernize your network in 2025 | NTT DATA Group, accessed April 16, 2025, https://www.nttdata.com/global/en/insights/focus/2025/how-to-modernize-your-network-in-2025

142. Fully-validated and optimized AI High-Performance Storage (HPS) solutions for cloud service partners featuring NVIDIA HGX™ H100 8-GPU based platforms. - DDN, accessed April 16, 2025, https://www.ddn.com/wp-content/uploads/2024/08/FINAL-DDN-NCP-RA-20240626-A3I-X2-Turbo-WITH-NCP-1.1-GA-1.pdf

143. The Importance of Connectivity for AI Applications in Public Safety and Critical Response, accessed April 16, 2025, https://www.thefastmode.com/expert-opinion/39566-the-importance-of-connectivity-for-ai-applications-in-public-safety-and-critical-response

144. Verizon advances its fiber network for AI workloads | News Release, accessed April 16, 2025, https://www.verizon.com/about/news/verizon-advances-its-fiber-network-ai-workloads

145. Windstream Wholesale and Ciena Boost 400G Network Capacity in the Southeast, accessed April 16, 2025, https://www.stocktitan.net/news/CIEN/windstream-wholesale-and-ciena-boost-400g-network-capacity-in-the-12xv3lauwr72.html

146. State of IoT 2024: Number of connected IoT devices growing 13% to 18.8 billion globally, accessed April 16, 2025, https://iot-analytics.com/number-connected-iot-devices/

147. Leveraging Machine Learning and Artificial Intelligence for 5G - CableLabs, accessed April 16, 2025, https://www.cablelabs.com/blog/leveraging-machine-learning-and-artificial-intelligence-for-5g

148. Transforming Telecom Networks to Manage and Optimize AI Workloads - NVIDIA Developer, accessed April 16, 2025, https://developer.nvidia.com/blog/transforming-telecom-networks-to-manage-and-optimize-ai-workloads/

149. Qualcomm Edge AI Box | IoT Edge Computing, accessed April 16, 2025, https://www.qualcomm.com/products/technology/artificial-intelligence/edge-ai-box

150. 5G and Edge Computing: Empowering the Future of IT - ZNetLive, accessed April 16, 2025, https://www.znetlive.com/blog/5g-and-edge-computing-empowering-the-future-of-it/

151. Solving AI Foundational Model Latency with Telco Infrastructure - arXiv, accessed April 16, 2025, https://arxiv.org/html/2504.03708v1

152. Next-Generation Low-Latency Architectures for Real-Time AI-Driven Cloud Services, accessed April 16, 2025, https://www.researchgate.net/publication/387669054_Next-Generation_Low-Latency_Architectures_for_Real-Time_AI-Driven_Cloud_Services

153. Enhanced AI-Native routing for private and service provider WAN: the cornerstone of autonomous networking - Juniper Blogs, accessed April 16, 2025, https://blogs.juniper.net/en-us/ai-native-networking/enhanced-ai-native-routing-for-private-and-service-provider-wan-the-cornerstone-of-autonomous-networking

154. Learning Cache Coherence Traffic for NoC Routing Design - arXiv, accessed April 16, 2025, https://arxiv.org/html/2504.04005v1

155. Framework for Integrating Machine Learning Methods for Path-Aware Source Routing, accessed April 16, 2025, https://arxiv.org/html/2501.04624v1

156. INTELLIGENT ROUTING ALGORITHMS FOR 6G WIRELESS NETWORKS: A REVIEW AND ANALYSIS - International Journal of Advanced Research in Computer Science, accessed April 16, 2025, https://ijarcs.info/index.php/Ijarcs/article/download/7119/5825/15254

157. Evolution of AI Infrastructure From On-Premises to the Cloud and Edge - Gcore, accessed April 16, 2025, https://gcore.com/learning/evolution-of-ai-infrastructure

158. AI Cloud: What, Why, and How? | CNCF, accessed April 16, 2025, https://www.cncf.io/blog/2025/03/06/ai-cloud-what-why-and-how/

159. Optimize Your Network for AI: Bandwidth, Latency & Scalability, accessed April 16, 2025, https://www.networkpoppins.com/blog/optimize-your-network-for-ai-bandwidth-latency-scalability

160. Intelligent Resource Allocation: AI Strategies in Infrastructure Automation, accessed April 16, 2025, https://datahubanalytics.com/intelligent-resource-allocation-ai-strategies-in-infrastructure-automation/

161. 2024 Cloud and AI Business Survey - PwC, accessed April 16, 2025, https://www.pwc.com/us/en/tech-effect/cloud/cloud-ai-business-survey.html

162. Multi-Cloud Challenges with Security and Strategy - F5, accessed April 16, 2025, https://www.f5.com/company/blog/multi-cloud-networking-challenges-and-opportunities

163. Understanding Virtualization: A Comprehensive Guide - CloudOptimo, accessed April 16, 2025, https://www.cloudoptimo.com/blog/understanding-virtualization-a-comprehensive-guide/

164. SOFTWARE-DEFINED NETWORKING CHALLENGES AND RESEARCH OPPORTUNITIES FOR FUTURE INTEREST Santhosh Katragadda, Bukola Hallel - ResearchGate, accessed April 16, 2025, https://www.researchgate.net/profile/Lorenzaj-Harris/publication/383697222_SOFTWARE-DEFINED_NETWORKING_CHALLENGES_AND_RESEARCH_OPPORTUNITIES_FOR_FUTURE_INTEREST/links/66d7859fb1606e24c2df9518/SOFTWARE-DEFINED-NETWORKING-CHALLENGES-AND-RESEARCH-OPPORTUNITIES-FOR-FUTURE-INTEREST.pdf

165. Software Defined Networking (SDN) Advantages - ServerWatch, accessed April 16, 2025, https://www.serverwatch.com/guides/software-defined-networking-advantages/

166. Networking for AI workloads | Nokia, accessed April 16, 2025, https://www.bell-labs.com/data-center-networks/networking-for-ai-workloads/

167. Arm AI Readiness Index, accessed April 16, 2025, https://www.arm.com/-/media/Files/pdf/report/arm-ai-readiness-index-report-part1.pdf?rev=2f8c6d73c3464702ac91cff6c245372f&revision=2f8c6d73-c346-4702-ac91-cff6c245372f

168. Enhancing Communication Networks in the New Era with Artificial Intelligence: Techniques, Applications, and Future Directions - MDPI, accessed April 16, 2025, https://www.mdpi.com/2673-8732/5/1/1

169. Artificial Intelligence (AI) for Network Operations - IETF, accessed April 16, 2025, https://www.ietf.org/archive/id/draft-king-rokui-ainetops-usecases-00.html

170. AI-powered network optimization: Unlocking 5G's potential with Amdocs - Google Cloud, accessed April 16, 2025, https://cloud.google.com/blog/topics/telecommunications/ai-powered-network-optimization-unlocking-5gs-potential-with-amdocs

171. Scaling Bandwidth for AI/HPC Multi-Die Designs with Ethernet, PCIe & UCIe IP - Synopsys, accessed April 16, 2025, https://www.synopsys.com/articles/ai-hpc-multi-die-designs.html

172. SDN & NFV Network Development Solutions - ACL Digital, accessed April 16, 2025, https://www.acldigital.com/industries/communications/network-modernization/sdn-nfv

173. Unveiling the Symphony of Network Orchestration in Telecoms: A Comprehensive Guide, accessed April 16, 2025, https://metavshn.com/unveiling-the-symphony-of-network-orchestration-in-telecoms-a-comprehensive-guide/

174. Exploring the Impact of SDN/NFV on Telecommunications Infrastructure - metavshn, accessed April 16, 2025, https://metavshn.com/exploring-the-impact-of-sdn-nfv-on-telecommunications-infrastructure/

175. Overview of Software Defined Networking (SDN) Risks Background, accessed April 16, 2025, https://hpc.mil/images/hpcdocs/ipv6/Overview-of-Software-Defined-Networking-SDN-Risks.pdf

176. Decarbonization with Network Functions Virtualization - AT&T Business, accessed April 16, 2025, https://www.business.att.com/learn/articles/decarbonization-with-network-functions-virtualization.html

177. The Current State of NFV: Standards | Itential Network Automation, accessed April 16, 2025, https://www.itential.com/blog/company/automation-strategy/the-current-state-of-nfv-standards/

178. The Role of SDN/NFV in the Telecom Industry - METAVSHN all-in-one ISP software, accessed April 16, 2025, https://metavshn.com/the-role-of-sdn-nfv-in-the-telecom-industry/

179. Network Functions Virtualisation (NFV) - ETSI, accessed April 16, 2025, https://www.etsi.org/technologies/nfv

180. 5G - Wikipedia, accessed April 16, 2025, https://en.wikipedia.org/wiki/5G

181. 5G: Understanding the Technology that's Changing Connectivity - TDK, accessed April 16, 2025, https://www.tdk.com/en/tech-mag/past-present-future-tech/what-is-5g-and-why-is-it-important

182. Opportunities and challenges of 5G network technology toward precision medicine - PMC, accessed April 16, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC10651640/