Mobile Networks Evolution in the AI Era

Introduction .....
Software-based Flexible Network .............

2.1. Network Evolution: from Purpose-built Hardware to Software-based Network
2.2. Flexibility of Software-based Networks

Integration of AI and Mobile Networks .... . 12

3.1. Why Software-based Networks are best suited for AI
3.2. AI-powered Automation
3.3. AI-powered Core and Radio Access Network

Personalized Service-centric Network ...... . 18

4.1. Personalized Network-as-a-Service
4.2. AI-powered Autonomous Service Assurance
4.3. On-Demand Edge AI-as-a-Service

5. Conclusion... 22

Abbreviations

AI Artificial Intelligence
API Application Programming Interface
CAPEX Capital Expenditure
COTS Commercial off-the-shelf
CPU Central Processing Unit
CSR Cell Site Router
C-RAN Centralized Radio Access Network
CU Centralized Unit
D-RAN Distributed Radio Access Network
DU Distributed Unit
E2E End to End
GPU Graphics Processing Unit
KPI Key Performance Indicator
LLM Large Language Model
MIMO Multiple Input, Multiple Output

mmWave Millimeter wave MNO Mobile Network Operator NaaS Network-as-a-Service NF Network Function OPEX Operational Expenditure RAN Radio Access Network SLA Service Level Agreement SoC System on-Chip UE User Equipment UPF User Plane Function vCore virtualized Core network vCSR virtualized Cell Site Router vDU virtualized Digital Unit vRAN virtualized Radio Access Network

Introduction

Since its inception in the 1970s, the mobile industry has undergone a remarkable transformation driven by the aspiration to transmit more data in less time. The evolution began with first-generation (1G) networks, which offered only a few kilobits per second of voice. With each subsequent generation, the capabilities expanded exponentially: 2G ushered in speeds up to hundreds of kilobits per second, while 3G peaked at several megabits per second. The advent of 4G revolutionized access, supporting highspeed data transmission of up to 1 Gbps. Now, with the arrival of 5G, users can experience lightning-fast services exceeding 10 Gbps. This leap in technology enables high-definition video streaming, immersive mobile gaming, and seamless mobile office experiences anytime and anywhere.

What does the future hold for the mobile industry?

Samsung envisions a groundbreaking evolution in mobile networks that will further enhance transmission capacity. These networks will be characterized by their flexibility, automation, intelligence, and have a strong focus on delivering service-oriented solutions. This progressive approach promises to revolutionize how we connect and interact in an increasingly digital world.

2

The future of mobile networks lies in transcending traditional dedicated hardware setups and embracing a transformative shift toward hardware-agnostic software solutions. By decoupling hardware from software, mobile network operators (MNOs) can leverage commercial off-theshelf (COTS) or general-purpose hardware, enabling them to install and run network systems seamlessly without the need for hardware modifications. Many MNOs are already leveraging software-driven core and RAN equipment, and this promising trend will undoubtedly gain momentum with its strengths. In the very near future, a majority of mobile infrastructure, including the core network, RAN, and transport network equipment, will transition to a softwarecentric model, driving innovation and efficiency across the industry.

Mobile networks are on the brink of a transformative shift towards automation, significantly reducing the need for human intervention. As we face increasing complexity in mobile networks—driven by evolving capacity demands and functional advancements, these diverse components must work harmoniously. Network automation streamlines every phase, from initial installation and ongoing operations to practical problem analysis and resolution. This methodology reduces costs and minimizes the likelihood of human errors, resulting in enhanced network quality and stability. Embracing this automated approach is crucial for the future of mobile networks.

3

Mobile networks will become more intelligent through the application of artificial intelligence (AI), taking over tasks traditionally performed by humans or tasks requiring human expertise. For instance, with large language models (LLMs) and intention-based automation, networks can deliver more efficient services based on data-driven insights. Additionally, these systems will facilitate faster and more accurate decision-making by proposing optimized solutions for human consideration. With advanced learning capabilities, AI can minimize network downtime and enhance service continuity by effectively identifying the root causes of problems when they arise.

4

Mobile networks and products will be tailored to be serviceoriented and personalized. Software-based flexible networks enable the creation of customized applications that satisfy more specific needs. When paired with autonomous networks and innovations driven by AI, these systems enhance resource allocation, predict usage patterns, and optimize service level agreements (SLAs). This combination systematically improves service quality and ensures an excellent end user experience.

The evolving landscape and shifts also present new business opportunities for MNOs. These mobile infrastructures, where all network equipment operates through software, rely on general-purpose central processing units (CPUs) and graphics processing units (GPUs). Samsung envisions this innovative infrastructure as essentially transforming mobile networks into a data center and cloud platform capable of handling a variety of workloads, including various AI applications. Mobile communication equipment is located wherever users are, from central data centers to antenna sites, ensuring services are delivered close to users. By leveraging future networks, MNOs can unlock new revenue streams not only from mobile communication services but also from cloud operations [1].

2. Software-based Flexible Networks

The evolution of mobile communication technology has enabled it to meet the increasing user demand for new services – and these advancements wouldn’t be possible without chip innovations. Continuous developments in chip technology have introduced micro-processing, which enabled the implementation of increasingly complex and advanced technologies at lower costs, smaller sizes, and reduced power consumption. Smartphones released in 2010 were equipped with chipsets made using 65nm (nanometers) process technology, while smartphones released in 2025 utilize chipsets based on 3nm process technology, enhancing performance and user experience. Mobile network equipment also has a purpose-built system-onchip (SoC) for processing large-capacity data, and equipment providers are consistently integrating the latest SoCs into their base stations.

However, costs associated with developing new chips have grown exponentially, particularly for high-tech processes. For instance, it cost \$48 million to create a 28nm chip in 2011, whereas the development cost for a 2nm chip now stands at \$725 million. This evolution represents a staggering 15-fold increase in chip development costs over the past decade (see reference Figure 1), which prompted a call for changes in the development of mobile communication equipment.

Moreover, the release cycle for purpose-built SoCs produced by mobile equipment companies typically ranges from three to four years, notably longer than the one to two-year cycle for general-purpose CPUs and GPUs. As a result, it is unclear whether chips designed for specific purposes will remain competitive or continue to be utilized in the near future. In contrast, general-purpose chips already hold significant market share, have a robust ecosystem, and are rapidly advancing. They may serve as a more effective alternative in the mobile communications sector.

source:Arm Inc 2023 IPO prospectus (IBS, july 2022)
Figure 1. Chipset market trend – chip design cost per process node

That said, operating software-based mobile equipment on generic chip-based hardware allows for substantial cost savings compared to developing dedicated chips. There are a number of other advantages that are driving the transition to softwarebased solutions on general-purpose chip platforms:

。 COTS equipment can be utilized regardless of the hardware brand or underlying architecture type. Hardware resources can be used more efficiently, as capacity can be easily expanded or reduced according to market demands. It enables resource sharing with general applications, such as AI, which helps to minimize redundant investments.

2.1 Network Evolution: From Purpose-built Hardware to Software-based Network

Efforts are underway to transition existing communication equipment toward a software-based network across all domains, RAN, core, and transport.

01 Virtualized Core Network

Core network equipment has traditionally handled large-capacity, high-speed packet processing using dedicated hardware specifically designed for the unique attributes of each network entity. However, with the performance improvements observed with general-purpose CPUs and servers, it is now possible to virtualize and implement core network entities. This shift allows for high processing capacity with lower power consumption.

The 5G standard introduces several changes aimed at enabling the
virtualization of the core network and enhancing efficiency: The connection method between different NFs has been
standardized through a servicebased interface that utilizes
HTTP/2, allowing for easier use and
the reuse of services for each NF.

By implementing the network entities within the core network domain – defined as network functions (NFs) – as microservices in a cloud-native architecture, they can be scaled in or out based on the CPU load or the number of user sessions being supported.

User equipment (UE) context can now be stored in a separate database rather than managed individually by each NF. This design enables all NFs to operate without maintaining a consistent state, as they can retrieve and update UE contexts through the database.

As a result of these transformations, the processes and storage for NFs are decoupled, leading to simpler operations and improved scalability.

02 Virtualization of RAN

Traditionally, the RAN relied on
purpose-built chips due to the
complex signal processing
requirements at the physical
layer, which involve sophisticated
timing and high-speed data
processing. Dedicated hardware
was perceived as the only cost
effective and energy-efficient
solution for these operations.
While that may have been
true previously, with the rapid
advancements in general-purpose
CPUs and ongoing efforts to implement
virtualized RANs (vRAN), it has become
feasible to virtualize RAN equipment to
compete with traditional dedicated
hardware solutions [2-4].

In particular, vRAN offers optimal options for various installation environments and requirements, thanks to its flexible and scalable attributes, without the need for additional or modified hardware.

Various versions of 3GPP standards: each MNO must support 4G LTE and 5G NR, depending on the service status, while also maintaining 2G GSM for machine-tomachine services. Furthermore, it should be possible to minimize the need for hardware changes even as technology evolves toward 6G in the future.

Various types of radio units: as multiple input, multiple output (MIMO) technology evolves, multiple types of radio units, ranging from 2T2R radio units to 64T64R massive MIMO units, can be connected.

Various combinations of frequency bands: recent cellular networks are required to support a variety of bandwidths and frequency bands, depending on the MNOs and local environmental conditions. These may include narrow bandwidth at FDD low-band, wide bandwidth at TDD mid-band, and extremely wide bandwidth at millimeter wave (mmWave). Additionally, multiple carriers may need to be supported within each band.

Other configuration options: in 5G, RAN functions
are divided into three components: centralized unit
(CU), distributed unit (DU), and radio unit (RU). With the exception of the RUs, which require separate hardware for handling radio frequencies, each
distributed RAN (D-RAN) can include both CU and DU.
Alternatively, a D-RAN may consist only of a DU, with the CU centralized in a separate data center.

These various combinations of RAN requirements can lead to either excess or insufficient capacity if only limited hardware options are used. In contrast, vRAN can utilize a variety of commercially available COTS servers, allowing for optimal configurations through vRAN software optimization tailored to meet complex needs. Additionally, more powerful and capable general-purpose CPUs designed for communication networks are expected to be released in the future. As a result, vRAN is anticipated to continue to advance further, improving its performance.

03 Disaggregated Transport Network

As other communication equipment evolves into virtualized software on COTS servers, there is a growing momentum to adopt systems that separate hardware and software in transport networks.

In cloud or MNO data centers, large-capacity routers are necessary to connect high-capacity servers and core equipment to the public network and D-RAN sites. These large-capacity routers require numerous physical Ethernet ports and high-performance packet processing capabilities to handle significant network traffic simultaneously.

Consequently, it is practically challenging to replace them solely with COTS servers.

Instead, white box routers can be used as devices that meet these specifications. A white box router combines standard networking chips for high-performance packet processing with a general-purpose CPU for software operation, effectively separating hardware from software. This design allows network managers to control, manage, automate, and optimize the entire transport network from a centralized location using an open network operating system and a software-defined network while supporting large-capacity packet switching.

You can check the video by clicking on the image.

Another transport network device used in mobile networks is the cell site router (CSR) at the D-RAN site. Unlike traditional large-capacity routers, D-RAN sites do not require a high number of Ethernet ports or extensive simultaneous network switching processing. As a result, a virtualized CSR (vCSR) can be implemented using computing resources from a COTS server where a vRAN has already been installed. The introduction of a vCSR at the D-RAN site eliminates the need for a separate hardware CSR, integrating all functions into a single box. This integration is based on the capacity requirements of the D-RAN site and advancements in general-purpose CPU technology, leading to significant savings in both capital expenditures (CAPEX) and operational expenditures (OPEX) [5].

2.2 Flexibility of Software-based Networks

01 Virtualization of Core and Transport Networks and the Role of Edge Computing

The core domain functions—including the control plane function and the user plane function (UPF) for processing high-capacity calls from each RAN site—are located in the MNO’s central data center. This data center also manages RAN sites and handles application processing for services provided directly by the specific MNO. Virtualizing NFs in the central data center offers several advantages, including dynamic resource allocation and adjustment in response to changes in subscriber numbers or traffic volume. This flexibility enables the network to adapt efficiently to new RAN sites or services.

02 Future Networks and Integration of AI

When we think of a future network where services are continuously evolving, edge applications will incorporate AI processing. If the AI operations are sensitive to latency, the AI processes can be added to the virtual NF within the edge data center. Alternatively, if the AI operations require substantial learning from diverse and extensive data collected from multiple edge data centers, the AI processes might be housed as a virtual NF in the central data center. In this case, the computing resources in the edge data center can be repurposed for other functions. As the computing resources of the edge data center evolve and expand in various ways, this development presents MNOs with an opportunity to establish a “distributed edge cloud.”

Figure 2. Software-based network functions and AI workloads running on general-purpose platforms

04

03 Reducing Latency with Edge Data Centers

Shortening the traffic path between users and application servers is essential to support latency-sensitive services. This routing can be achieved by relocating application servers to edge data centers closer to users. Instead of routing traffic through the central data center of a national network, the edge data center can be introduced to substantially reduce the distance between users and service applications. If the RAN sites connected to the edge data center consist solely of the DU and the RU, the edge data center must include several necessary NFs. These include the CU of the RAN domain and the UPF of the core domain.

Since the 5G standard defines the CP and UP of the core domain separately, the edge data center can connect with the edge application server even if it includes only UP NFs without core CP NFs. The advantages of virtualizing all NFs on general-purpose hardware in the edge data center are evident. Being situated between the central data center and the RAN site allows workloads from several NFs to be distributed across each RAN site or concentrated in the central data center based on the available computing resources at each location. Additionally, the required capacity for each NF, including edge applications, can vary due to several factors, such as the initial setup of the edge service, spikes in service subscriptions, the addition of new frequency bands at the RAN sites, or increased application complexity as the edge service evolves. The flexibility in redistributing resources when adjusting the allocation among NFs within the edge data center is significantly greater than that of conventional hardware-based networks. This setup includes the ability to add more resources through additional hardware installations

Virtualization in Distributed RAN (D-RAN) Sites

The NFs of a D-RAN site, composed of various functions within the DU, can be virtualized on the same generalpurpose hardware platform. Depending on the network topology design, the NFs corresponding to the CU may also be included. Additionally, a virtualized NF can also replace the CSR used to connect the D-RAN site to the backhaul network. This structure allows the D-RAN site to be configured using only COTS hardware. In the future, as general-purpose CPUs continue to improve and software optimization advances accordingly, it is anticipated that, in many cases, a single COTS server will be able to support both RAN functions and CSR functions for all of the frequency carriers in a D-RAN site. Furthermore, the remaining computing resources can enable the integration of AI applications within RANs, eliminating the need for separate equipment.

05 Benefits of Virtualization in Private Network Deployment

Benefits of virtualizing the RAN, core, and transport elements in a COTS server can be seen in the establishment of private networks in industrial factories, schools, and enterprise campuses. Unlike national macro networks, these private networks typically consist of dozens to hundreds of cells. Using traditional hardware-based network equipment with fixed configurations and traffic capacity would require the separate manufacturing of dedicated equipment for each NF based on the needed capacity. This approach often results in either unnecessary large-scale equipment or additional expenses. In contrast, a virtualized network can operate with just one COTS server to support all components, including vRAN, vCore, and vCSR, even for extremely small capacities. Additionally, to meet specialized service requirements, it’s possible to create a customized configuration that increases the capacity of specific NFs in certain domains while maintaining overall flexibility.

3. Integration of AI andMobile Networks

Like other industries, the mobile network business has recently utilized AI to conduct research in various fields. These investigations aim to enhance operator convenience, user experience, and typical network performance, such as data speed.

This section discusses the relationship between AI and mobile networks, primarily focusing on two major questions:

What advantages and outcomes might result from combining AI and mobile networks?

What preparations
should be made in
advance to support
various AI applications in
the future?

You can check the video by clicking on the image.

3.1 Why Software-basedNetworks are best suited for AI

For AI technology to be effectively utilized,three key components are necessary: computing resources, data, and algorithms.

When applying AI within a mobile network—excluding algorithms specific to the AI application’s functions—securing the computing resources required for AI operations within the network system is essential. Additionally, significant amounts of relevant network data must be collected and provided to AI applications for model training and inference.

First, we examine the issue of securing network data. To fully leverage the potential of AI, it is essential to collect various types of network data so that AI applications can learn effectively. In traditional monolithic hardware environments, all network entities must report entity-level network data through a standard-defined interface, regardless of configuration. These network entities, which may take the form of dedicated hardware equipment, can be broken down into internal layers, including hardware devices, operating systems, middleware, and software functions.

However, the exact structure and status of these internal layers are primarily accessible only to the equipment provider. This segmentation might create a limitation for MNOs, who must rely on the solutions offered by the equipment provider to gather network data necessary for assessing the system’s normalcy or identifying issues at this level. Consequently, the MNOs’ ability to determine the normal state of the network system depends heavily on the information processed and selected by the equipment provider. Moreover, there are challenges related to the collection and monitoring of new network data, as well as the ability to perform new analyses and optimizations.

Conversely, in software-based networks constructed using COTS general-purpose servers and open software platforms, network entities are created by integrating technologies across various layers, including operating systems and vir tual ized or containerized NFs. Since different providers often supply the hardware and software for each layer, the configuration and status information for these layers is accessible to the MNO using the network entities. This access allows the MNO to collect and integrate detailed network data from all entities, functions, and their respective layers. Please refer to below Figure 3.

Figure 3. Unified network data collection of software-based network

The data gathered at this level not only enhances the understanding of each network entity, which may include multiple network functions, but also contributes to the creation of big data for the entire network. By comparing historical and current network data or analyzing the relationships between various data points, the interconnections between different network entities can be understood more clearly and in greater depth.

Given the extensive and intricate interrelationships represented in this big network data, there is an increasing need for AI applications to assist in analyzing these associations. From the perspective of network operations, AI can be employed to predict potential issues that may be challenging for MNOs to detect directly, as well as to identify risks associated with the interworking of network functions.

AI applications require complex processing of large datasets and understanding the semantic relationships between them. If MNOs are required to allocate dedicated computing resources, such as separate CPUs and GPUs exclusively for AI, it could significantly increase CAPEX and create substantial entry barriers for the initial implementation of AI technology in networks. In software-defined networks, mobile network workloads and AI workloads can share the same hardware, simplifying the introduction and expansion of AI functionalities. AI workloads can be distributed from central data centers to edge data centers and D-RAN sites. The decision on where to place the AI workload depends on the functional attributes of the AI application, including latency requirements, the size of the associated network data, and the target application for the analysis results.

Another important consideration is how to allocate the available computing resources at each location between network functions and AI workloads. For instance, a central data center with the highest computing capacity can handle AI model training and management, while edge data centers or D-RAN sites can receive the trained AI models and focus on specific AI tasks to optimize resource utilization.

Additionally, software-defined networks can centralize the management of distributed computing resources across all these locations for both network and AI workloads.

3.2 AI-powered Automation

Operating a mobile network requires deep expertise across multiple domains. As the network evolves, it is divided into various services, such as IoT services, latency-sensitive services, fixed wireless access, and traditional voice and data. In the past, experts managed and operated each domain individually, focusing on the equipment specific to that area. However, there is now a need for service-level operations encompassing the entire network. MNOs must continuously monitor numerous cells, mobile network equipment, subscribers, and their connections to assess normal operating conditions. This monitoring enables them to quickly detect and recover from serious issues, such as equipment failures or service disruptions. Maintaining and managing such a large-scale mobile network demands significant time, staffing, and financial resources.

Network automation has emerged as a key technology in the telecommunications industry in recent years. Automating network operations significantly reduces both management time and costs for MNOs. Beyond just lowering operational expenses, efficiency is enhanced by decreasing the time required to resolve complex network issues. Additionally, the risk of human error is minimized by reducing the reliance on subjective decision-making. A stable automated network can be managed by nonspecialized personnel, making it more accessible. Routine and repetitive tasks essential for network management can be automated and performed periodically, which helps to maintain network stability. As a result, these advancements improve and ensure the network’s reliability.

AI-based network automation can offer recommended actions, cost-effective analysis, and valuable insights. By integrating not only flat relationships between collected network data but also multi-dimensional relationships, AI models can uncover hidden patterns and meanings that are useful for detecting and resolving problems. Therefore, it is crucial to establish a comprehensive data set that reflects the operational context from a network management perspective. This extensive data set enhances visibility into network management, enabling analysis tools and AI applications to understand the fundamental operating contexts more profoundly and extract insights. Software-defined networks eliminate silos created by hardware dependence and provide integrated visibility and management points, which facilitate AI-based analysis, interference resolution, and network control.

The typical use cases are outlined in Table 1 below. In addition to continuous end-to-end data monitoring, AI can detect abnormalities in advance and analyze their root causes, significantly reducing the average repair time. When a network issue arises, traditional recovery methods often require domain experts to be dispatched based on the identified cause and the type of customer complaint. However, AI-enhanced recovery technologies utilize insights from AI to address network problems, events, and performance more effectively. Network optimization can be improved by incorporating individual AI models for tasks such as prediction and cause analysis. Furthermore, generative AI technologies, including LLMs, can support MNOs in their network operations.

Table 1. AI example use cases for autonomous network

Category	Use case	Description
Planning	Cell capacity andcongestion forecast	Traffic demand is predicted at the site level andforecasts when congestion occurs given the currently deployed sites and carriers.
Deployment	Service-awareresource provisioning	Network resources are allocated to meet the service quality requirements and KPl targets.
Operation	Anomaly detection	Anomalous patterns in KPl changes aredetected, which would be challenging to identifythrough manual analysis.
	Alarm correlation andsuppression	Massive alarms are collected and filteredout based on the relation of each alarm andcausality analysis. Operators can focus just themost critical alarm on the incident at hand.
	Service continuity	Al predicts the age of network equipment andcomponents and proactively recommendsreplacements to avoid service breakdown.
Optimization	Energy saving	Al predicts the transient traffic demands and required network resources to find the best cell on/off policy, potentially reducing energy consumption while maintaining service quality.
	RAN speedoptimization	RAN parameter optimization improves thetarget KPl, such as downlink throughput, withoutrequiring manual analysis or configuration.
	RAN digital twin	"What-if" analysis is available using RAN performance emulation - caled RAN digital twin. MNOs can decide when existing equipment should be upgraded to manage the increasingtraffic demands.
Convenience	Network Copilot	The Al assistant aids customers via naturallanguage-based chat, offering briefings,summaries, step-by-step guides,explanations,and answering questions.
Convenience	Network Autopilot	Multi-Al agents cooperate to derive decisionsfor both operations and optimization. Each Al agent takes the role of an expert in networkmanagement. Through a step-by-stepconversational procedure, human operators can follow up on the history of decision inference.

3.3 AI-powered Core and Radio Access Network

In software-based networks, AI functionalities can be seamlessly extended across all network domains, including the RAN and core network, as well as automation processes.AI workloads and applications can be efficiently deployed wherever needed, from central data centers for network management to edge data centers for latency-sensitive AI applications and even D-RAN sites.

Performance improvement and optimization can be achieved by applying AI functions to the core network and RAN. However, unlike central data centers, core and RAN sites often have limited computing infrastructure available for AI workloads. If computing resource bottlenecks occur, it could lead to poor performance or service interruptions across the entire network. Therefore, it is essential to manage and distribute computing resources flexibly and efficiently, taking into account the features of various use cases, to enhance performance through AI. Legacy purpose-built hardware-based networks have been restrictive in developing and integrating new AI functions due to their rigidity. In contrast, software-based networks facilitate the application of AI functionalities across all network domains and seamlessly support the expansion and sharing of computing resources more flexibly via open hardware ecosystems and virtualized functions. AI workloads and applications can be efficiently deployed wherever needed, from core networks to D-RAN sites.

Figure 4. Examples of Samsung AI applications for automation

In the core network,

AI can analyze user mobility patterns and apply this information to optimize the paging operation. By doing so, the number of paging requests sent to the RAN can be reduced, increasing the paging success rate and minimizing unnecessary use of wireless resources for broadcasting messages over the air. AI also enables effective load balancing among multiple user plane instances by predicting traffic usage and patterns to optimize the distribution of workloads across available resources. This capability allows for preventative measures against possible overloading scenarios involving the UPF.

Figure 5. Examples of Samsung AI applications for vCore

At cell sites,

AI has the potential to enhance channel estimation on the physical layer, improving receiver performance under weak RF conditions. AI functions can capture downlink channel characteristics such as line-of-sight, Doppler spread, and frequency selectivity from PMI feedback or SRS response. This enables the base station to configure the most suitable reference signal type and select the optimal MIMO mode, ultimately increasing user throughput. A wide range of studies is underway to leverage AI for enhancing the L2 operations of next-generation systems, as well as the physical layer. One example is AI-based link adaptation, which helps determine the optimal MCS levels for each user, resulting in maximized user experience [6-7].

Figure 6. Examples of Samsung AI applications for edge data centers

4. PersonalizedService-centric Network

The evolution of mobile networks is shifting from the traditional concept of a common public network to more personalized and dynamic service networks.

Currently,

the services offered on public networks primarily consist of connection services that support internet access and video streaming. These services are provided to all subscribers under a standardized SLA, although there may be variations in SLA terms and conditions based on service plans. This approach can be described as a “one-size-fits-all” model, where network resources such as bandwidth, data speed, and services are uniform for all subscribers. While this model simplifies network management and ensures basic service levels, it often falls short of meeting the diverse SLA needs and expectations of various consumers and businesses in the future.

The emergence of personalized networks

Figure 7. Network trends towards personalized service-centric network

represents a shift from traditional, connection-based public networks to service-centric models. These models enable the dynamic creation of a virtualized service network tailored to individual needs and SLAs. This transition can be achieved through advanced technologies, such as containerized and virtualized NFs operating on a universal computing platform. This evolution allows equipment providers to move beyond their proprietary solutions and eliminate equipment vendor dependency in creating a virtualized service network. As a result, MNOs can create flexible service networks anytime and anywhere using these technologies, enabling them to provide customized SLAs for various user segments.

4.1 Personalized Network-as-a-Service

Network slicing is a technology that enables MNOs to create virtualized networks within a single physical network, allowing them to provide value-added end-to-end (E2E) services. Using network slicing technology, MNOs can offer more personalized and customized services specific to the request applications or SLAs, aligning with the trend of increasingly diverse customer groups. On the other hand, through the network-as-a-service (NaaS) application programming interface (API), MNOs’ enterprise and government customers can achieve operational benefits similar to those of having their own dedicated networks. In this setup, enterprise and government customers can delegate the management of virtualized service networks to MNOs and focus on the service quality received from the virtualized service networks.

Figure 8. Dynamic creation of personalized services anytime and anywhere

In an area within 2 kilometers of the factory in Munich, allow 200 connections with a maximum delay of 50 milliseconds and a transmission rate of 50 Mbps.

As illustrated in Figure 8, when a customer initiates the creation of a virtualized service network, this process involves placing a service creation order on the MNO with simple service intention.

The MNO system then translates these service orders into a virtualized service network creation procedure on the physical network by utilizing network slicing technology.

This conversion process may become even more simplified and human-friendly in the future, thanks to the implementation of an AI-powered intent-handling engine or an LLM. Ultimately, the MNO can customize the virtualized service network according to the specific needs of the customer, such as area, usage duration, and SLAs. This tailored approach explains the provision of personalized network services. To evolve into future networks, MNOs are transforming cell sites and communication data centers into distributed telco edges and AI infrastructures, enhancing user experience and supporting AI workloads by leveraging proximity to users. This conversion enables real-time interactions with nearby edge and AI sites, significantly improving service experience. By placing data processing closer to end-users, edge computers deliver faster response times for applications, particularly beneficial for real-time services such as online gaming, video conferencing, and augmented reality.

4.2 AI-powered AutonomousService Assurance

MNOs face growing demand for personalized virtualized network services and diversification of various virtual network services, which implies that MNOs will soon encounter limitations that arise from an increasing number of these services. This is particularly true when MNOs create virtualized network services and adhere to the manual or rule-based operations only. Ensuring SLAs despite numerous service failures and performance degradation issues becomes a critical blocker to the mass deployment of services. Finally, the data associated with the massive numbers of network resource usages of users, and network quality from the multitude of virtual network services are expected to increase exponentially. To effectively manage this complexity, it is essential to have simple and intelligent ways of automating the virtualized service network management and SLA guarantees with AI.

AI-powered virtualized service network management can predict resource usage and achievable SLA levels for individual services and identify which resources and parameters are crucial for ensuring SLAs. The AI-powered system becomes more intelligent and advanced, considering both the current and future states of the network. Figure 9 illustrates examples of how AI can enhance virtualized service network management in terms of usage prediction, SLA forecasting, and resource provisioning. By integrating such AI capabilities, MNOs can offer a more intelligent and autonomous service guarantee system, leading to more efficient resource utilization and improved SLA guarantees for more diversified personalized virtual service networks.

Figure 9. AI-powered autonomous service assurance architecture

AI-powered prediction of resource usage and SLA: by analyzing traffic volume and usage trends, AI can proactively adjust resources, either expanding or reducing them as necessary, to ensure optimal performance. Additionally, AI continuously monitors the system’s status, automatically recovering from errors by reallocating resources or rerouting traffic, thereby maintaining service continuity. AI can identify correlations and patterns between traffic and overall service performance by tracking key SLAs such as delay time, transmission rate, and error rate.

AI-powered resource provisioning: intelligent decisionmaking is conducted based on current, past, and predicted network usage in relation to the target SLA. AI determines optimal parameter values and suggests the best subsequent actions to take. By utilizing historical data collected from the network, AI can anticipate potential performance issues before they arise and proactively provide automatic recommendations to the network.

4.3 On-demand Edge AI-as-a-Service

Transition to Distributed Edge AI Infrastructure

MNOs are working to transform their cell site locations and communication data centers for future networks into distributed telco edge and AI infrastructures, aimed to enhance user experience and support AI workloads. Unlike traditional AI infrastructure that relies on large data centers, this approach focuses on developing distributed “edge AI infrastructure” utilizing the existing communication resources in each MNO region. By converting their current distributed communication infrastructure into edge AI infrastructure, MNOs can significantly improve service quality by combining AI and edge computing. One of the innovative services that can be offered in the future distributed edge AI infrastructure is on-demand edge AI-as-a-Service. This service enables the dynamic creation of AI applications or services (referred to as AI workloads) over distributed CPU and GPU computing resources available within the mobile network. MNOs can generate distributed AI workloads even where CPU and GPU resources are physically spread across different sites and edge data centers.

Service Management and Orchestration for AI-as-a-Service

To enable the on-demand edge AI-as-a-Service, it is essential to develop a new service management and orchestration capability for creating and deploying AI services. This functionality requires adequate inventory and topology management of the CPU and GPU computing resources within the edge AI infrastructure. When an AI service is requested, the necessary resources are automatically allocated in both the CPU and GPU layers to meet the SLA requirements of edge AI-as-a-Service. This resource allocation is determined considering the inventory and topology information related to the available CPU and GPU resources and communication delays between the nodes. It also analyzes post-deployment performance and adjusts or relocates resources as necessary to dynamically support various AI services by observing the actual traffic demand for AI services.

5. Conclusion

Samsung’s forward-looking vision in network and AI centers around two key concepts: Network for AI and AI for Network.

To pave the way for a network infrastructure designed to better serve AI—referred to as Network for AI—Samsung extends its successful vRAN model and experience across the entire network [8-10]. This end-to-end network virtualization approach enables operators to build a multipurpose platform capable of accommodating diverse workloads, including AI applications and services, which are expected to grow exponentially. Without the need for separate servers or platforms, this approach enables operators to seamlessly deploy CPUs and GPUs as required for AI at any layer of the network, from central offices to edge locations and cell sites, ensuring AI capabilities are available wherever they are needed. Additionally, this multipurpose and distributed edge platform, spanning millions of locations, enables operators to efficiently leverage location data—a critical advantage and catalyst for the next big leap in the AI era.

Furthermore, an end-to-end software-based network serves as the optimal environment to maximize intelligent and AIpowered connectivity. With a data center-like architecture featuring uniform software running on COTS servers, a software-based network simplifies intelligent automation, reduces operational burdens (OPEX), and creates new revenue streams. This positions operators for long-term competitiveness in the evolving AI landscape.

AI for Network focuses on utilizing artificial intelligence to enhance network performance and efficiency. As an example, Samsung’s CognitiV Network Operations Suite (CognitiV NOS) [11] provides a diverse library of AI-powered applications designed for end-to-end network automation, covering planning, installation, operation, and optimization phases. By streamlining manual operations with automation, the cost of network management is significantly reduced, enabling a shift toward more advanced network models. End-to-end network automation is key to evolving from traditional common public networks to personalized and dynamic service networks aimed at driving revenue increase. AI-powered functionalities spanning L1/L2, L3, and core network layers will further enhance performance, delivering seamless connectivity and improved user experiences.

Samsung’s strategy encompasses AI integration across the entire network, from core to edge, empowering operators to build more intelligent and efficient networks. By harmoniously integrating Network and AI, Samsung is not only leading but also shaping the future of mobile communication networks, positioning them as data-driven and truly AI-ready next-generation systems.

References

[1] GSMA M360 APAC Seoul 2024 Samsung Keynote Speech: AI & Telco, https://www.youtube.com/watch?v=oXNwiYE03BM
[2] “Samsung’s Fully Virtualized RAN Ensures Flexibility, Scalability and Reliability”, https://tinyurl.com/2dmba6q4
[3] “Virtualized RAN-Vol.2”, https://tinyurl.com/26v9gesk
[4] “vRAN makes future upgrades easy”, https://tinyurl.com/27jvqjz4
[5] “From Boxes to Bits: The Evolution of the Virtualized Network”, https://tinyurl.com/28gl5262
[6] “From AI Concept to Real World: AI-powered Modem for Next Samsung Cellular Networks”, https://research.samsung.com/
blog/From-AI-Concept-to-Real-World-AI-powered-Modem-for-Next-Samsung-Cellular-Networks#none
[7] “AI-Driven Modem Innovations for Shaping the Future RAN”, https://research.samsung.com/blog/-Blog-AI-Driven-Modem
Innovations-for-Shaping-the-Future-RAN
[8] “Samsung and Verizon Charge Ahead with vRAN”, https://www.samsung.com/global/business/networks/insights/
blog/1209-samsung-and-verizon-charge-ahead-with-vran/
[9] “Verizon starts C-Band equipment deployment”, https://www.samsung.com/global/business/networks/insights/press
release/0419-verizon-starts-c-band-equipment-deployment/
[10] “Open up new business opportunities with Samsung’s vCore”, https://www.youtube.com/watch?v=lzzug1S9osk
[11] “CognitiV NOS: Harnessing AI for Autonomous Networks”, https://www.samsung.com/global/business/networks/cloud
network/intelligent-autonomous-network/

About Samsung Electronics Co., Ltd.

Samsung inspires the world and shapes the future with transformative ideas and technologies. The company is redefining the worlds of TVs, smartphones, wearable devices, tablets, digital appliances, network systems, and memory, system LSI, foundry and LED solutions.

ⓒ 2025 Samsung Electronics Co., Ltd.

All rights reserved. Information in this leaflet is proprietary to Samsung Electronics Co., Ltd. and is subject to change without notice. No information contained here may be copied, translated, transcribed or duplicated by any form without the prior written consent of Samsung Electronics.

2025人工智能时代的移动网络演进技术白皮书.pdf免费PDF下载