过去150年，我们一直连接着人类的智能。如今，我们正在赋予网络以智能，使其能够自主行动。
四十余年来，网络性能一直是移动技术演进的主要衡量标准，并将在未来持续发挥作用。然而，单纯的性能已不再是衡量网络能力的唯一标尺。智能网络底座代表着网络演进的下一次根本性跃迁。作为一类在设计之初即具备可编程、自主与AI原生属性的基础设施，它将为网络奠定基础，使网络不仅能够传输数据，还能够在实时环境中进行感知、预测和行动。
每一代移动通信技术都持续拓展着“连接的人类智能”能够做到的事情。爱立信率先推出早期交换系统，使大规模通信路由成为可能，并在20世纪70年代中期进入数字交换时代，在80年代前后开启移动通信时代。如今，5G已在全球180个国家商用部署，拥有超过30亿签约用户，而6G也已进入视野，我们正在迈入一个“连接智能”的新时代。
智能网络底座不仅是移动网络架构的简单延伸，更代表着一次重要跨越：它将演变为通用的AI网络框架，连接并无缝整合分布式AI系统、云服务、终端、传感器以及遍布网络、城市和企业的边缘算力。借助这一底座，AI系统能够持续交互、协同学习，并以机器速率进行运行，同时在先进的5G与未来6G连接能力支持下，保证安全性、可靠性以及可保障的性能。
通过强大的无线网络与云/数据中心以及边缘计算的结合，再叠加全面自治的网络能力与健全的数据治理机制，智能网络底座将使网络能够在数十亿并发交互之间实时进行感知、预测和行动——支持的应用范围从数字孪生与实时翻译，到自动驾驶和任务关键型应急通信服务。在这一过程中，智能网络底座将从根本上重塑网络的作用，使AI能够在真正的全球尺度上改变家庭、城市、企业和行业。
值得信赖的技术栈
在AI时代，数字基础设施的快速演进——涵盖算力、存储、网络以及半导体——是任何具有竞争力的企业和国家的共同基石。这一基础设施必须作为一个整体系统协同运作，形成值得信赖的技术栈。
在这一技术栈中，电力生产与电网、半导体、频谱资源以及包含稀土在内的原材料，共同构成共享的关键基础。技术栈将建立在开放系统与开放接口之上，把各个域连接起来。这种开放而模块化的方法能够让企业在各自领域内展开竞争，同时在系统层面开展协作，从而在全栈范围实现全球规模效应。
在这一栈中，连接本身正演进为智能网络底座：与云/数据中心域和终端域一起，为AI模型层提供所需能力——既包括集中部署在数据中心的能力，也包括在全国范围内分布于先进连接与终端中的能力。这些能力将被下一代用户所使用的应用所消费，并通过先进的连接能力，为大规模AI应用提供安全性、性能与效率保障。
技术栈的设计原则包括：开放性、可信度与灵活性。开放性确保创新不受接入门槛限制，任何开发者、企业或政府都能够在其之上构建、竞争与发展；可信度确保安全与韧性，能够在威胁环境不断演变的情况下保持稳健；灵活性则使技术栈具备足够的适应力，为AI世界提供可保障的性能与差异化服务。
在应用层面，智能网络底座将向消费者、企业与政府提供差异化服务，通过开放网络API对外暴露网络能力，并通过自治运维进行管理，成为开发者与创新者打造下一代AI原生应用的平台。我们正在围绕开放性、可信度与灵活性三个核心原则设计未来网络，将其作为技术栈的底层属性，引导网络向智能网络底座演进。
我们也正在为实现完全自治的网络进行架构设计：网络将具备完整的自治能力，高级服务可以通过意图而非具体指令进行管理，网络将在实时环境中以人类运维团队无法企及的规模与速度实现自优化。与此相伴的是能效的核心地位。随着AI在全球范围内扩展，“智能”的成本将成为一项战略变量：在每次交易上实现最低能耗成本的网络，将获得结构性优势，并随着时间推移持续累积。
What is forming today and what your network must be ready for
State of play 2026
The next era of connectivity is already taking shape. From best-effort consumer services to AI-native, mission-criticaland sovereign applications, the demand profile of mobile networks is shifting structurally and irreversibly. The four trends examined in this article reflect what we see forming, and what we are designing to meet it.
We expect agentic AI to continue its strong trajectory into new categories
Most discussions about AI traffic focus on new generative demands. However, the real shift is behavioral. We are adding a new traffic category to networks optimized for human consumption – and AI-driven applications interact continuously, not intermittently. Agentic AI traffic is a fraction of overall mobile broadband volume today, but it is growing at a rate that makes its future network impact unpredictable. The shift from centralized AI models to distributed autonomous agents will place fundamentally different demands on uplink capacity and latency. We are designing for those demands now, not when they arrive.
Mobile is becoming the primary access layer for AI-native services, and the implications go beyond throughput
Smart wearables, personal agents and multimodal AI applications are converging on mobile as their primary connectivity layer, driven not by today’s traffic volumes but by the high and continuous latency demands they will generate as adoption scales. Agents that must act in the world in real time require guaranteed, low-latency connectivity as a baseline condition, not a premium option.
A new class of network demand is forming and the window to design ahead of it is now
As autonomous AI agents scale across manufacturing, logistics, health care and public services, they will generate a distinct class of network demands that are continuous, latency-demanding and uplink-heavy, and they will require verified identity at machine speed. These demands do not yet dominate network traffic, but the architecture required to serve these takes years to build. The decisions made now will therefore determine whether networks are ready for these new demands when needed.
Post-quantum cryptography is standardized, available and ready to deploy
Large-scale quantum computing will eventually break today’s public-key cryptography, and adversaries are already harvesting encrypted data in anticipation of that capability. NIST (The National Institute of Standards and Technology) published its first three quantum-resistant standards in 2024, and Ericsson is driving adoption across the 3GPP (the 3rd Generation Partnership Project), O-RAN (the Open RAN Alliance) and the GSMA (GSM Association), with full postquantum cryptography (PQC) support confirmed for 6G from its first release. Quantum resilience must now be reframed as an infrastructural commitment, not a compliance exercise, with the need for communication service providers (CSPs) to ensure migration by 2030 at the latest.
2026 telecom industry trends shaping the intelligent fabric
In this CTO Technology Trends article, I examine the four technology trends that together define the intelligent fabric: what drives its adoption, what secures its foundation, whatextends its perception, and what makes it computationally viable at groundbreaking scale.
Trend 1: Physical AI
The network’s primary user is changing. Alongside consumers and enterprises, autonomous agents are becoming significant actors on the network: generating traffic, consuming capabilities and making decisions at machine speed. This is giving rise to a new commercial reality and a new set of demands that existing network architecture must evolve to meet.
Trend 2: Security in transformation
Trust is not a feature of the intelligent fabric; it is the condition on which everything else depends. The security architecture of next-generation networks must be built to protect today’s data from threats that do not yet exist. AI is transforming how vulnerabilities are found and exploited. Quantum computing will eventually break the public-key cryptography that today’s networks rely on. Both demand a proactive response and building quantum-resilient, AI-native security into the intelligent fabric is our industry’s collective imperative.
Trend 3: The programmable world
The world’s networks are becoming a programmable platform. Through open APIs, integrated sensing and cognitive digital twins, the network is evolving into a platform that perceives, exposes and acts on the physical world. For developers and enterprises, this means building directly on network capabilities at scale. For industries, it means a real-time model of the physical world that adapts as conditions change.
Trend 4: Heterogeneous compute
The intelligent fabric will run on a heterogeneous compute architecture, an optimized combination of purpose-built and general-purpose accelerators, each deployed where it delivers the greatest efficiency. Ericsson’s custom silicon is optimized specifically for network processing tasks and AI inference, ensuring that the most demanding workloads run with the performance, energy efficiency and cost structure that make global-scale deployments viable.
Trend 1: Physical AI
Physical AI is reshaping the network’s primary user: From people consuming data to autonomous agents sensing, deciding and acting at machine speed
The new dynamic
Exponential agentic growth will generate demand that existing networks were not designed to serve.
Alongside business-to-consumer and business-to-business, a third commercial relationship has emerged: business-to-agent, where software systems buy services, negotiate capabilities and execute tasks autonomously on behalf of humans. The growth of agentic AI traffic has been exponential. In this new business-to-agent economy, competitive position is not secured by contract. It is earned continuously, as agents evaluate and switch based on performance, capability and cost at machine speed.
The next wave is already visible: AI-enabled wearables such as smart glasses place unrelenting demands on networks with the ability to continuously sense surroundings, transmit contextual data to cloud or edge models and act on behalf of users in real time, at any place.
Physical AI is accelerating this transition further. In 2026, we expect to see the first commercial humanoid deployments operating in controlled industrial settings, performing structured, repetitive tasks in logistics and manufacturing environments, with broader use in residential and urban outdoor environments remaining experimental. These systems offload critical algorithms, including real-time vision and control, to 5G edge cloud infrastructure, with onboard silicon augmented by frontier capabilities delivered over the network. What is genuinely new is not their physical capability but their perception. Computer vision, depth sensing and AI reasoning let these systems interpret their environment in real time, enabling autonomous behavior that fixed automation cannot match. Once an agent learns something new, every connected agent in the world has access to that knowledge. In doing so, the network will serve as the shared nervous system of a distributed, continuously self-improving system – and in that sense, an operating system for the physical world, abstracting the complexity of sensors, compute and AI into a single programmable layer.
Architectural trends
The networks will evolve into an intelligent fabric that senses, decides and acts at machine speed.
Continuous sensing by agents and devices is reshaping traffic patterns, increasing the strategic importance of uplink capacity, low round-trip latency, edge-cloud integration and, critically, security and trust. Networks will evolve into platforms where agents are authenticated, authorized and granted differentiated treatment, with CSP capabilities exposed through standardized, machine-readable interfaces. As a founding member of the Linux Foundation’s Agentic AI Foundation, Ericsson is working to enable autonomous agents to interoperate securely across vendors, industries and national boundaries, building the open and trusted foundation the agent economy requires.
Next steps for the industry
Communication service providers that are first to make their networks available to machines will gain a significant edge in the agent economy.
Agents will increasingly write their own integration code against CSP APIs, discovering and invoking network capabilities autonomously, with no human engineer in the loop.
At Ericsson, we are building the open, standardized interfaces that make this possible. The CSPs whose networks for machine-to-machine transactions first will set the commercial terms on which the emerging agent economy will run.
Trend 2: Security in transformation
Building quantum-resilient, AI-native security into the intelligent fabric is our industry’s collective imperative
The new dynamic
AI is reshaping the security landscape across multiple dimensions: accelerating both threat detection and exploitation, introducing new challenges around agent identity and access governance, and demanding a fundamentally broader security response. At the same time, large-scale quantum computing is expected to break today’s widely deployed public-key cryptography. Both require a proactive industry response.
AI is simultaneously the most powerful tool available for network defense and the source of a fundamentally new class of threat.
AI-powered systems can now automate vulnerability discovery, detect previously unknown zero-day threats and generate remediation at machine speed. As the window between threat discovery and exploitation continues to narrow, these capabilities are becoming a baseline requirement for critical infrastructure. The challenge on the threat side is equally significant. As agentic AI proliferates across enterprise environments, it introduces attack surfaces that traditional security models were not built to address. Agents authorized to access systems and execute decisions autonomously require a rethink of authorization, intent verification and action auditing across the full operational chain.
The cryptographic challenge for our industry is structural and long-range. Telecom networks are long-lived, globally interconnected and carry the communications of critical services. Adversaries are already harvesting encrypted data today in anticipation of the quantum capability that will eventually make today’s public-key algorithms obsolete.
Architectural trends
We are developing next-generation network architecture on a foundation of AI-native security and quantum-resistant cryptography, across every deployment model and product life-cycle phase.
Security is built into all Ericsson network products from the design stage, with AI-based automation and assurance integrated to enable real-time response to both known and emerging threats. To address the cryptographic transition, Ericsson has been an active contributor to the global PQC standardization process and is driving adoption across the 3GPP, O-RAN and the GSMA. Full PQC support is confirmed for 6G from its first release.
Next steps for the industry
The migration timelines are defined, the standards are in place and the tools exist. The decisions made in this period will determine the security posture of the networks on which the intelligent fabric runs.
Agentic AI is already proliferating across enterprise environments, and the threat landscape it introduces cannot be addressed by traditional security models. The proactivity of CSPs, vendors, enterprises and governments in addressing this very real threat will be critical: reevaluating authorization, intent verification and action auditing before agentic deployments scale, not in response to the vulnerabilities they expose.
On PQC, high-priority systems should complete migration to quantum-resistant algorithms by 2031, with full transition by 2035. The geopolitical context demands earlier action on the most critical deployments. I expect PQC adoption to begin in the evolution of 5G and to be a baseline requirement in first-generation 6G systems. The networks carrying the world’s most sensitive communications must be quantum-resilient before quantum computing capability matures.
Trend 3: The programmable world
The intelligent fabric converges programmability, sensing and real-time simulation into a single network layer, making the physical world programmable at scale
The new dynamic
Network capabilities at scale, such as differentiated connectivity and integrated sensing, demand autonomous operations programmed on intents and supported by AI, thus transforming the network from a pipe into a platform. This will be further emphasized by agentic API exposure capabilities, enabling a new way to consume network services.
Network programmability will be critical to efficiently support open agentic network APIs, enabling applications to directly express their demands to the network. Early commercial traction in fraud prevention, identity verification and quality-on-demand has established the logic for what follows. The next step is intent-based networking, where CSPs declare desired outcomes and the network fulfills them autonomously, shifting the commercial model from selling connectivity to monetizing guaranteed outcomes.
Alongside this, integrated sensing and communication (ISAC) extends the network into environmental awareness, building on shared radio sites, spectrum and hardware, while introducing new sensing-specific signals, control mechanisms and processing capabilities to detect objects, map conditions and generate sensing data as a service across traffic monitoring, infrastructure protection and urban management.
Cognitive digital twins anchor this programmable layer to the physical world, evolving from passive simulations into self-adapting models that respond to real-world change in real time. On the factory floor of Industry 5.0, collaborative robots (cobots) and human workers collaborate, guided by twins that adapt continuously to production changes, with machines handling hazardous and repetitive tasks, while humans focus on higher-value work. Spatial computing extends this further, softening the boundary between physical and digital worlds across health care, logistics and manufacturing.
Architectural trends
We are building the programmable network layer by layer: from API exposure and intent-based management today, through cognitive digital twins and AI-powered autonomous operations, to full ISAC-sensing capabilities in 6G.
Autonomous networks are evolving from traditional rule-based automation toward systems capable of learning, adapting and making business-critical decisions in real time. Through global ecosystem partnerships, we are creating standardized, machine-readable network APIs through which developers, enterprises and AI agents can access and build on operator capabilities at scale.
As the primary consumers of those interfaces shift from human developers to autonomous agents, we are designing the authentication and validation frameworks that will determine who shapes participation in the programmable network and who captures the value it generates. ISAC is progressing through 3GPP standardization, and we are actively contributing to the architecture that will deliver its full sensing potential at 6G network scale.
Next steps for the industry
By 2030, the autonomous, intent-driven intelligent network will be the invisible yet indispensable fabric of a fully programmable, human-centered world. The commercial decisions being made today will determine who builds it and who benefits.
Near-term, the priority is API monetization at scale, building the developer and enterprise ecosystems that make differentiated connectivity commercially viable. The medium-term opportunity is sensing-as-a-service and cognitive digital-twin integration. Realizing this requires addressing the challenges that remain: cross-vendor interoperability, zero-trust security across programmable layers, and the skills development that widespread deployment demands. The CSPs and ecosystem partners that address these challenges systematically, rather than deferring them, will be those that define the programmable world rather than inherit it.
Trend 4: Heterogeneous compute
Heterogeneous compute, built on purpose-designed silicon, is the foundation on which a viable, scalable intelligent fabric will be built
The new dynamic
The shift to AI-native networks demands a compute strategy purpose-built to deliver the performance, real-time responsiveness and energy efficiency that CSPs require at global scale.
Advanced antenna systems in 5G mid-band marked a step change in compute demand that general-purpose architectures were never designed to meet; multiple antenna branches, wide-band beamforming and real-time signal processing in parallel demand a fundamentally different compute architecture. The evolution toward 6G and deeper AI integration across every layer of the RAN (radio access network) will increase that demand further. As AI inference workloads scale across the network, CSPs face a direct consequence: networks built on general-purpose hardware will face a widening gap between what they can deliver and what AI-native applications require, with rising energy costs that compound that disadvantage over time. The compute strategy that CSPs define today will determine their performance ceiling, energy economics and competitive positioning for the decade ahead.
Architectural trends
Purpose-built heterogeneous compute architectures, combining specialized silicon, tensor processing and telecom-specific accelerators, are the design foundation of AI-native RAN.
The compute architecture required for AI-native RAN combines specialized digital signal processors, tensor processing cores and telecom-specific accelerators, implemented through chiplet technologies and advanced packaging on leading-edge semiconductor nodes.
This heterogeneous approach enables AI workloads to execute directly within the RAN stack, at the microsecond-level latency that real-time radio functions demand, while minimizing energy consumption across the network footprint. Ericsson Silicon and the Ericsson Many-Core Architecture represent Ericsson’s implementation of this approach, with neural network accelerators now integrated directly into Massive MIMO radios to enable on-site AI inference at scale. The Tensor Core and DSP (Digital Signal Processing) compute elements are designed specifically for increasing AI algorithm use across the RAN stack, providing the computational foundation for the AI-native 6G transition. Looking further ahead, neuromorphic and quantum computing architectures represent long-horizon research investments, with near-term focus on heterogeneous chiplet strategies that scale with evolving network requirements.
Next steps for the industry
The organizations that define their compute strategy for AI-native networks now will be significantly better positioned to absorb the demands of 6G than those that address it reactively.
Purpose-built heterogeneous platforms enable the growing demands of AI inference to be absorbed within the existing network footprint, avoiding the costly reinvestment that reactive approaches require. As AI workloads scale, architectures that deliver the lowest cost of intelligence will generate an operational cost differential that widens over time. The transition to 6G will be the definitive test of these decisions, and CSPs with purpose-built compute foundations will be best placed to meet it.
Conclusion
My vision for the next decade is a world where the network does more than transport the data. It senses, decides and acts at the speed of intelligence itself, supporting humans, enterprises and autonomous systems alike.
The four trends examined in this article are not parallel forecasts. They are interrelated dimensions of a single transformation, each a precondition for the next. Agentic AI defines what the intelligent fabric must serve. Security and post-quantum cryptography define the conditions under which it earns trust. The programmable world defines what it can perceive and expose. And heterogeneous compute defines what makes it computationally and economically viable at global scale.
In my view, the intelligent fabric is to the AI economy what the best-effort mobile broadband service was to the internet economy in the past: the connective layer that makes everything else possible, and the differentiator that determines who leads and who follows. It is not built by extending what exists. It is built by designing for what is forming, with openness, programmability and trust as foundational properties from the outset.
Ericsson will not build this alone, and we have never intended to. The intelligent fabric will be shaped through deep collaboration across CSPs, enterprises, developers, governments and academia, all aligned on a common architecture and a shared direction. The decisions made in the next two to three years will determine how this transformation unfolds, and who is positioned to define it.
We are committed to leading that effort, and wherever you sit in the ecosystem, we invite you to shape it with us.
Further reading
- Ericsson Technology Review, Exploring the potential advantages of quantum computing in telecommunication networks
- Ericsson Technology Review, Enhancing developer experience to accelerate network automation
- Ericsson white paper, ICT energy evolution: Telecom, data centers, and AI
- Ericsson white paper, Differentiated connectivity: Unleashing the full potential of 5G
- Ericsson white paper, Intent-driven networks is a key step in the journey to autonomous networks
- Ericsson white paper, From CPaaS to a global network API platform, enabling CSPs to monetize on 5G
作者简介
作为集团首席技术官，Erik Ekudden负责爱立信集团的战略和技术发展方向。他在全球领先技术领域的丰富工作经验影响了公司在5G、6G、边缘计算、人工智能（AI）、增强/虚拟现实和物联网方面的战略决策及投资。Ekudden的领导才能建立在他在技术战略和行业活动领域的数十年经验之上。他于1993年加入爱立信，曾在公司担任多个管理职位，包括技术战略主管、美洲首席技术官（美国圣克拉拉）和标准化与行业负责人。他还是瑞典皇家工程科学院的成员，也是《爱立信技术评论》的出版人。