As Physical AI Dawns, the Mobile Network Must Be Reinvented

For the past year, the industry has been preoccupied with a single question: how much smarter can large language models become? Agents, multimodal capabilities, and edge deployment have dominated the AI narrative.

Yet as AI migrates into physical form factors—glasses, automobiles, robots, cameras—and begins to interact with humans on roadways, factory floors, and in homes, a last-mile gap between the real world and large AI models has become apparent.

AI glasses must read text in real time, robotaxis must assess surrounding traffic, and robots must determine their next action. These systems depend on more than algorithms alone—they require image, voice, location, spatial, device-state, and environmental data.

Data must first be captured at the terminal, uploaded via the network, and processed by cloud- or edge-based models. The model’s output must then travel back to the terminal before it can drive real-world action.

This places new demands on mobile networks in the AI era.

Historically, network jitter might have caused a video to buffer for seconds. Going forward, the same instability could derail real-time recognition, delay remote control, or prevent an agent from completing a task.

Mobile networks can no longer function as mere content-delivery conduits. They must identify the connected entity, understand the type of service, assess latency requirements, and prioritize connections when resources are constrained.

In the physical-AI era, the network cannot remain merely a “wider pipe.”

For much of its history, the mobile network’s primary constituency was people.

The voice era saw networks facilitate mobile calls for real-time communication. With 3G and 4G, smartphones and online content proliferated, and the network’s primary mission shifted to high-bandwidth data delivery. Web pages, images, music, video, live streaming, gaming, and online conferencing defined the mobile internet’s signature use cases.

Smartphones, in turn, became highly integrated gateways. Functions once handled by standalone devices—GPS, music players, cameras, voice recorders, and productivity tools—were absorbed into a single handset. A hallmark of the mobile internet era was the consolidation of terminal form factors around the phone, with humans as the principal connected subjects.

AI has upended this premise.

For large models to comprehend the physical world, they require richer and more continuous real-world input. Images, voice, location, motion, spatial relationships, environmental state, and device data will underpin model recognition, reasoning, and feedback.

The network’s service recipients have expanded from “humans” to “agents.” Large models and generative capacity have emerged in the cloud, while the edge now hosts a growing array of AI endpoints—on-device assistants, AI glasses, wearables, cameras, robots, in-vehicle systems, home appliances, industrial terminals, and sensors—all of which may become active network subscribers.

This also reshapes the content networks carry. Historically, mobile networks primarily carried information flows—text, images, video, and app data. In the AI era, networks must also carry data streams tied to reasoning, decision-making, and feedback, imposing stricter requirements on latency, uplink throughput, connection reliability, and resource allocation.

The expansion of connected entities alters connection scale. Tomorrow’s mobile networks will serve not only smartphone users but a vast array of intelligent terminals embedded in cities, homes, transit systems, industrial sites, and public spaces.

Zhao Dong, Vice President of the Wireless Network Product Line at Huawei ICT BG, told Leiphone, “By 2035, the number of deployable agents could reach 900 billion—far surpassing today’s human-centric connection scale.”

The future mobile network, in other words, will confront a system defined by greater device volume, broader geographic distribution, and denser connectivity. Its mandate will extend beyond human browsing experiences to encompass the full cycle of agent perception, reasoning, and feedback.

As service recipients evolve, so too must the capabilities of the mobile network.

Historically, the mobile internet’s dominant traffic direction was downlink. Users opened their phones to pull video, images, music, games, and web pages from the cloud—the network functioned as a content-distribution pipeline. Network quality, in the user’s mind, was measured by download speed, video clarity, streaming smoothness, and gaming responsiveness.

The AI era has fundamentally reshaped traffic patterns. Edge devices are no longer mere download consumers—they must continuously upload images, voice, environmental data, and sensor readings, then await large-model output. The mobile network’s role has shifted from a predominantly unidirectional “information retrieval” link to a closed loop: edge sensing, data upload, cloud inference, and result feedback.

Consequently, communication network requirements now extend beyond downlink bandwidth to encompass uplink throughput, latency, jitter, reliability, and the ability to handle high-concurrency connections.

At MWC Shanghai, Huawei ICT BG distilled the AI era’s demands on mobile networks into three shifts.

First, connection scope expands. Agents will not be confined to densely populated areas. They will be distributed across urban infrastructure, residential devices, industrial sites, transport systems, energy grids, commercial spaces, and public-service settings. Mobile networks must therefore cover more complex environments and serve a broader array of non-handset terminals.

Second, connection density rises. In exhibition halls, malls, stadiums, transit hubs, libraries, and factories, phone users, cameras, AI glasses, robots, sensors, and other terminals may connect concurrently and stream data continuously. Historical network planning based on average human behavior is ill-equipped for such high-concurrency access.

Third, connection-quality requirements tighten. AI workloads are more sensitive to network instability. Historically, degraded network performance mainly compromised content consumption. Going forward, a disrupted intelligent workflow could directly prevent a recognition, control, or execution task from completing. Networks must evolve from offering generic connectivity toward delivering assured, schedulable, and perceptible connections.

Yet today’s mobile networks fall short on all three fronts.

Historically, mobile internet has been downlink-dominated, with network build-out and user experience metrics revolving around download speeds. As AI terminals proliferate, image, video, voice, and sensor data will stream continuously from the edge. Insufficient uplink capacity means agents cannot relay the physical world to cloud models in a timely manner, dragging down subsequent inference and feedback.

Latency and reliability are emerging as new pressure points. Conventional mobile networks deliver respectable average experiences, but AI workloads prioritize deterministic performance at critical moments. Real-time interaction, remote control, robotic coordination, and in-vehicle scenarios cannot be assessed by average throughput alone. A transient jitter spike or a single connection drop can compromise recognition, decision, or control outcomes.

Network complexity is also on the rise. With the stacking of 5G, 5G-A, multiple frequency bands, multiple radio-access technologies, diverse scenarios, and heterogeneous terminal types, operators must balance coverage and capacity while discerning service priorities and guarantee needs. As AI ushers in still more service types and connected entities, traditional manual practices and reactive fault resolution are no longer adequate.

More fundamentally, AI services impose divergent network demands. Some require high uplink throughput, others depend on low latency, still others demand high reliability, and many prioritize connection density or cost efficiency. Networks can no longer allocate resources in a coarse, uniform fashion; scheduling must be driven by service type, scenario priority, and real-time network conditions.

Transforming mobile networks for the AI era, therefore, cannot rely on capacity expansion alone. Adding bandwidth, sites, and spectrum can mitigate some capacity constraints, but it cannot address the structural challenges of service differentiation, real-time resource orchestration, and rising network-wide complexity. Networks require stronger sensing, reasoning, and scheduling capabilities—they must identify divergent service-level requirements and orchestrate coordinated guarantees across base stations, network management systems, and the core.

This is the premise behind Huawei’s AI for Network approach.

Zhao stressed that Huawei views AI not merely as a service that runs atop the network but as a capability that should permeate the network itself—becoming foundational to wireless efficiency, resource optimization, operational simplification, and new-service enablement.

“Bringing AI into wireless communication networks requires addressing three issues: effectiveness, reliability, and cost of introduction. If the cost is prohibitive and the network cannot bear it, AI will ultimately have no role to play.”

Zhao argues that communication networks are critical infrastructure; any new technology integrated into them must first survive a triple test of effectiveness, reliability, and cost. Beyond resource scheduling, channel estimation, beam control, O&M efficiency, and service assurance, AI must satisfy two additional prerequisites: it must be reliable and it must be deployable.

Reliability precludes a system that is opaque, ungovernable, and prone to frequent errors. Deployability means cost and energy consumption must remain within manageable bounds.

Huawei’s strategy is to introduce AI into mobile networks layer by layer: first into base stations and physical pipes to lift network efficiency; then into the network-wide management system to enable autonomous O&M and resource allocation; and ultimately into the core network and carrier business platforms to serve the AI era’s new operating models.

The base station is where this transformation begins.

Applying AI to base stations does not mean converting them into massive compute platforms. Rather, it embeds AI algorithms into resource management, channel estimation, beam control, and other internal communication processes—making the network pipe itself more efficient and energy-conscious while containing costs as it scales to serve more users.

Resource scheduling is a canonical use case. Conventional networks rely on complex mathematical algorithms for scheduling, but as user count, cell density, and service complexity rise, these algorithms face mounting computational pressure. AI enables dynamic resource allocation across broader scopes and finer time grains, equipping the network to manage more users, more cells, and more intricate service mixes concurrently.

Huawei envisions AI enabling tenfold increases in users and cells, and a hundredfold expansion in the dimensionality of problem-solving.

Channel estimation and beam management follow a comparable logic.

Wireless channels are influenced by distance, blockage, reflection, velocity, and environmental dynamics. Conventional measurement entails trade-offs between precision and service overhead. AI can reconstruct a fuller channel characterization from limited sampled data, boosting measurement accuracy without materially affecting service performance.

Building on this, AI can integrate user location, mobility patterns, and service context to anticipate imminent changes, enabling beams to preemptively adjust direction and allocate wireless resources with greater precision.

With more sensors integrated into base stations, the infrastructure gains the ability to perceive surrounding structures, obstructions, reflections, and coverage shifts, and to determine optimal signal direction. For high-rise urban environments, indoor spaces, transit hubs, and complex public venues, such capabilities directly shape network quality.

Beyond individual base stations, Huawei aims to leverage AI for end-to-end network autonomy.

This tackles operators’ mounting O&M burden. Sites are scattered across buildings, basements, mountains, rural areas, transit corridors, and dense urban zones, often demanding extensive field labor. As 5G, 5G-A, multi-band, multi-standard, and multi-terminal environments accumulate, network complexity escalates—legacy manual practices and reactive fault-finding can no longer sustain the scale of tomorrow’s networks.

Network autonomy aims to minimize human intervention as long as site power, transmission links, and hardware remain intact.

AI can continuously monitor network conditions, detect anomalous experiences, forecast congestion and failure risks, and autonomously perform parameter tuning and resource adjustments. Network O&M will progressively transition from “humans hunting for issues” to “the network detecting, diagnosing, and resolving issues.”

Higher up the stack, AI will also penetrate the core network and carrier business platforms.

Historically, the core network primarily handled voice and packet-switched data—a linchpin for connection management and data forwarding. In the AI era, it can further incorporate compute, storage, models, and generative capacity, unlocking new core-network service domains.

Huawei has dubbed this platform Agenty Core, aiming to add an AI plane alongside the core’s existing user and signaling planes, and to introduce AI Service Functions.

AI Service Functions orchestrate capabilities spanning fixed and mobile broadband—including voice, memory, intent comprehension, and experience-management modules. For carriers, this means the core network transcends its traditional roles of connectivity and forwarding to engage in intelligent distribution, experience assurance, and new-service production.

Technology upgrades, of course, must ultimately translate into commercial reality. What roles network-equipment vendors and operators should assume in the AI era is a question Huawei is actively exploring.

For operators, the most immediate opportunity from AI-infused mobile networks still lies in connectivity and traffic.

A growing agent population means network connections will continue to multiply. On-device assistants, AI glasses, robots, in-vehicle systems, cameras, wearables, industrial terminals, and diverse sensors all require persistent network connectivity. They must transmit image, voice, location, environmental, and device-state data to cloud models and receive model output in return. For operators, this translates into more connections, higher density, and greater uplink traffic.

Yet AI’s impact extends beyond traffic growth. Intelligent services impose tiered demands on network capabilities. Routine Q&A, non-real-time interaction, and generic content generation require modest network guarantees and can be treated as ordinary information delivery. Real-time control, embodied intelligence, autonomous driving, and multimodal real-time interaction, by contrast, demand higher performance across speed, latency, synchronization, and reliability dimensions.

Different intelligent services, therefore, command different levels of network value. The more mission-critical, real-time, and integrity-dependent the intelligent workflow, the higher the network assurance tier it demands—and the greater its revenue potential. Operators’ business model can thus evolve from selling bandwidth to selling assured quality of experience.

In real-time interaction, remote control, embodied intelligence, in-vehicle AI, industrial operations, and multimodal collaboration, operators that deliver complete, stable, and synchronized transmission assurance stand to unlock new revenue streams.

Huawei is working with China Mobile and China Telecom to introduce tiered management in 5G-A networks. Basic-tier users receive baseline service; premium-tier users receive deterministic, multiplier-based service. The model emphasizes relative assurance: if a basic user receives only 1 Mbps in a given scenario, a premium user gets double; if the baseline is 1K, the premium allocation doubles accordingly.

Dynamic slicing represents a further capability. Conventional network slicing typically requires pre-provisioned resources and bilateral agreements with OTT providers, posing deployment challenges. Dynamic slicing, by contrast, allocates RB resources on the fly based on network capacity estimates and real-time service status, safeguarding the premium-tier experience.

This enables a new stratified operating model: basic traffic for universal connectivity, multiplier-rate plans for relative experience guarantees, and dedicated dynamic slices for deterministic capabilities serving high-value workloads.

The spread of AI terminals will also reshape operators’ market engagement. Historically, operators expanded their subscriber base through handset subsidies and terminal discounts. Going forward, a more probable model involves bundling devices, connectivity, and services through rate plans, eSIM, network access points, and experience-assurance features.

AI glasses, for instance, could be paired with dedicated uplink plans; robot control, with secure private networks; and in-vehicle AI, with low-latency assurances. Operators would thus move beyond access provision to participate in the design of intelligent terminal use cases.

Core-network AI capabilities will also open new frontiers for traditional telecom services. Voice calls can be augmented with anti-fraud, secure identity verification, private calling, noise cancellation, and voice enhancement—enabling operators to layer higher-grade experience and security atop their existing communication offerings.

Viewed this way, operators are no longer merely “SIM card” distribution channels. They can instead provide connectivity, experience, security, and service assurance—all organized around the use cases of intelligent agents.

This is the AI era’s most consequential shift for mobile networks: it has redefined the fundamental mission of mobile communications.

Historically, the network’s primary purpose was to help humans access information faster and more reliably. Going forward, it must also enable agents to perceive reality, comprehend their environment, reason through tasks, and execute feedback.

Throughout this transformation, AI serves as both a new class of traffic for networks to carry and a transformative instrument for network reinvention.