Lumen Targets Terabit Fiber Mesh as AI Workloads Overwhelm WANs

AI Has a Brain. Who Owns the Nervous System? With Lumen — Ai4

Ai4

As AI models commoditize, enterprise value shifts to programmable high-capacity networks that prevent multi-million-dollar compute clusters from sitting starved of real-time data.

AI-assisted summary of AI Has a Brain. Who Owns the Nervous System? With Lumen — original source. Check the original for full context.

Source published: 2026-08-17

Detailed notes

## Background Published on YouTube by Ai4 on August 17, 2026, this keynote features Jim Fowler, Chief Technology and Product Officer at Lumen. The presentation explores why telecom must evolve into core AI infrastructure by providing the high-capacity, low-latency 'nervous system' that allows AI models to sense, decide, and act in real time.

## TL;DR Jim Fowler argues that while industry attention centers on AI models as the 'brain,' these models are rapidly commoditizing and remain useless without a high-speed network 'nervous system' to feed them real-time data. With enterprise data sprawled across multi-cloud environments and AI workloads doubling annually, traditional Wide Area Networks (WAN) have become severe bottlenecks for expensive compute clusters. To address this, Lumen is rolling out 'Cloud 2.0' across three pillars: expanding high-capacity physical fiber, building a programmable on-demand multi-carrier mesh, and integrating software-defined network services. Ultimately, the successful deployment of autonomous, agentic AI depends on delivering secure, terabit-scale capacity with sub-5-millisecond latency between distributed machines and data.

## Atmosphere & style The presentation carries a pragmatic, authoritative, and grounded tone that cuts through speculative AI hype with hard infrastructure realities. Fowler blends humor—such as referencing Steve Martin in *The Man with Two Brains*—with nostalgic career anecdotes about floppy disks and manual Unix kernel compilation to contextualize enterprise architectural shifts. It feels like an insider masterclass on data-center plumbing and network engineering, pivoting from past computing eras to an urgent call for modernizing the physical foundations powering next-generation intelligence.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=bmXj16lJDjE&t=4s) AI models represent only the brain; without a physical and digital nervous system to transport real-time data, models remain isolated and unable to sense or act. - [00:04](https://www.youtube.com/watch?v=bmXj16lJDjE&t=4s) The model layer is rapidly becoming a commoditized space, making the underlying network infrastructure the true determinant of real-world AI utility. - [04:05](https://www.youtube.com/watch?v=bmXj16lJDjE&t=245s) To achieve autonomy, enterprise AI requires a closed-loop biological architecture that senses data in real time, decides via inference, and acts through automated execution. - [05:13](https://www.youtube.com/watch?v=bmXj16lJDjE&t=313s) Enterprise IT has evolved from simple, centralized on-premise networks to extreme sprawl across 50,000 remote locations, multiple public clouds, and unmapped SaaS databases. - [08:58](https://www.youtube.com/watch?v=bmXj16lJDjE&t=538s) AI workloads are projected to grow at a 100% compound annual growth rate through 2030, effectively doubling in volume each year. - [08:58](https://www.youtube.com/watch?v=bmXj16lJDjE&t=538s) Over 50% of global internet traffic is now machine-to-machine, driven by autonomous agents and distributed data center expansion set to grow 10x by 2030. - [11:25](https://www.youtube.com/watch?v=bmXj16lJDjE&t=685s) While Cloud 1.0 automated compute and storage, Cloud 2.0 must make the Wide Area Network programmable through software-defined, on-demand mesh architectures. - [11:25](https://www.youtube.com/watch?v=bmXj16lJDjE&t=685s) Inadequate WAN bandwidth can bottleneck multi-million-dollar GPU clusters, leaving advanced compute assets idle while waiting for training data. - [11:25](https://www.youtube.com/watch?v=bmXj16lJDjE&t=685s) Cloud-to-cloud and data-center-to-data-center machine traffic is growing at 13% annually, vastly outpacing traditional premise-to-cloud traffic growth of 1%. - [16:51](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1011s) Telecommunications providers must move away from static pipe sales to dynamic, cloud-like consumption models managed through a single pane of glass. - [18:02](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1082s) Hyperscale AI demands a massive step-function in bandwidth, shifting purchasing patterns toward multi-terabit blocks and 400G, 800G, and 1.6 Tbps circuits. - [18:02](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1082s) AI models are being actively weaponized by adversaries to discover vulnerabilities at the physical and firmware levels of legacy, unsupported network hardware. - [18:57](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1137s) Achieving real-time enterprise AI performance requires software-defined networks capable of delivering sub-5-millisecond latency across distributed data centers.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=bmXj16lJDjE&t=4s) Beyond the Brain: Why AI Needs a Nervous System

Jim Fowler, Chief Technology and Product Officer at Lumen, opens by addressing the prevailing obsession across the AI industry: the relentless focus on models. While conference stages are filled with discussions about models growing larger and faster, benchmark records, and apocalyptic predictions about the end of human jobs, Fowler argues that a critical half of the equation is being neglected. An AI model is merely a brain, and that brain is only as intelligent as the data feeding it. As enterprise data sprawls across disparate environments, the true bottleneck becomes how quickly and reliably that data can be transported to the model so it can take real-time action.

``` [ Sensing ] ---> [ Deciding ] ---> [ Acting ] IoT / Cash Regs AI Models Automated Code Plant Machinery (The Brain) Network Healing ^ | |========== NERVOUS SYSTEM =============| (WAN / Fiber) ```

Drawing a humorous comparison to the Steve Martin comedy *The Man with Two Brains*, Fowler illustrates the futility of an isolated model: a brain sitting in a glass jar can do nothing on its own. Without eyes, it cannot see; without hands, it cannot touch or act; without a mouth, it cannot speak; and without feet, it cannot walk. In the exact same way, an AI model disconnected from IoT sensors on vehicles, factory machinery on the plant floor, or retail point-of-sale registers is utterly blind to reality. In Fowler's view, the models themselves are rapidly becoming a commoditized layer where providers constantly leapfrog one another. The enduring, complex challenge lies in building the physical and digital nervous system that delivers real-time data into the model.

- AI models depend entirely on their data fuel, yet enterprise data sprawl is largely overlooked. - Emerging use cases span autonomous coding agents, human-sounding customer service bots trained on hundreds of thousands of hours of call audio, automated grocery replenishment (e.g., Walmart, Target, Kroger), and scientific breakthroughs mined from decades-old research. - The model layer is commoditizing; real enterprise value depends on the nervous system closing the loop between data collection and execution.

---

[04:05](https://www.youtube.com/watch?v=bmXj16lJDjE&t=245s) Sense, Decide, Act: The Autonomous Feedback Loop

To make artificial intelligence truly actionable, infrastructure must mirror biological nervous systems through a continuous three-part loop: sensing, deciding, and acting. In the human body, holding a hand over a hot surface instantly triggers the sensory nerves; the brain recalls past pain, decides on a course of action, and signals the muscles to pull the hand away. AI requires this exact real-time feedback mechanism. Data must be captured as it happens, transported to a model trained to make an inference decision, and immediately translated into automated downstream execution—such as autonomously deploying code, resolving a customer trouble ticket, or executing self-healing actions within a telecommunications network.

---

[05:13](https://www.youtube.com/watch?v=bmXj16lJDjE&t=313s) The Evolution of Complexity: From LAN Administrator to Multi-Cloud Sprawl

Fowler reflects on his early career as a Local Area Network (LAN) administrator at NCR to highlight how drastically enterprise architecture has transformed. In those days, setting up Windows PCs with software to enable basic email and file sharing felt like high-tech magic. In hindsight, that environment was extraordinarily simple: all employees worked on-site inside corporate-owned facilities, networks were completely self-contained, and physical security gates served as the primary firewall because there was no public internet connectivity. The most severe security threat was an employee bringing a infected floppy disk from home, and the entire enterprise ran on two mainframes in data centers separated by merely a hundred yards.

Today's landscape represents total decentralization and sprawl. Corporate users are distributed across 50,000 remote locations and home offices rather than centralized company buildings. Instead of two mainframes holding all enterprise data, information is scattered across hundreds of thousands of disparate databases and third-party SaaS platforms where organizations often have no visibility into the physical location of their data. Compute infrastructure is fragmented across hybrid environments, multiple public cloud providers, multiple regions, and multiple availability zones. Meanwhile, cybersecurity threats have escalated from floppy-disk viruses to sophisticated nation-state actors actively attempting to infiltrate networks, exfiltrate intellectual property, and cause systemic disruption.

---

[08:58](https://www.youtube.com/watch?v=bmXj16lJDjE&t=538s) The Explosion of AI Workloads and Data Center Growth

The shift toward hyper-distributed multi-cloud infrastructure is not a distant prediction—it is already the baseline reality for modern enterprise computing. Today, 90% of enterprises operate across multiple cloud providers, largely because diverse SaaS platforms and specialized workloads dictate their own cloud environments. At the same time, the transition toward autonomous agentic workflows is driving an exponential surge in data traffic, with AI workloads projected to expand at a 100% compound annual growth rate (CAGR)—effectively doubling in volume every single year through 2030.

``` +------------------------------------------------------------------+ | THE DATA SPRAWL REALITY | +------------------------------------------------------------------+ | - 90% of enterprises are Multi-Cloud | | - 100% CAGR for AI workload growth through 2030 | | - 10x surge in global data centers by 2030 | | - 50%+ of all internet traffic is machine-to-machine | +------------------------------------------------------------------+ ```

This workload explosion is mirrored by massive physical expansion: global data center capacity is slated for a 10x increase by 2030 based on facilities already planned or under construction. These new facilities are increasingly dispersed to wherever power, water, and favorable tax laws are available, pushing the computational edge further outward. Furthermore, over 50% of all global internet traffic has already transitioned to machine-to-machine communication, a figure set to climb dramatically as autonomous AI agents interact directly with other machines without human intervention.

---

[11:25](https://www.youtube.com/watch?v=bmXj16lJDjE&t=685s) Cloud 2.0 and the Three Pillars of the AI Nervous System

Fowler draws parallels between the current networking bottleneck and his second job as a Unix server administrator, when provisioning compute required manually partitioning physical hard drives, compiling the kernel, and swapping CDs. Cloud 1.0 automated and virtualized compute and storage into programmatic, on-demand APIs, but it left the Wide Area Network (WAN) behind, stranded in physical meet-me rooms and manual router configurations. Cloud 2.0 must bring the network into the programmable era, and Lumen is executing on this vision across three core pillars:

1. **A High-Capacity Physical Fiber Foundation:** Lumen operates 17 million miles of fiber across the United States, expanding to 58 million miles by the end of 2031. The company has upgraded intercity photonics to 400 Gbps and 800 Gbps on the path to 1.6 Tbps across 65 major metropolitan markets, targeting 90% coverage of cloud on-ramps. Fowler recounts a conversation with the CIO of a Fortune 5 bank who rented high-end GPUs to train a proprietary model, only to have those multi-million-dollar compute assets sit idle because the WAN lacked the bandwidth to feed training data from remote storage to the cluster in real time. 2. **A Programmable, Cross-Carrier Mesh (Network as a Service):** Instead of network engineers manually touching 20 routers to provision bandwidth, WAN connectivity must operate as on-demand software. Through its acquisition of Alkira, Lumen is deploying a single software-defined control plane with virtual Points of Presence (POPs) embedded directly inside AWS, Microsoft Azure, Google Cloud Platform, and Oracle Cloud, creating multi-carrier mesh routing across competitive networks. 3. **Integrated Software Ecosystem Services:** Basic network functions such as firewalls, load balancing, and DNS routing must be deployed dynamically through software rather than deployed as hardware appliances.

This architectural shift matches the commercial reality of network traffic: legacy premise-to-cloud connectivity is growing at a modest 1% annually due to remote work and shrinking corporate office footprints, whereas machine-to-machine cloud-to-cloud and data-center-to-data-center connectivity is surging at 13% annually.

[16:51](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1011s) Moving from Point-to-Point Connections to an AI Nervous System

As agentic AI takes off, connectivity demands are set to escalate rapidly. The core challenge for infrastructure providers is making it as seamless as possible for customers to spin up, configure, and tear down high-speed connections in real time. In an AI-driven landscape, legacy point-to-point telecommunications models are obsolete. Network providers are no longer just selling static pipes; they are operating the biological equivalent of a real-time nervous system that bridges the gap between sensing, deciding, learning, and acting.

Meeting this demand requires an on-demand consumption model where bandwidth and routing can be dynamically altered to match fluctuating compute workloads. Rather than managing fragmented vendor tooling, enterprises require unified control from a "single pane of glass." This interface must orchestrate not only raw fiber throughput but also the surrounding edge and transport ecosystem—including software-defined firewall policies, security controls, and load balancing.

- **On-Demand Consumption:** Cloud-like flexibility to spin up and spin down network capacity on the fly. - **Unified Management:** Centralized configuration of networking, firewalls, and load-balancing services from a single console. - **The Core AI Mandate:** Serving as the connective tissue linking sensing, acting, and continuous model learning.

---

[18:02](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1082s) Terabit-Scale AI Capacity and Modern Infrastructure Security

AI infrastructure requirements demand an unprecedented step-function increase in network bandwidth, transitioning rapidly from 400-gigabit links to 800 gigabits and onward to 1.6 terabits. Hyperscalers and large enterprises are no longer purchasing bandwidth incrementally; they are securing multi-terabit capacity. Fowler illustrates this paradigm shift with a recent order Lumen booked from a hyperscaler deploying four hundred 100-gigabit circuits simultaneously to interconnect their AI clusters.

``` BANDWIDTH EVOLUTION FOR HYPERSCALE AI WORKLOADS

[ 400 Gbps ] ───────> [ 800 Gbps ] ───────> [ 1.6 Tbps ] │ │ │ └─────────────────────┼─────────────────────┘ ▼ Terabit-Scale Orders (e.g., 400x 100G) ```

At the same time, network security must evolve alongside throughput. Threat actors and next-generation AI models are actively weaponized to probe for vulnerabilities, scanning not just application-layer software but the physical and firmware layers of legacy networking gear. Running outdated, end-of-life, or end-of-support hardware creates critical exposure, as automated models can effortlessly identify the smallest "chink in the armor" to compromise enterprise environments.

- **Massive Pipeline Scaling:** Direct architectural transition from 400G to 800G and 1.6T circuits. - **Hyperscaler Volume:** Routine procurement of multi-terabit blocks (e.g., 400 x 100G circuits in a single deployment). - **AI-Driven Vulnerability Hunting:** Automated adversarial models targeting legacy, unsupported network appliances.

---

[18:57](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1137s) The Future of Connectivity: Connecting People, Machines, and Data Under 5 Milliseconds

In closing, Fowler emphasizes that while every organization is racing to make their products and systems smarter, "intelligence alone, though, is not enough." Raw algorithmic intelligence cannot function in isolation; it fundamentally relies on an agile, low-latency, high-capacity infrastructure that seamlessly binds people, autonomous machines, and distributed data pools together.

This operational reality looks entirely different from the traditional telecom environment of previous decades. Modern customers are demanding software-defined architectures capable of delivering sub-5-millisecond latency to every one of their distributed data centers. Building this responsive, hyper-low-latency fabric is what will ultimately dictate the speed, scope, and real-world impact of enterprise AI deployment.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=bmXj16lJDjE&t=4s) Everyone talks about AI models as the brain. But brains don’t matter without a nervous system. — Jim Fowler

> [00:04](https://www.youtube.com/watch?v=bmXj16lJDjE&t=4s) The real question is: who builds and controls the infrastructure that lets AI sense, decide, and act in real time. — Jim Fowler

> [00:04](https://www.youtube.com/watch?v=bmXj16lJDjE&t=4s) An AI model is merely a brain, and that brain is only as intelligent as the data feeding it. — Jim Fowler

> [00:04](https://www.youtube.com/watch?v=bmXj16lJDjE&t=4s) The model layer is commoditizing; real enterprise value depends on the nervous system closing the loop between data collection and execution. — Jim Fowler

> [11:25](https://www.youtube.com/watch?v=bmXj16lJDjE&t=685s) Cloud 1.0 automated and virtualized compute and storage into programmatic, on-demand APIs, but it left the Wide Area Network (WAN) behind, stranded in physical meet-me rooms and manual router configurations. — Jim Fowler

> [11:25](https://www.youtube.com/watch?v=bmXj16lJDjE&t=685s) Cloud 2.0 must bring the network into the programmable era — Jim Fowler

> [16:51](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1011s) Network providers are no longer just selling static pipes; they are operating the biological equivalent of a real-time nervous system that bridges the gap between sensing, deciding, learning, and acting. — Jim Fowler

> [16:51](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1011s) single pane of glass — Jim Fowler

> [18:02](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1082s) Hyperscalers and large enterprises are no longer purchasing bandwidth incrementally; they are securing multi-terabit capacity. — Jim Fowler

> [18:02](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1082s) chink in the armor — Jim Fowler

> [18:57](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1137s) intelligence alone, though, is not enough. — Jim Fowler

> [18:57](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1137s) Modern customers are demanding software-defined architectures capable of delivering sub-5-millisecond latency to every one of their distributed data centers. — Jim Fowler

## Mentioned Jim Fowler, Lumen, Steve Martin, The Man with Two Brains, Walmart, Target, Kroger, NCR, Alkira, AWS, Microsoft Azure, Google Cloud Platform, Oracle Cloud

## Video description Jim Fowler - Chief Technology and Product Officer, Lumen

Everyone talks about AI models as the brain. But brains don’t matter without a nervous system. The real question is: who builds and controls the infrastructure that lets AI sense, decide, and act in real time. That’s where telecom moves from connectivity provider to AI infrastructure company. I’ll talk about what it really means to build the nervous system for AI, and why the companies that get this right will define what AI can actually do.

[Source](https://www.youtube.com/watch?v=bmXj16lJDjE)

中文简报

AI工作负载压垮广域网，Lumen瞄准太比特光纤网格随着AI模型逐渐商品化，企业价值转向可编程的高容量网络，以防止价值数百万美元的计算集群因缺乏实时数据而处于闲置饥饿状态。 Jim Fowler认为，尽管行业焦点集中在作为“大脑”的AI模型上，但这些模型正在迅速商品化，如果没有高速网络“神经系统”为其输送实时数据，它们将毫无用处。随着企业数据分散在多云环境中，且AI工作负载每年翻一番，传统广域网（WAN）已成为昂贵计算集群的严重瓶颈。为解决这一问题，Lumen正通过三大支柱推进“Cloud 2.0”：扩建大容量物理光纤、构建可编程的按需多运营商网格，以及整合软件定义网络服务。归根结底，自主型Agentic AI的成功部署取决于在分布式机器与数据之间提供安全、太比特级容量以及低于5毫秒的低时延。 AI模型仅代表大脑；若没有传输实时数据的物理和数字神经系统，模型将陷入孤立，无法感知或采取行动。 (0:04) 模型层正迅速走向大宗商品化，底层网络基础设施因而成为决定AI现实实用价值的真正关键。 (0:04) 为了实现自主性，企业级AI需要一种闭环生物学架构，能够实时感知数据、通过推理作出决策，并通过自动化执行采取行动。 (4:05) 企业IT已从简单、集中的本地网络，演变为极度分散于50,000个远程分支、多个公有云及未纳入资产图谱的SaaS数据库中。 (5:13) 预计到2030年，AI工作负载将以100%的复合年增长率增长，其规模实际上每年翻倍。 (8:58)

Reading summary

Lumen Chief Technology and Product Officer Jim Fowler argues that artificial intelligence models are commoditizing into isolated "brains" whose real-world utility is constrained by enterprise networking bottlenecks. Fowler contends that enterprise value depends on building a programmable, high-capacity "nervous system" that transports distributed data in real time to complete the loop between sensing, deciding, and acting. The Sense-Decide-Act Autonomous Feedback Loop (0:04) Fowler asserts that the tech industry's intense focus on AI model parameters and benchmark records overlooks the essential infrastructure needed to feed those models. Comparing an isolated model to a brain in a jar, he argues that algorithms disconnected from real-time edge inputs—such as factory machinery, IoT sensors, or retail point-of-sale systems—are blind and ineffective. In his view, foundational models will rapidly commoditize as providers leapfrog one another, shifting enterprise differentiation toward network infrastructure that closes the operational loop. To create actionable autonomy, Fowler argues enterprise architecture must mirror biological systems across three continuous stages: sensing, deciding, and acting. Rather than merely processing retrospective queries, AI systems must ingest live telemetry, compute inferences, and immediately execute automated workflows. Real-world applications of this closed loop include autonomous code deployment, customer service bots trained on call audio, automated grocery replenishment at retailers like Walmart and Target, and self-healing telecommunications networks. Decentralized Sprawl and Expanding Machine Traffic (5:13) Tracing the evolution of enterprise computing from early on-premises local area networks to modern distributed environments, Fowler highlights an extreme fragmentation of data and compute. Corporate users are distributed across tens of thousands of remote locations, while 90% of enterprises operate across multiple cloud providers. Enterprise data is scattered across hundreds of thousands of databases and SaaS platforms, often without visibility into its physical storage location. This hyper-distributed footprint coincides with an explosion in machine-driven workloads. Fowler cites industry forecasts projecting that AI workloads will expand at a 100% compound annual growth rate through 2030, doubling yearly, alongside a planned 10x surge in global data center capacity by 2030. Furthermore, machine-to-machine traffic already accounts for over 50% of all internet traffic. This structural shift is reflected in enterprise bandwidth trends: premise-to-cloud traffic is growing at only 1% annually, whereas cloud-to-cloud and data-center-to-data-center connectivity is surging at 13% annually. Cloud 2.0: Virtualizing the Wide Area Network (11:25) Fowler observes that while "Cloud 1.0" virtualized compute and storage into on-demand APIs, wide area networking remained stranded in manual router configurations. This network bottleneck directly threatens high-performance computing; Fowler recounts a Fortune 5 bank CIO whose multi-million-dollar rented GPU clusters sat idle because wide area networks lacked the throughput to feed training data from remote storage in real time. To address this, Fowler outlines Lumen's "Cloud 2.0" strategy across three pillars: physical fiber scaling, on-demand programmable meshing, and integrated software services. Lumen plans to expand its U.S. intercity fiber footprint from 17 million to 58 million miles by the end of 2031, upgrading photonics from 400 Gbps and 800 Gbps toward 1.6 Tbps across 65 metropolitan markets. A software-defined control plane, bolstered by Lumen's acquisition of Alkira, embeds virtual points of presence directly inside AWS, Azure, GCP, and Oracle Cloud, enabling multi-carrier mesh routing and on-demand provisioning of firewalls and load balancers. Terabit-Scale Bandwidth and Hardware Security Vulnerabilities (18:02) Supporting agentic AI requires a step-function increase in network bandwidth, shifting enterprise and hyperscaler procurement toward multi-terabit capacity. Fowler illustrates this demand with a hyperscaler deployment ordering 400 simultaneous 100-gigabit circuits from Lumen to interconnect AI clusters. To support real-time machine-to-machine execution, modern enterprise architectures increasingly demand network fabrics that deliver sub-5-millisecond latency across distributed data centers. Simultaneously, Fowler warns that network expansion introduces heightened security risks. Threat actors and nation-states are deploying adversarial AI models to scan for vulnerabilities across infrastructure. Rather than targeting only application software, automated models actively probe physical and firmware layers of legacy, unsupported networking hardware, turning end-of-life appliances into critical points of enterprise exposure. Fowler concludes that raw algorithmic intelligence cannot function in isolation. The ultimate scope and impact of enterprise AI will depend on whether organizations deploy agile, software-defined network fabrics capable of binding dispersed data, people, and machines under sub-5-millisecond latency thresholds.

精炼阅读摘要

Lumen首席技术与产品官Jim Fowler指出，当前AI产业对算法模型的单向追逐忽略了数据传输瓶颈；随着模型逐渐商品化，真正的企业竞争力取决于能否构建低延迟、可编程的物理与数字“神经系统”，实现感知、决策与执行的实时闭环。突破模型孤岛与自主闭环机制 (0:04) Fowler认为，脱离外部数据输入的AI模型如同置于玻璃罐中的孤立大脑，若缺乏与车辆物联网传感器、工厂流水线或零售终端等边缘设备的数据连接，模型便无法感知物理现实，更无法采取有效行动。在开源与商业模型迭代迅速、各厂商交替领先的背景下，模型算法层正迅速演变为高度同质化的大宗商品。企业实现AI落地的核心壁垒，在于打通由物理光纤与广域网构建的数字神经通道。为了使人工智能具备自主行动力，网络基础设施必须重现生物神经系统的三元闭环：感知（Sensing）、决策（Deciding）与行动（Acting）。数据在产生时必须被实时捕获并传输至模型端进行推理决策，随后直接触发自动化执行，例如自主部署代码、线上客服工单化解，或在电信网络内部完成自动化自愈修复，从而摆脱对人工介入的依赖。企业多云扩散与机器间流量爆发 (5:13) 企业计算架构已从早期局域网（LAN）时代的单点部署彻底演变为高度复杂的去中心化多云生态。过去企业员工集中在固定办公区，数据汇聚于物理隔离的两台大型机；而如今，大批远程办公节点分散接入，企业数据散落在成百上千个外部SaaS平台与分布式数据库中，边界防火墙的物理防护完全瓦解。Fowler引述行业数据指出，目前已有90%的企业在跨多个公有云运行各类业务。伴随AI自主智能体（Agent）工作流的落地，全球网络流量特征正在发生结构性剧变。据预测，到2030年AI计算负载将保持每年100%的复合年增长率（CAGR），即数据量每年翻倍；同时全球数据中心容量将因算力需求扩大10倍，并向具备能源与税收优势的边缘区域分散。目前全球已有超过50%的互联网流量属于机器对机器（M2M）通信。反映在网络需求上，传统的本地办公点到云端连接年增长率已萎缩至1%，而云与云、数据中心之间的机器互联需求年增长率则达到13%。云2.0网络支柱与按需编排能力 (11:25) 以往的“云1.0”虚拟化了计算与存储资源，却将广域网（WAN）遗留在依赖人工配置路由器与机房跳线的物理时代。Fowler透露，一家全美前五的大型银行CIO曾斥巨资租用高端GPU集群训练专有模型，却由于广域网带宽受限无法从异地存储向集群实时喂入数据，致使昂贵的算力资源长期空转闲置。为此，Fowler主张通过三大支柱推进网络向“云2.0”演进：第一是扩展超大容量光纤，Lumen计划将其全美光纤网络从1700万英里扩展至2031年底的5800万英里，并在65个核心大都市区将城际光通信升级至400 Gbps、800 Gbps乃至1.6 Tbps，实现对90%云接入点的覆盖；第二是通过收购Alkira，在各大公有云内部署软件定义虚拟接入点，建立跨运营商的网状网络即服务（NaaS）；第三是将防火墙、DNS路由和负载均衡等硬件设备完全转换为软件化服务。针对AI工作流的突发性与不可预测性，企业网络管理正从固定链路转向云原生的按需消耗模式。基础设施必须通过“单窗格”（single pane of glass）统一控制台，使企业能够像申请云算力一样，在几分钟内动态调配、扩展或释放高带宽网络管道及关联的安全策略，以匹配瞬息万变的训练与推理负载。太比特级扩容诉求与硬件安全隐患 (18:02) 超大规模云服务商（Hyperscalers）对网络吞吐的需求正在经历阶梯式跃迁，采购单位已从常规带宽升级直接跨入多太比特（Terabit）级别。Fowler举例称，Lumen近期承接了一家超大规模客户的采购订单，该客户单次即部署了400条100 Gbps电路，以支撑其超大AI集群间的海量互联需求。与此同时，高带宽扩张也带来了更严峻的基础设施安全挑战。攻击者已开始利用AI模型自动化扫描并探测网络缺陷。这种对抗性模型不仅攻击顶层应用，还会主动探测底层网络设备的物理层与固件漏洞。Fowler警示，企业若继续在骨干链路上运行已过保修期或停止软件支持的老旧网络硬件，微小的底层固件缺陷就可能被自动化攻击工具迅速捕获，从而威胁整个AI生态的系统安全。 Fowler总结认为，单凭算法智能无法单独支撑产业变革，企业AI的价值最终依赖于能否建立连接分散数据、分布式计算与终端用户的极低延迟网络架构。尽管客户已普遍提出对分布式数据中心实现低于5毫秒超低延迟的要求，但如何在巨额重资产光纤建设与敏捷软件编排之间取得平衡，仍是通信基础设施服务商面临的核心挑战。

Detailed notes — zh

## Background 该视频由频道 Ai4 于 2026 年 8 月 17 日发布在 YouTube 平台上。内容为 Lumen 首席技术与产品官 Jim Fowler 的主题演讲，主要探讨在各界聚焦 AI 模型作为“大脑”的背景下，谁来掌控让 AI 能够实时感知、决策与行动的“神经系统”，以及电信网络如何从传统连接服务商转型为核心 AI 基础设施提供商。 ## TL;DR 演讲指出 AI 模型正逐步走向商品化，若缺乏强韧的网络神经系统，模型将沦为无法感知与执行任务的“缸中之脑”。随着机器间通信占据半数以上互联网流量以及数据中心爆发式扩张，产业正从算力与存储云化的 Cloud 1.0 迈向广域网全面可编程的 Cloud 2.0 时代。Lumen 通过建设数千万英里太比特级物理光纤网络、收购 Alkira 打造云原生虚拟 PoP 点与软件控制平面，以及全面软件化安全与负载服务，构建起 AI 神经系统的三大支柱。现代 AI 基础设施已进入太比特级容量采购、主动抵御 AI 漏洞攻击以及追求跨数据中心 5 毫秒超低延迟的新范式阶段。 ## Atmosphere & style 演讲兼具技术老兵的幽默怀旧与敏锐前瞻，Jim Fowler 巧妙借用《双脑人》电影隐喻和早期局域网管理员的经历开场，以轻松自嘲的语调拉近与现场观众的距离。随着话题切入海量数据蔓延、国家级网络威胁以及算力闲置等严峻现实，叙事氛围迅速转为硬核务实且充满紧迫感。结合一线超大规模采购订单与具体银行客户案例，整场演讲呈现出兼具宏大行业视野与扎实工程落地底蕴的专业气场。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=bmXj16lJDjE&t=4s) AI 的大脑与被忽视的“神经系统” Lumen 首席技术与产品官（Chief Technology and Product Officer）Jim Fowler 在开场时介绍了 Lumen 作为全球最大电信光纤网络提供商之一的背景，强调其拥有全球规模最大、互联与对等互联（peered）最广泛的网络基础设施之一。Jim Fowler 分享了他在大会现场观察到的普遍现象：几乎所有人都在热火朝天地讨论大模型——模型变得多大、多快，跑分基准如何，或者模型是否会摧毁世界、终结人类现有的工作。然而，他指出当前技术对话中遗漏了一个至关重要的核心：**模型的能力完全取决于喂养它的燃料，而这种燃料就是数据**。当前 AI 领域最被低估的痛点是“数据的无序蔓延”（data sprawl）。如果将 AI 比作大脑，那么真正决定 AI 能否有效运转的核心问题在于：究竟谁拥有并掌控着连接海量数据与信息的“神经系统”？从软件开发（训练了数百万行代码的模型已能理解整个代码库并由智能体代写代码），到客服中心（利用数十万小时通话数据训练出声音与行为极像真人、能自主操作后台系统的智能体），再到商业零售（如 Walmart、Target、Kroger 正在布局让消费者的智能冰箱在牛奶耗尽时自动完成跨网络静默下单），乃至科学探索（深入挖掘 20 年前未被重视的研究数据攻克世纪难题），AI 带来的所有颠覆性体验，其底层都依赖于海量数据如何在网络神经系统中高效、实时地传输。 - 模型本身无法脱离数据独立产生商业价值，及时的实时数据流输入是模型执行任务的前提。 - 软件工程、拟真客服智能体、自主物联网零售交易及科研数据挖掘等新场景，正在全面重塑企业的生产与运营模式。 --- ### [02:42](https://www.youtube.com/watch?v=bmXj16lJDjE&t=162s) 缸中之脑隐喻与“感知-决策-行动”闭环为了形象地解释网络连接的重要性，Jim Fowler 幽默地提到了喜剧演员 Steve Martin 主演的老电影《双脑人》（*The Man with Two Brains*）。他将脱离基础设施的 AI 模型比喻为一个被泡在玻璃罐里的“缸中之脑”：一个孤立的大脑即使再聪慧，如果没有眼睛就无法观察，没有双手就无法触摸感知，没有嘴巴就无法说话呼吸，没有双脚就无法行走。AI 模型亦是如此——如果它不能连接到车载 IoT 传感器、工厂车间的机械设备或零售店的收银机，它就无法接收任何感知数据，自然也无法做出任何实际行动。 Jim Fowler 甚至直言，从网络基础设施的视角来看，**AI 模型本身正在迅速走向“商品化”（Commodity）**。各大模型之间频繁出现性能反超（leapfrog），构建模型已不再是最艰巨的技术瓶颈；真正的挑战在于如何将数据在毫秒级时延内安全、稳定地输送给模型。一套完整的 AI 神经系统必须具备并闭合人类神经反射的三个核心阶段： ```mermaid flowchart LR accTitle: AI 神经系统闭环流程 subgraph SENSE["1. 感知 (Sense)"] S1["IoT 传感器 / 车间设备 / 收银机"] S2["实时业务流 / 数据蔓延端点"] end subgraph DECIDE["2. 决策 (Decide)"] D1["AI 大模型 / 推理引擎"] end subgraph ACT["3. 行动 (Act)"] A1["生成代码 / 工单处理"] A2["网络自治自愈 / 自主交易"] end S1 --> D1 S2 --> D1 D1 --> A1 D1 --> A2 ``` 1. **感知（Sense）**：就像手悬在滚烫物体上方感知到高温一样，AI 神经系统需要实时捕获边缘与各终端涌入的数据。 2. **决策（Decide）**：大脑结合过往被烫伤的记忆与经验进行判断；AI 模型根据预先训练的知识和最新推理上下文做出决策。 3. **行动（Act）**：手迅速缩回以防烧伤；AI 智能体自主编写一段代码、流转处理故障工单，乃至对通信网络执行自愈（autonomous self-healing）操作。 --- ### [05:13](https://www.youtube.com/watch?v=bmXj16lJDjE&t=313s) 从 NCR 局域网管理员时代到现代数据孤岛的反差 Jim Fowler 回顾了自己走出校园后的第一份工作——在 NCR 担任局域网管理员（LAN Administrator）。他调侃台下发出会心笑声的观众“暴露了年龄”，并回忆起当年自己为 Windows PC 一台台安装网络软件、配置企业最早的局域网以实现内部电子邮件、文件共享和日程协作的场景。在当时看来，这项工作犹如具有魔力一般复杂而前沿，但以现代视角审视，当时的 IT 拓扑其实极其简单封闭：所有员工都在公司自有的物理办公楼内上班，没有远程办公；员工连接的是企业自行搭建并管理的专用网络；物理围墙与园区大门构成了天然的安全边界，连防火墙都不需要，因为根本没有连接外部互联网；当时最大的安全威胁仅仅是有员工把软盘（floppy disk）带回家感染病毒后再拷回公司；整个公司的全部数据甚至只存放在相距仅 100 码的两座数据中心的大型机里。Jim 感慨道，面对今天 AI 基础设施的极端复杂性，他甚至无比怀念过去那个简单直白、边界清晰的局域网时代。 --- ### [07:14](https://www.youtube.com/watch?v=bmXj16lJDjE&t=434s) 现代混合多云现实与 AI 算力爆发式增长今天的企业现实与几十年前的孤岛架构相比发生了翻天覆地的剧变。现代企业面临着前所未有的“数据、用户与应用彻底解耦与分散”的复杂局面： ```mermaid flowchart TD accTitle: 现代企业分散架构与安全威胁 subgraph USERS["全球化分散用户"] U1["50,000+ 远程/居家/客户现场办公节点"] end subgraph DATA_APPS["碎片化应用与数据蔓延"] APP1["SaaS 软件服务商 (位置未知)"] DB1["分布全球的数十万数据库 / Mainframe"] end subgraph INFRA["混合多云基础设施"] C1["跨云厂商 (Multi-Cloud / 90% 企业采用)"] C2["多 Region / 跨可用区 (AZ)"] C3["本地私有数据中心 (Private DC)"] end subgraph THREATS["国家级网络威胁"] SEC1["国家级黑客组织 (数据窃取 / IP 盗窃 / 破坏)"] end USERS <--> INFRA DATA_APPS <--> INFRA THREATS -. 攻击尝试.-> INFRA ``` 用户不再集中于自有办公楼，而是分散在世界各地数以万计的家庭或客户现场；数据不再封存于两台大型机，而是散落在全球数十万个数据库中；核心业务应用被外包给各类 SaaS 供应商，企业甚至难以确知业务数据具体驻留在哪个地理节点；底层计算架构由跨多个云服务商、多地域（Regions）、多可用区（Availability Zones）及本地私有数据中心组成的混合多云环境支撑；网络安全威胁也从“防范带病毒的软盘”演变为抵御试图窃取核心知识产权、破坏关键基础设施的国家级黑客（nation-states）。 - **90% 的企业采用多云架构**：早期企业试图押注单一云厂商的设想已被证伪，SaaS 软件自身的部署决策直接迫使企业必须面对多云治理。 - **100% 的 AI 工作负载年复合增长率（CAGR）**：在智能体（agentic workflows）加速接管生产业务的驱动下，AI 工作负载在 2030 年前预计每年规模翻倍。 - **数据中心容量需求爆发**：为了承载机器与机器之间海量数据的实时互通，全球数据中心的扩建与能耗规划正在以肉眼可见的惊人速度增长。 ### [10:00](https://www.youtube.com/watch?v=bmXj16lJDjE&t=600s) 数据中心扩张与机器流量爆发：从 Cloud 1.0 到网络可编程的 Cloud 2.0 从现在到 2030 年，全球已规划并正在建设的数据中心数量将迎来整整 10 倍（10x）的爆发式增长，而这还只是目前账面上的确定数据。数据中心的选址不再受传统逻辑束缚，而是哪里有充足的电力、充足的水源、优惠的税收政策，数据中心就会落地在哪里。这也意味着企业的数据资产将以前所未有的速度在地理上极度分散，边缘节点变得越来越碎片化。回顾自身职业生涯的起点，Jim Fowler 感慨当年 AI 仅仅是教科书里的一个概念，或者是电影《2001太空漫游》（*2001: A Space Odyssey*）里的科幻想象；而如今，全网已有超过 50% 的互联网流量属于“机器对机器”（Machine-to-Machine, M2M）通信，且随着机器间自动化事务的爆发，这一比例只会不断飙升。AI 模型本身固然强大，但它的智能上限完全取决于能否及时获取并处理海量数据。这就需要一个前所未有的强韧“神经系统”，而这种基础设施目前在全球范围内尚未真正成型。这一演进路径与 Cloud 1.0 的发展如出一辙。Jim Fowler 回忆自己的第二份工作是 Unix 系统管理员，当时必须手动给硬盘分区、配置 Unix 内核、编写底层代码，甚至还要手动换光盘和软盘来安装系统；而如今的系统管理员只需在云端输入几条指令，虚拟机、存储资源和 Web 服务器就会自动就绪。Cloud 1.0 将算力与存储彻底“云化”和“可编程化”，但广域网（WAN）却被遗忘在接线机房（Meet-me rooms）中，处于物理割裂状态。进入 Cloud 2.0 时代，核心使命就是将广域网络全面纳入可编程的云化体系中。 - **数据中心规模预测**：2030 年前全球数据中心数量将实现 10 倍增长，选址由电力、水资源与税收法规驱动。 - **流量结构转变**：目前 50% 以上的互联网流量为机器对机器（M2M）通信。 - **Cloud 1.0 与 2.0 的对比**：Cloud 1.0 实现了算力与存储的即时编排；Cloud 2.0 则必须攻克广域网的可编程性。 --- ### [12:49](https://www.youtube.com/watch?v=bmXj16lJDjE&t=769s) AI 神经系统的三大支柱与超大规模光纤网络建设 AI 的神经系统主要由三大支柱构成，其根基是一张超大容量的物理光纤网络。为了应对超乎预期的海量数据迁移需求，Lumen 正在全美范围内对其光子网络（Photonics）进行跨代升级，从 400G 迈向 800G，并向 1.6T（Terabit）城际互联容量演进，覆盖全美 65 个主要大都市圈，并在年底前接入全美 90% 的云接入点（Cloud On-ramps）。谈及底层带宽不足的痛点，Jim Fowler 分享了近期与一家全美前五大银行（Fortune 5 Bank）CIO 的交流：该银行为训练专属大模型租用了大量昂贵的 GPU 算力，但由于可用 GPU 与核心训练数据不在同一个地理位置，网络专线带宽严重受限，数据传输速度远远跟不上计算需求，导致最昂贵的 GPU 资产长时间处于闲置等待状态。为了彻底解决这一瓶颈，Lumen 目前在全美运营着 1700 万英里的光纤网络，并计划在 2031 年底前将其扩张至 5800 万英里，以此打通各大分散的数据中心节点。 ```mermaid flowchart TD accTitle: AI 神经系统三层架构图 subgraph L3["第三层：生态增值服务 (Ecosystem Services)"] FW["软件定义防火墙"] LB["自适应负载均衡"] DNS["可编程 DNS 与安全套件"] end subgraph L2["第二层：软件控制平面 (Control Plane / Alkira)"] NaaS["网络即服务 (NaaS) 编排"] VPoP["云原生虚拟 PoP 点 (AWS / Azure / GCP / OCI)"] Mesh["跨运营商多网格协同 (Multi-Carrier Mesh)"] end subgraph L1["第一层：底层超宽物理光纤 (Physical Network)"] Fiber["5800万英里全美光纤骨干 (至2031年)"] Bandwidth["400G / 800G / 1.6Tbps 城际超高速互联"] DC["覆盖 65 个核心都市圈 & 90% 云接入点"] end L1 --> L2 L2 --> L3 ``` --- ### [14:14](https://www.youtube.com/watch?v=bmXj16lJDjE&t=854s) 软件定义、多运营商网格与云间互联的高速增长 AI 神经系统的第二大支柱是“全面可编程性”。过去跨越全美配置一条专线需要网络工程师手动调试 20 台路由器的时代已经终结。现代网络必须以“网络即服务”（Network as a Service, NaaS）的形式运行，允许用户根据实时负载按需拉起（Spin up）或释放（Spin down）端到端带宽。此外，任何单一电信运营商都无法独立承载未来 AI 的全局流量，必须建立跨运营商的网状网络（Multi-carrier mesh network）。为此，Lumen 近期完成了对网络虚拟化公司 **Alkira** 的收购。Alkira 颠覆了传统“将 PoP 点（网络接入点）设在物理布线中心”的模式，将其重塑为直接嵌入在 AWS、Google Cloud、Microsoft Azure 及 Oracle Cloud 等主流公有云内部的“虚拟 PoP 点”。通过统一的软件控制平面，企业可以像调度云资源一样自由构建云到云、云到本地的连接。第三大支柱则是将生态增值服务（防火墙、负载均衡、DNS）完全软件化。随着办公模式向居家及混合办公转型，传统的“企业本地到云端”（Premise-to-Cloud）专线市场年复合增长率仅剩 1% 左右；相反，“云到云 / 数据中心到数据中心”（Cloud-to-Cloud / Data Center-to-Data Center）的机器互联市场正在以每年 13% 以上的高速增长。随着自主智能体（AI Agents）与分布式计算的爆发，云间流量的增长还将进一步加速。 | 业务维度 | 传统网络体系 (Legacy Telco) | AI 时代神经系统 (Cloud 2.0 / NaaS) | |:--- |:--- |:--- | | **连接模式** | 静态、人工逐台配置路由器的点对点专线 | 软件定义、按需伸缩的端到端可编程连接 | | **PoP 点形态** | 物理机房与接线柜（Wiring Centers） | 部署在公有云内部的逻辑虚拟 PoP（如 Alkira 技术） | | **网络附加服务** | 物理防火墙、硬件负载均衡器、独立 DNS 盒子 | 随网络链路一键下发的云原生软件定义服务 | | **主要增长市场** | 本地到云端（Premise-to-Cloud，年增长约 1%） | 云到云/数据中心间（Cloud-to-Cloud，年增长超 13%） | --- ### [17:09](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1029s) AI 时代网络新范式：太比特级容量、主动安全与 5 毫秒低延迟面对由 AI 主导的产业格局，网络基础设施必须重塑其商业模式与技术标准。企业不再只是购买一条静态的点对点连接通道，而是在接入一套让 AI 实现“感知、决策与行动”闭环的神经系统。所有带宽容量与外围安全策略都必须支持单一口径（Single pane of glass）的统一配置。超高带宽需求已经成为现实。Jim Fowler 透露，就在两周前，某家超大规模云服务商（Hyperscaler）刚刚向 Lumen 下达了一笔采购 **400 条 100G 电路** 的超级大单，整体容量直接进入太比特（Terabits）级别，许多大型企业客户也在跟进同等量级的订单。与此同时，网络安全迎来了全新挑战：下一代 AI 模型正被用于主动探测企业软件及老旧网络硬件的底层漏洞。如果企业仍依赖大量处于停产（End-of-Life）或停止支持（End-of-Support）状态的陈旧硬件，AI 将轻而易举地找出这些“盔甲上的裂缝”（"chink in the armor"）并发动入侵。在演讲的结尾，Jim Fowler 强调：“单靠模型自身的智能是远远不够的，我们必须以极低的延迟和超高的带宽，将人类、机器与海量数据紧密连接在一起。”如今的头部客户在签订下一代网络合同时，明确要求网络运营商必须具备为其所有分布式数据中心提供 **5 毫秒（5ms）以内** 互联延迟的硬核能力。唯有具备超大容量、超低延迟及软件定义敏捷性的基础设施，才能真正支撑起 AI 驱动的未来世界。 - **超算级采购案例**：某 Hyperscaler 单笔订购 400 条 100G 线路，网络吞吐直接迈向数十 Terabits 级别。 - **AI 时代硬件安全**：下一代 AI 漏洞挖掘工具正专门针对过期的 EOL/EOS 网络硬件进行渗透攻击。 - **下一代互联硬指标**：实现跨分布式数据中心群 5 毫秒（5ms）以内的极致超低延迟。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=bmXj16lJDjE&t=4s) 模型的能力完全取决于喂养它的燃料，而这种燃料就是数据。 — Jim Fowler > [02:42](https://www.youtube.com/watch?v=bmXj16lJDjE&t=162s) AI 模型本身正在迅速走向“商品化”（Commodity）。 — Jim Fowler > [02:42](https://www.youtube.com/watch?v=bmXj16lJDjE&t=162s) 如果它不能连接到车载 IoT 传感器、工厂车间的机械设备或零售店的收银机，它就无法接收任何感知数据，自然也无法做出任何实际行动。 — Jim Fowler > [10:00](https://www.youtube.com/watch?v=bmXj16lJDjE&t=600s) Cloud 1.0 将算力与存储彻底“云化”和“可编程化”，但广域网（WAN）却被遗忘在接线机房中，处于物理割裂状态。进入 Cloud 2.0 时代，核心使命就是将广域网络全面纳入可编程的云化体系中。 — Jim Fowler > [10:00](https://www.youtube.com/watch?v=bmXj16lJDjE&t=600s) 目前全网已有超过 50% 的互联网流量属于“机器对机器”（Machine-to-Machine, M2M）通信，且随着机器间自动化事务的爆发，这一比例只会不断飙升。 — Jim Fowler > [14:14](https://www.youtube.com/watch?v=bmXj16lJDjE&t=854s) 过去跨越全美配置一条专线需要网络工程师手动调试 20 台路由器的时代已经终结。现代网络必须以‘网络即服务’的形式运行。 — Jim Fowler > [17:09](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1029s) 如果企业仍依赖大量处于停产或停止支持状态的陈旧硬件，AI 将轻而易举地找出这些“盔甲上的裂缝”（chink in the armor）并发动入侵。 — Jim Fowler > [17:09](https://www.youtube.com/watch?v=bmXj16lJDjE&t=1029s) 单靠模型自身的智能是远远不够的，我们必须以极低的延迟和超高的带宽，将人类、机器与海量数据紧密连接在一起。 — Jim Fowler ## Mentioned Jim Fowler, Lumen, Steve Martin, 《双脑人》 (The Man with Two Brains), NCR, Walmart, Target, Kroger, 《2001太空漫游》 (2001: A Space Odyssey), Alkira, AWS, Google Cloud, Microsoft Azure, Oracle Cloud ## Video description Jim Fowler - Chief Technology and Product Officer, Lumen Everyone talks about AI models as the brain. But brains don’t matter without a nervous system. The real question is: who builds and controls the infrastructure that lets AI sense, decide, and act in real time. That’s where telecom moves from connectivity provider to AI infrastructure company. I’ll talk about what it really means to build the nervous system for AI, and why the companies that get this right will define what AI can actually do. [Source](https://www.youtube.com/watch?v=bmXj16lJDjE)

Source quotes

[0:04] Jim Fowler: 模型的能力完全取决于喂养它的燃料，而这种燃料就是数据。 [2:42] Jim Fowler: AI 模型本身正在迅速走向“商品化”（Commodity）。 [2:42] Jim Fowler: 如果它不能连接到车载 IoT 传感器、工厂车间的机械设备或零售店的收银机，它就无法接收任何感知数据，自然也无法做出任何实际行动。 [10:00] Jim Fowler: Cloud 1.0 将算力与存储彻底“云化”和“可编程化”，但广域网（WAN）却被遗忘在接线机房中，处于物理割裂状态。进入 Cloud 2.0 时代，核心使命就是将广域网络全面纳入可编程的云化体系中。 [10:00] Jim Fowler: 目前全网已有超过 50% 的互联网流量属于“机器对机器”（Machine-to-Machine, M2M）通信，且随着机器间自动化事务的爆发，这一比例只会不断飙升。 [14:14] Jim Fowler: 过去跨越全美配置一条专线需要网络工程师手动调试 20 台路由器的时代已经终结。现代网络必须以‘网络即服务’的形式运行。 [17:09] Jim Fowler: 如果企业仍依赖大量处于停产或停止支持状态的陈旧硬件，AI 将轻而易举地找出这些“盔甲上的裂缝”（chink in the armor）并发动入侵。 [17:09] Jim Fowler: 单靠模型自身的智能是远远不够的，我们必须以极低的延迟和超高的带宽，将人类、机器与海量数据紧密连接在一起。

Source quotes

[0:04] Jim Fowler: Everyone talks about AI models as the brain. But brains don’t matter without a nervous system. [0:04] Jim Fowler: The real question is: who builds and controls the infrastructure that lets AI sense, decide, and act in real time. [0:04] Jim Fowler: An AI model is merely a brain, and that brain is only as intelligent as the data feeding it. [0:04] Jim Fowler: The model layer is commoditizing; real enterprise value depends on the nervous system closing the loop between data collection and execution. [11:25] Jim Fowler: Cloud 1.0 automated and virtualized compute and storage into programmatic, on-demand APIs, but it left the Wide Area Network (WAN) behind, stranded in physical meet-me rooms and manual router configurations. [11:25] Jim Fowler: Cloud 2.0 must bring the network into the programmable era [16:51] Jim Fowler: Network providers are no longer just selling static pipes; they are operating the biological equivalent of a real-time nervous system that bridges the gap between sensing, deciding, learning, and acting. [16:51] Jim Fowler: single pane of glass [18:02] Jim Fowler: Hyperscalers and large enterprises are no longer purchasing bandwidth incrementally; they are securing multi-terabit capacity. [18:02] Jim Fowler: chink in the armor [18:57] Jim Fowler: intelligence alone, though, is not enough. [18:57] Jim Fowler: Modern customers are demanding software-defined architectures capable of delivering sub-5-millisecond latency to every one of their distributed data centers.

Original transcript

[0:04] My name is Jim Fower. I'm the Chief Technology and Product Officer for Lumen and for those of you who [0:08] don't know who Lumen is, uh we are one of the world's largest providers of telecommunications [0:13] fiber. Uh we have one of the largest fiber networks in the world. We have one of the most [0:16] connected and most peered networks in the world. One of the things that's kind of struck me today [0:21] as I've been going around to the different sections is all the discussion on models. [0:26] About how the models are getting bigger, faster, talking about the benchmarks. [0:31] Talking about how the models are going to be the end of the world or the end of jobs as we know it. [0:35] Um but I think there's a piece of the conversation that's really missing that I'm going to try to [0:39] cover today and that's really that the model is only as good as the fuel that feeds it and [0:44] that fuel is data. And I think one of the most underlooked components of what's happening in AI [0:49] today is the sprawl of data. So, I'm going to take a a few moments and I'm going to try to talk about [0:55] AI and if AI is the brain, uh who owns the nervous system that connects all of the data, all of the [1:02] information that AI needs to actually operate effectively? Um we all know where this is going, [1:08] right? AI is going to change everything that we know about the way that work today. We're going to [1:13] do what Okay. We're seeing that today from a software development perspective where we've [1:19] trained millions of lines of code into models and we can now infer all of the code that sits in our [1:24] code bases and and have agents write code for us going forward. We're going to see new customer [1:29] experiences. Many of us are already experimenting with AI driven agents that are answering the phone [1:35] and sound and act like a human. They interact with systems on the back end and trained with [1:40] hundreds of thousands of hours of call center data that they're using to be able to understand how to [1:45] deal with a customer call when it comes in. We're seeing new business models. You've got companies [1:49] like Walmart and Target and Kroger that are out there thinking about how is your refrigerator [1:55] going to process an order for milk when you run out of it this week without you knowing [1:59] it. Like they're thinking about the way that those transactions are going to flow back and forth over [2:03] the network, over the nervous system of of AI. And we're seeing breakthroughs in science that [2:09] are really driven by tomes of data that have been collected over years where we're going back and [2:14] finding discoveries that are buried in research done 20 years ago and using it to solve problems [2:20] that have gone unsolved forever. Right? This this AI model that we're talking about the the [2:25] the brain, the AI, the model that's getting a lot of the billing is really only as smart as the data [2:32] that you can feed into it. And one of the biggest problems about the data you feed into it is how [2:36] do you get it all there in a in a timely manner so that it can act upon it. Um I'm going to apologize [2:42] now cuz I was watching uh A Man with Two Brains for any uh Steve fans out there. So, I apologize [2:49] I got this in my mind and I thought this was a good analogy of what's happening right now. It's [2:53] if you've got a brain, you think about the model is the brain and you put a brain in a in a jar, [2:58] it really can't do much. The human brain, if it's not connected to eyes, it can't see. Hands, [3:03] it can't touch. It can't feel. It can't act. Mouth, it can't talk. It can't breathe. Feet, [3:08] it can't walk. Right? A brain that's disconnected from the rest of the body that doesn't have a [3:12] nervous system that connects all of the sensing data really isn't that valuable to us. And AI is [3:17] really not that different. A model can be the smartest model in the world, but if it's not [3:22] connected to the IoT sensors that are sitting on vehicles or it's not connected to the machinery on [3:27] the plant floor or connected to the cash registers in the in the in the retail establishment, [3:32] it's not getting that sensed data to be able to take action. You need a nervous system. [3:38] You need a body that can actually pull into it. I would also argue, since you knew you were coming [3:44] to more of a networking discussion, that the model part of this is actually becoming the commodity. [3:50] That we're seeing the leapfrog event of the models and the model anymore is not the difficult part. [3:55] It's getting the data to the model for the model to do its job. That That is becoming one of the [4:00] most complicated pieces of the world that we work in. And so when you think about what that nervous [4:05] system is for me, there's three basic components to it. The first is sensing. You want to You want [4:10] a nervous system that can sense. When I hold my hand over something hot, I sense that it's hot and [4:15] my mind goes to decide what do I want to do. In the past when I felt something hot, I got burned. [4:21] And so it goes to decide to act. I pull back. Sense, decide, and act. That's the way the human [4:26] body works. That's the way the nervous system of the body works. That's also the way that AI [4:31] is intended to work. We want to be able to sense something to be able to take data that is coming [4:36] at us in real time. We want to send it to a model that's going to decide what to do based on all the [4:42] training that we've given it, all the inference of information that we're feeding into it. [4:46] And we want it to have it take action, to do something for us, to write a piece of code, [4:50] to process a trouble ticket, to to as you heard in the last section if you were here to to maybe [4:54] take an autonomous action to do a a healing action in a network. Those are all examples of [4:59] what the nervous system has to do. It has to close the loop. You have to have all three of these [5:04] elements working for AI to really be valuable. Uh and I'm going to back up for a second. Oops. [5:13] Um to my first job. So my first job out of school was working for NCR and I had the job title of LAN [5:21] administrator. A few of you laughed cuz you're old enough to know. So just know I know that now. I [5:27] was a local area network administrator. My job was to install the software on the Windows PCs that we [5:32] had to be able to connect the PC to the network to be able to do simple things like email, to be able [5:38] to get to file shares, to be able to actually kind of collaborate across the company. It was [5:43] the first network that existed in NCR was the one that I helped put in when I was in that that job. [5:49] As you know, that job title doesn't exist anymore. But when I go back to that job title, [5:54] the role that I thought I was doing, I thought it was really complicated. I thought it was magic. I [5:58] heard them putting the software on this machine and then now can talk to every other machine and [6:02] we can now have shared schedules and email and I remember thinking that was so cool. I I look back [6:07] on that and what I thought was so difficult and daunting at the time, it really wasn't. I mean, [6:12] think about it. All of our workers sat in buildings we owned. Nobody worked from home. [6:19] All of our workers sat in buildings we owned that were connected to a network that we owned [6:23] and managed. They all sat within a perimeter that was protected by the way, by the gates [6:29] of the building. Forget about firewalls because there was no internet connectivity. We weren't [6:33] connected to anything on the outside. Our biggest security risk was somebody taking a disc home, a [6:40] a floppy disc home, putting it in their home machine and bring a virus back in. Right? That [6:44] world that we thought was so complicated, that I thought was so complicated in my first job, [6:49] I look at that now and go, it's not complicated at all. In fact, I wish that was the network nervous [6:55] system that we're working with AI today. It would make life so much easier. Two data centers. Two [6:59] data centers that were literally 100 yards apart from each other. Right? You wouldn't You wouldn't [7:03] have that today. Right? So, that world that I worked in, I look back and I go, "Oh my gosh, I [7:08] really wish we could go back to that because life would be easier." But, that's not today's reality. [7:14] Today's reality in the AI-driven world that we're in first says you've got data, users, [7:18] and apps that are not sitting next to next to each other anymore. Um in fact, all of those users that [7:24] used to sit in the facilities that we owned, they now sit in 50,000 locations around the world many [7:30] times working from a home location or a customer's office location. The data that they're working off [7:37] of is now sprawl. We used to have kind of two big mainframe systems. All the data you ever needed [7:41] was in those two big mainframe systems. I say big compared to today, not, but that that was all the [7:45] data you needed. Today, data sits in hundreds of thousands of databases that are spread across the [7:49] world, right? So, you have that complexity. And applications that we wrote, we owned, we [7:54] understood, are now relegated to SaaS applications where you've got to be able to get the data out [7:59] of the SaaS provider who's running it someplace else around the world, and you may not have any [8:03] clue where that data is that you need to be able to run your business process. And oh, by the way, [8:08] sitting on computer infrastructure that's in multiple clouds that potentially is in a hybrid [8:14] where you've got something sitting in your data center, something sitting in cloud providers in [8:18] multiple regions, in multiple availability zones, in multiple cloud providers around the world. [8:24] And security has gone from worrying about the the person with the bad floppy disk to nation-states [8:31] who are actively trying to break into your network and cause havoc. And they might cause havoc, [8:36] they might steal your data, they might steal your IP, but you are now dealing with a completely [8:40] different landscape. And the complexity that you're working in really now doubles down the [8:46] importance of as you're moving in an AI world, how do you manage all of that infrastructure? And [8:53] and this isn't a future, like this isn't a view of where we're headed. This is where we're at today. [8:58] Today, 90% of enterprises use multiple clouds. 90%, right? This I This idea that that I had early [9:07] on on cloud that you'd pick one provider and stick with it. No way. You You can't control it. This [9:12] providers are deciding where they're going to run their applications. You have very little control [9:16] of being able to be in one cloud. So, you're automatically going to be if you're going to be [9:20] cloud, and I don't know why you wouldn't be cloud at this point, you're going to be multi-cloud. [9:24] 100% is going to be the the cager for workload growth in AI. So, AI is going to double in size [9:32] every year between now and 2030. So, we're just at the beginning of of how much data, how much [9:38] transaction is going to be driven in an agentic world where agents are doing more and more of the [9:43] work. So, you're going to see this huge increase in capacity demands for data to be able to flow [9:49] back and forth for those machines to be able to operate. The proof point that you can look at [9:55] to say, "Is that real, Jim, or not?" is the data center growth. Already on the books, we're going [10:00] to see a 10x increase in data centers in the that are built around the world between now and 2030, [10:05] 10x. Right? That's already what's on the books. Like that there's more coming. That's not all that [10:10] we're going to have. And so, that sprawl picture that I kind of drew on the page before of where [10:15] your data sits, it's going to get worse. It's going to be in more places. It's going to follow [10:19] power. It's going to follow water. It's going to follow taxes tax law. Right? That's what's [10:24] going to determine where the next data centers are going to go. And there going to be more of them, [10:28] and they're going to be in located in more locations. That edge is going to get more and more [10:33] distributed. And and this idea that AI back when I started, you know, when I started my career, [10:40] AI was a page in a book in a textbook, and it was an idea in 2001: A Space Odyssey. Right? [10:47] That That was all AI was when I started my career. Today, 50% of the internet traffic is [10:54] machine-to-machine. Now, is that all AI? No, it's machine-to-machine. But it's machine-to-machine. [10:59] And that number's only going to get higher. As we do more transactions where machines are talking [11:03] to other machines, that number's only going to continue to get bigger and bigger. And so, [11:08] I go back to my thesis at the beginning of this. The models are They're really smart, [11:13] but they're only as smart as the ability to be able to get information and data to them [11:18] in a timely way to be able to process it. And that requires a really robust nervous system, [11:25] one that we really have not built yet around the world. This isn't new. Cloud 1.0 kind of proved [11:33] this out for us. So, my second job is I was a Unix server administrator. I actually had to partition [11:39] hard drives. I had to lay down the Unix kernel. I had to build out code. I actually had to swap in [11:43] CDs and disks to be able to install it. When was the last time a systems administrator did that? [11:48] They go out in the cloud environment and they say, "Set up my virtual machine." and it just shows up. [11:53] Like, I don't even know that somebody would know how to partition a drive anymore. Right? All of [11:56] that in Cloud 1.0 got cloudified. It got turned into a programmatic experience where you were able [12:02] to type a few commands and get the compute and the storage that you needed to be able to run your [12:06] business. It added on additional capabilities. You wanted to be able to have a web server, [12:11] it was automatically there. You didn't need to go install the web server. There were value-added [12:15] services that got installed in the cloud. That's what's going to happen in the network. Cloud 2.0 [12:21] is going to bring the network along for the ride because it didn't come. It stayed disconnected. [12:26] It stayed in the meet-me rooms. It was completely disconnected from the cloud. The cloud providers [12:31] managed what they were doing inside their network, but nobody did anything on the wide area network. [12:36] And that's what we believe this this Cloud 2.0 has to be about. It has to be about making the [12:42] network programmable at a wide area network level. So, so what does that nervous system look like? [12:49] Um we think the nervous system of AI has three components to it. It starts with a really robust [12:54] physical network. At Lumen, uh we've been building this out. So, we've upgraded the [13:00] photonics across the United States where we're now at 400 going to 800 gig on our way to 1.6 terabit [13:06] capacity intercity across the entire United States, across 65 major metropolitan areas, [13:12] all the way out by the end of this year to cover 90% of the cloud on-ramps. Why? It's that capacity [13:18] discussion. We see it coming at us faster than we can get ready for it. We see our customers wanting [13:23] to move around large amounts of data. I had a discussion with a CIO of a of a Fortune 5 bank [13:29] about 2 months ago and she was talking about training her own model and she was talking about [13:36] the problem that she was in is she rented the GPUs, but the GPUs were available to her were not [13:40] in the same same place as the data that she needed to train it and she was being gated by the network [13:46] lines. The network didn't have enough capacity to be able to feed the GPUs fast enough and she had [13:51] idle time for the most expensive asset she had because she couldn't get data to it. We have to [13:56] build out a physical network that's able to do that. Today, we have 17 million miles of fiber [14:02] that run across the United States. By the end of 2031, we'll have 58 million miles of fiber that [14:08] are connecting all of these data centers, really preparing for that future growth. The second piece [14:14] of of building out the the the nervous system is really you have to make it programmable. [14:20] The days of touching 20 routers to be able to set up a connection that goes across the United States [14:26] are over. It needs to work like the cloud. You need network as a service where you can consume [14:32] it, where you can control it, where you can spin up and spin down the capacity that you need, where [14:36] you can create the point-to-point connectivity you need, all defined by software that executes [14:42] the network on the back end. So, it first has to be programmable. It has to be a multi-carrier [14:48] mesh network. Lumen is not the only carrier that's going to be needed to be able to make this work. [14:53] Our competitors have to be part of that dialogue. There needs to be a cross-carrier mesh network [14:59] that makes all of this work and we're building that out. We actually just recently finished the [15:03] acquisition of a company called Alkira who did this really magnificent thing of building out a [15:08] single control plane that allows a customer to be able to create cloud-to-cloud connectivity, [15:13] cloud-to-premise connectivity through software. They they changed their whole point of view of [15:19] saying instead of a pop or a gateway center being a wiring center someplace, we're actually going [15:24] to think about it as being logically inside of a cloud. So, in AWS, in Google, in Azure, [15:29] in Oracle Cloud, they actually repositioned the the the pop, the point of presence, to be virtual [15:36] sitting inside a cloud provider. That's how you're going to make this multi-carrier, that's [15:40] how you're going to make this be programmable, is to be able to have it to be software defined. [15:45] And then lastly, it has to have the ecosystem partners. Remember the discussion I had about when [15:49] you go to the cloud, you don't lay down your web servers anymore? Then why are we still installing [15:54] firewalls and load balancers? And why are we still putting our own DNS capabilities in place? [15:59] Why is that not also just a software function of what you put down when you lay out your wide area [16:05] network? That's the last piece that has to be in place to be able to have the nervous system, and [16:11] that's frankly what we're building. We're building a network that is programmable, that has a single [16:16] control plane, that's cross-carrier, that's cross-cloud, that allows for connectivity, and [16:22] we're really focused on cloud-to-cloud. When you look at the marketplace today, premise-to-cloud, [16:28] which was that world that I built, you know, 30 years ago when I first got in, that market's [16:33] only growing at about 1% a year. Why? More people from home, office space is reducing. [16:39] That's not where the growth is. Where the growth is is cloud-to-cloud, data center to data center, [16:44] growing at about a 13% growth rate on on being able to put the connectivity between the machines. [16:51] And as Aquantik takes off, that number we believe is going to grow. And so, our focus is how do we [16:56] make that as easy as possible for our customers to be able to create those connections in real time, [17:02] spin them up, spin them down, and be able to move forward from a a nervous system perspective. [17:09] So, maybe where I leave you with is what what I think you'd expect in an AI-driven [17:14] world when When thinking about what the future of networking looks like, one, I think you got to you [17:20] got to take the days of point-to-point connectivity and say those are done. [17:25] You're no longer running a point-to-point connection business, you're running a nervous [17:29] system for AI. You are the You are the connection between the ability to sense and act and learn [17:37] that is happening from an AI perspective. You should be able to have on-demand consumption. You [17:41] should be able to spin up and spin down capacity as you need it. You should be able to have control [17:46] over everything in one place and not just the network connectivity, the services that sit around [17:51] it, the firewall services, the security services, the load balancing services. They should all be [17:56] configurable from a single pane of glass that you configure and run. It should be built for AI. [18:02] And when I say built for AI, that 400 gig going to 800 gig going to 1.6 terabit, that's real. [18:10] We just took an order this We just took an order 2 weeks ago from a hyperscaler that's putting in 400 [18:16] 100 gig circuits. I mean, we're talking We're taking orders in the terabits of size at this [18:22] point. They're already planning for this data connectivity. The enterprises are starting to [18:27] also put those types of orders in place. You have to have infrastructure that's built out for that [18:32] level of capacity. And you've got to take the security real. Um if you are thinking about [18:38] the next generation of models they're already being used to find the deficiencies not just in [18:42] the software that runs your business, but the networking equipment that runs your business. [18:45] And so, if you're sitting out there with a ton of legacy end-of-life end-of-support hardware, [18:50] I guarantee that those next models are looking for the [ __ ] in the armor to be able to break into [18:56] your network. And so, maybe I'll leave you with this as my time is almost up. One final thought. [19:02] Today, everything is really competing to be smart and wants to be smarter with AI. Intelligence [19:08] alone, though, is is not enough. We have to connect the people, the machines, and the data [19:15] with very low latency and very high capacity. And that infrastructure looks very different [19:22] than the Telco world that many of us grew up in. This is going to be software defined. It's [19:28] going to be high capacity. It's going to be low bandwidth. I'm having conversations with customers [19:33] who want to know that I have capacity with that I have capability to be within 5 milliseconds [19:38] of every one of their data centers. That's what they're asking me to think about as we contract [19:42] for what the next level of connectivity looks like. So, I hope this was helpful. [19:45] This is how we're thinking about what the future of connectivity looks like in an AI-driven world. [19:50] I hope you get a lot out of the next few days. Thanks very much. Appreciate it.