Meta Deploys AI Agents to Build Network Infrastructure as Power Hits 7 GW

@Scale

Meta uses domain-expert AI agents across its plan, design, deliver, and run pipeline to resolve the divergence gap between flat engineering headcount and exponential infrastructure complexity.

AI-assisted summary of Lighting Talk Intro: AI for NI | Mohab Gawish, Meta — original source. Check the original for full context.

Source published: 2026-09-13

Detailed notes

## Background This lightning talk was presented by Mohab Gawish from Meta and published by @Scale on YouTube on August 27, 2026. The video description invites viewers to learn more about the @Scale conferences via their official website.

## TL;DR Faced with a 5x explosion in network power footprint from 1.3 GW to 7 GW alongside flat engineering headcount, Meta addresses this 'divergence gap' by deploying domain-driven AI agents across their capacity buildout pipeline. Rather than relying on static scripts, Meta utilizes autonomous AI tools built directly by network engineers to manage multiplicative complexities across heterogeneous GPU generations and environmental variations. The talk outlines Meta's four-phase capacity pipeline—Plan, Design, Deliver, and Run—setting the stage for a series of deep-dive lightning talks by domain leads.

## Atmosphere & style The presentation is crisp, authoritative, and deeply technical, driven by a forward-looking, pragmatic engineering tone. Gawish delivers a focused, structured framing that bridges high-level strategic vision with grounded operational realities. The atmosphere feels fast-paced and efficient, setting an energetic baseline for an enterprise tech talk that replaces marketing hype with concrete infrastructure architecture.

## Key takeaways - [00:05](https://www.youtube.com/watch?v=nZgru_53Ask&t=5s) Meta's network capacity buildout is guided by the thesis that the infrastructure powering AI is increasingly delivered by AI. - [00:05](https://www.youtube.com/watch?v=nZgru_53Ask&t=5s) Capacity management at Meta is an end-to-end pipeline that transforms raw demand signals into live, operational network capacity. - [01:09](https://www.youtube.com/watch?v=nZgru_53Ask&t=69s) Meta's physical network power footprint experienced rapid expansion, scaling from 1.3 GW to 7 GW (~5x increase). - [01:09](https://www.youtube.com/watch?v=nZgru_53Ask&t=69s) Hardware heterogeneous complexity creates operational challenges, with multi-generational GPUs (N, N+1, N+2) coexisting simultaneously in production. - [01:09](https://www.youtube.com/watch?v=nZgru_53Ask&t=69s) Shifting environmental and physical constraints (power density, cooling technology, spatial footprints) add non-uniformity to infrastructure buildout. - [03:03](https://www.youtube.com/watch?v=nZgru_53Ask&t=183s) The 'divergence gap' describes the critical disconnect between exponential buildout complexity and flat engineering headcount. - [03:03](https://www.youtube.com/watch?v=nZgru_53Ask&t=183s) Traditional software abstractions, modeling, and simple script automation are no longer sufficient to bridge the divergence gap. - [03:03](https://www.youtube.com/watch?v=nZgru_53Ask&t=183s) AI-native, autonomous agents act as a force multiplier for human engineering expertise within the capacity management pipeline. - [04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) The capacity pipeline consists of four sequential phases: Plan, Design, Deliver, and Run. - [04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) The Plan phase converts raw demand signals and power availability into committed long-term investments and capacity targets. - [04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) The Design phase translates high-level planning signals into physical/logical network topologies, cabling, and rack blueprints. - [04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) The Deliver phase handles procurement, installation, provisioning, configuration, and validation to bring hardware online. - [04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) The Run phase manages ongoing operational health and maintenance across multi-generational hardware and software architectures. - [04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) Variability across the pipeline phases multiplies exponentially (represented by a multiplication sign between phases) rather than adding linearly. - [08:21](https://www.youtube.com/watch?v=nZgru_53Ask&t=501s) AI agents are authored directly by experienced network domain experts rather than isolated, central AI teams. - [08:21](https://www.youtube.com/watch?v=nZgru_53Ask&t=501s) Upcoming lightning talks by Ben Valis, Maz Mohammadi, Mohab Gawish, and Yuling Chen will detail the AI agents used in Plan, Design, Deliver, and Run phases respectively.

## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=nZgru_53Ask&t=5s) Intro & Setting the Frame: AI-Powered Infrastructure Buildout

At Meta, the team responsible for scaling network infrastructure faces a massive challenge: capacity buildout. Introducing a series of lightning talks on how Meta leverages AI to build its network infrastructure, presenter Mohab Gawish frames the overarching vision driving their work. The central motif of the entire talk series centers around one compelling reality: "The infrastructure that powers AI is increasingly delivered by AI." Gawish clarifies that while this phrase serves as a catchy theme, it is by no means just a marketing slogan—it has become an absolute operational necessity for modern tech companies.

To set the stage, Gawish establishes a clear definition of what capacity management actually means within Meta's ecosystem. It is the comprehensive, end-to-end pipeline that takes a raw demand signal and transforms it into a fully live, healthy, and operational network capacity.

- **Key Slogan & Thesis:** "The infrastructure that powers AI is increasingly delivered by AI." - **Scope of Capacity Management:** An end-to-end pipeline converting raw demand signals into live operational network infrastructure.

[01:09](https://www.youtube.com/watch?v=nZgru_53Ask&t=69s) The Scale Challenge vs. Compounding Complexity

To illustrate the sheer magnitude of Meta's growth, Gawish highlights the massive expansion of their physical footprint over the past few years. Meta's network infrastructure power footprint exploded from 1.3 gigawatts to 7 gigawatts—representing roughly a 5x increase. Standard engineering assumptions might suggest this is a straightforward, linear scale-out problem (simply deploying 5x more racks, 5x the network fabrics, and 5x the number of ports). However, in the context of rapidly advancing AI, that headline number conceals a much darker reality: the gigawatts being deployed are not interchangeable, and the underlying systems are far from uniform.

In the fast-evolving AI landscape, every single system axis varies simultaneously. Hardware relies on multiple generations of GPUs (e.g., $N$, $N+1$, $N+2$) coexisting in production at the exact same time. Each generation differs significantly in speed, memory capabilities, network interface specs, and NIC capabilities. Meta does not have the luxury of finishing one migration before the next hardware generation arrives. Furthermore, physical infrastructure parameters—such as liquid/air cooling capabilities, power density, and physical floor space—are constantly shifting, creating compounding variations in the physical buildout blueprint.

- **Power Footprint Scaling:** Expanded from **1.3 GW to 7 GW** (~5x growth). - **GPU Heterogeneity:** Coexistence of $N$, $N+1$, and $N+2$ hardware generations in live production without waiting for legacy migrations to complete. - **Environmental Variables:** Shifting power density, cooling infrastructure, and physical spatial constraints.

[03:03](https://www.youtube.com/watch?v=nZgru_53Ask&t=183s) The Divergence Gap and AI-Native Solutions

When mapping out these dynamics, a clear operational rift emerges: while power footprint grows linearly, infrastructure complexity grows exponentially. Crucially, engineering headcount remains relatively flat, as teams cannot realistically scale headcount in lockstep with gigawatts. Gawish defines this critical disconnect between exponential task complexity and flat staffing as the "divergence gap." Standard software abstractions, traditional modeling, and basic script automation are no longer sufficient to bridge this divide.

To solve this, Meta turned to autonomous, AI-native capabilities. Instead of treating AI merely as an external tool, they integrate AI agents directly into the pipeline to act as a force multiplier for human expertise. This creates a virtuous feedback loop: AI workloads drive massive capacity demand; that demand requires networks to be built at unprecedented scale; and specialized AI agents help design, deliver, and run those very networks.

- **The Divergence Gap:** The widening rift between linear power growth, exponential buildout complexity, and flat engineering headcount. - **Strategic Pivot:** Moving beyond traditional software tools and static automation toward autonomous, AI-native agent workflows.

[04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) The Four-Phase Capacity Pipeline (Plan, Design, Deliver, Run)

Meta’s capacity management pipeline is structured into four sequential phases, where each phase supplies critical inputs to the next:

1. **Plan:** Raw demand signals (such as XPU roadmaps, product forecasts, and regional power availability) are converted into committed investments on how many megawatts or gigawatts to deploy, specifying locations, spaces, and targeted execution quarters years in advance. 2. **Design:** Translates planning signals into physical and logical blueprints. This phase determines network topology, rack layouts, cabling schemas, and port mapping under defensive assumptions that every capacity mix is heterogeneous. 3. **Deliver:** Covers procurement, hardware installation, network provisioning, configuration, commissioning, and final validation to convert design blueprints into live traffic-carrying capacity. 4. **Run:** Encompasses all operational activities required to maintain fleet health across multi-generational hardware, software, and network architectures.

Gawish highlights a subtle structural detail in their pipeline model: rather than drawing simple directional arrows between phases, Meta places a **multiplication sign ($\times$)** between them. This denotes that variability does not merely add up across phases—it multiplies exponentially. For instance, a single new design blueprint creates entirely new vectors for configuration drift, software tooling breakage, and custom edge-case errors downstream.

- **Phase Breakdown:** **Plan** $\rightarrow$ **Design** $\rightarrow$ **Deliver** $\rightarrow$ **Run**. - **Multiplicative Complexity:** Variations in early phases compound exponentially in subsequent steps (signified by $\times$ rather than simple linear arrows).

[08:21](https://www.youtube.com/watch?v=nZgru_53Ask&t=501s) Domain-Driven AI Agents & Lightning Talk Overview

A key factor in the success of these AI agents is how they were built. Meta explicitly avoided handing the project over to an isolated, central AI team. Instead, the AI agents operating at each phase were authored directly by domain experts—the network engineers who have personally executed these tasks for years. As a result, each agent directly embodies the nuanced domain expertise and operational knowledge of its author.

Gawish concludes by introducing the upcoming presentations in the lightning talk series, where individual domain leads will deep-dive into how AI agents are applied within each specific phase:

- **Plan Phase:** Deep dive presented by **Ben Valis**. - **Design Phase:** Deep dive presented by **Maz Mohammadi**. - **Deliver Phase:** Deep dive presented by **Mohab Gawish**. - **Run Phase:** Deep dive presented by **Yuling Chen**.

## Notable quotes > [00:05](https://www.youtube.com/watch?v=nZgru_53Ask&t=5s) The infrastructure that powers AI is increasingly delivered by AI. — Mohab Gawish

> [03:03](https://www.youtube.com/watch?v=nZgru_53Ask&t=183s) divergence gap — Mohab Gawish

## Mentioned Meta, Mohab Gawish, Ben Valis, Maz Mohammadi, Yuling Chen, @Scale

[Source](https://www.youtube.com/watch?v=nZgru_53Ask)

中文简报

功耗达7 GW之际，Meta部署AI智能体构建网络基础设施 Meta在其规划、设计、交付和运行全流程中采用领域专家级AI智能体，以弥合工程人员规模持平与基础设施复杂性呈指数级增长之间的分歧差距。面对网络电力规模从1.3 GW激增至7 GW（增长约5倍）而工程人员编制保持不变的现状，Meta通过在其容量建设流程中部署领域驱动的AI智能体来解决这一“分歧差距”。Meta不再依赖静态脚本，而是利用由网络工程师直接构建的自主AI工具，来管理异构GPU代际并存及环境变化带来的多重复杂性。该演讲概述了Meta涵盖规划（Plan）、设计（Design）、交付（Deliver）和运行（Run）的四阶段容量流程，为各领域主管后续的一系列深入快闪演讲拉开序幕。 Meta 提出“驱动 AI 的基础设施正越来越多地由 AI 本身交付”的核心理念，将其作为端到端网络容量管理的关键运营原则。 (0:05) Meta 网络基础设施的物理电力规模从 1.3 GW 快速扩张至 7 GW（约 5 倍增长），同时生产环境中多代 GPU（N、N+1、N+2）与多种物理环境限制并存，导致异构复杂性急剧上升。 (1:09) 面对指数级增长的建设复杂度与保持平稳的工程人员规模之间的“分歧差距（divergence gap）”，Meta 放弃传统脚本与静态工具，转向将 AI 原生自主 Agent 直接引入交付管道。 (3:03) Meta 将容量管理划分为规划（Plan）、设计（Design）、交付（Deliver）与运行（Run）四个阶段，各阶段之间的变量呈乘数级级联放大，且相应 AI Agent 均由各领域的一线网络工程师亲自编写。 (4:12)

Reading summary

According to Meta's Mohab Gawish, managing the company's surging network infrastructure has become an operational necessity that requires AI itself, as compounding technical complexity outpaces flat engineering headcount. Footprint Growth and Compounding Hardware Heterogeneity (1:09) Meta's network infrastructure power footprint expanded roughly fivefold, scaling from 1.3 gigawatts to 7 gigawatts over recent years. While this growth might appear to be a linear buildout of additional racks and ports, Gawish notes that deployed gigawatts are not interchangeable and underlying systems lack uniformity. In live production, multiple hardware generations—such as N, N+1, and N+2 GPUs—must coexist concurrently without waiting for prior migrations to finish. Each generation introduces distinct memory capabilities, network interface specs, and speeds, while physical conditions like cooling methods (air versus liquid), power densities, and space constraints continuously shift. The Divergence Gap and Autonomous Agents (3:03) Because power demand grows linearly while infrastructure complexity grows exponentially, Meta faces what Gawish terms a 'divergence gap' against relatively flat engineering headcount. Traditional software modeling, script automation, and standard abstractions proved insufficient to keep pace. To bridge this divide, Meta integrates AI-native agents directly into operational workflows rather than treating AI as an external utility. This establishes a feedback loop where AI workload demands trigger massive network expansions, and specialized AI agents help design, deploy, and operate the resulting networks alongside human engineers. Multiplicative Complexity Across Capacity Phases (4:12) Meta organizes its capacity management pipeline into four sequential phases: Plan (translating raw demand into long-term megawatt or gigawatt commitments), Design (generating logical and physical blueprints such as cabling schemas and topologies), Deliver (procuring, installing, provisioning, and validating live capacity), and Run (maintaining operational fleet health). Gawish emphasizes that these stages relate through multiplication rather than simple linear addition. Variations introduced during the planning or design stages compound downstream, spawning new configuration drifts, edge-case errors, and software tooling failures throughout the delivery and run phases. Domain-Authored AI Implementation (8:21) Meta intentionally avoided delegating agent creation to a centralized, isolated AI team. Instead, the autonomous agents deployed throughout the pipeline were authored directly by practicing network engineers who possessed firsthand operational experience in each specific stage. The presentation frames a broader series of talks featuring Ben Valis on planning, Maz Mohammadi on design, Gawish on delivery, and Yuling Chen on run operations, aimed at detailing how specialized agents handle each phase of the capacity pipeline.

精炼阅读摘要

Meta 工程师 Mohab Gawish 指出，“支撑 AI 的基础设施正日益由 AI 本身交付”已成为当下的运营必然；Meta 正在通过将自主 AI 智能体直接嵌入端到端网络容量流水线，以应对物理规模剧增与异构系统交织所带来的系统性挑战。规模扩张与系统异构性挑战 (1:09) Gawish 介绍，Meta 网络基础设施的功耗规模在过去几年中从 1.3 吉瓦（GW）迅猛增长至 7 吉瓦，增幅约为 5 倍。常规工程视角常将此类扩张视作线性堆叠机架、结构网和端口的简单问题，但 AI 基础设施的技术演进使这一逻辑失效。Meta 部署的功耗规模并不具备可替代性与均质性，整个硬件层面上存在多重系统维度的动态交织。在生产环境中，N、N+1 以及 N+2 等多代 GPU 硬件同时并存运行，且各代设备在计算速度、内存容量、网络接口规范及网卡（NIC）性能上差异显著，工程团队无法等待前代迁移彻底完成后再引入新架构。与此同时，液冷与风冷散热、供电密度及机房物理空间等基础设施参数也在持续演变，导致物理构建方案的变异度不断累积加深。工程人力与复杂度脱节的“分歧差距” (3:03) 随着物理足迹的线性扩张，基础设施的构建复杂度呈现指数级增长，而工程团队的人员编制受现实条件约束基本保持平稳。Gawish 将指数级任务复杂度与扁平化人力规模之间的结构性失衡定义为“分歧差距”（Divergence Gap）。传统软件抽象层、静态建模以及常规脚本自动化已无法填补这一日益扩大的鸿沟。为了解决该矛盾，Meta 转向了原生 AI 自主能力的工程化应用。团队将 AI 智能体直接集成到作业流程中作为人类专业知识的效能放大器，从而构建正向反馈回路：AI 业务负载驱动超大规模容量需求，前所未有的网络规模催生了对专用 AI 智能体的依赖，而智能体反过来协助完成这些网络系统的设计、交付与运行。四阶段容量流水线与乘法效应 (4:12) Meta 将容量管理定义为将原始需求信号转化为健康可用网络容量的端到端管道，并划分为四个相扣阶段：规划（Plan）负责将芯片路线图、产品预测与区域供电可用性转变为未来数年的兆瓦/吉瓦投资与空间承诺；设计（Design）负责将规划信号映射为网络拓扑、机架布局、布线及端口映射等物理与逻辑蓝图；交付（Deliver）覆盖采购、安装、网络调优、配置配置、调试上线及最终验证；运行（Run）则涵盖维持多代软硬件架构健康度所需的一切日常运维。 Gawish 特别指出，各阶段之间的流转并非单纯的加法连接，而是呈现“乘法（×）”效应。前序阶段产生的微小变化或新设计蓝图，不会仅仅增加孤立变量，而是会在下游呈指数级放大，催生出大量配置漂移、软件工具链断裂及边缘特例错误。领域专家主导的智能体落地架构 (8:21) 在实施路径上，Meta 明确避免了将系统全权交由中心化通用 AI 团队研发的模式。相反，嵌入各阶段的 AI 智能体均由多年深耕一线的网络工程师直接编写与调优。这种模式确保了智能体能够完整继承并内化对应专业领域的操作细节与实战经验。系列技术演讲进一步展示了该架构在各个具体环节的分工部署：Ben Valis 负责深度解析规划（Plan）阶段的智能体应用，Maz Mohammadi 负责设计（Design）环节，Mohab Gawish 主讲交付（Deliver）流水线，而运行（Run）保障阶段的落地实践则由 Yuling Chen 进行剖析。 Gawish 强调，AI 驱动的网络容量管理已由前瞻探索转变为不可或缺的生产基石。然而，面对多代硬件混合部署与下游乘数级复杂度的累加，一线领域工程经验能否通过 AI 智能体完全实现自动化闭环，仍有待后续各专项落地数据的持续检验。

Detailed notes — zh

## Background 该视频由 YouTube 频道 @Scale 于 2026 年 9 月 3 日发布，内容为 Meta 工程师 Mohab Gawish 在 @Scale 大会上的闪电演讲开场引言。视频简介提供了 @Scale 大会官方网站链接，整体聚焦于 Meta 如何在超大规模网络基础设施（NI）建设中深度引入人工智能。 ## TL;DR 在 AI 浪潮驱动下，Meta 网络基础设施的电力足迹从 1.3 GW 激增至 7 GW，但多代异构硬件的并存与物理环境限制导致系统复杂度呈指数级爆发，与平缓的工程团队人力形成了严峻的“分歧鸿沟”（Divergence Gap）。Meta 将容量交付解构为规划（Plan）、设计（Design）、交付（Deliver）与运营（Run）四个阶段，指出各环节的未知数以乘法效应级联放大。为应对这一挑战，Meta 推动了“由 AI 亲自交付 AI 基础设施”的正向飞轮闭环，依靠深耕业务一线的网络工程师打造针对各阶段的专用 AI Agent 矩阵，从而突破规模化扩容的吞吐极限。 ## Atmosphere & style 演讲整体呈现出一种务实、冷峻且极具压迫感的硬核工程叙事风格。演讲者摒弃了宽泛浮夸的 AI 宣传话术，开门见山地用残酷的真实数据拆解超大规模算力落地时的工程混乱度与治理瓶颈，语调沉稳而坚定。整场分享洋溢着顶级技术团队在极限系统重压下主动突围的工程师思维，既透露出直面指数级复杂度的清醒审慎，又展现出以领域专家经验具象化为 Agent 破局的秩序感与决断力。 ## Bulletin - [00:05](https://www.youtube.com/watch?v=nZgru_53Ask&t=5s) Meta 网络基础设施电力足迹在数年间从 1.3 GW 飙升至 7 GW（增长近 5 倍），但工程团队人数基本持平，倒逼基础设施构建全面引入 AI 自主交付能力。 - [01:47](https://www.youtube.com/watch?v=nZgru_53Ask&t=107s) 不同于传统的可预测线性扩展，AI 时代数据中心面临多代 GPU/NIC 混部与物理环境制约，形成了系统复杂度指数爆发与人力持平之间的“分歧鸿沟”。 - [04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) 容量交付被定义为“规划（Plan）、设计（Design）、交付（Deliver）、运营（Run）”四阶段，各阶段的不确定性与变异度呈乘法效应级联扩散。 - [07:35](https://www.youtube.com/watch?v=nZgru_53Ask&t=455s) Meta 的破局方案是由扎根业务一线的网络工程师主导开发各阶段的 AI Agent，将专业排障经验实体化以形成生产力正向飞轮。 ## Key takeaways - [00:05](https://www.youtube.com/watch?v=nZgru_53Ask&t=5s) 核心工程命题是“支撑 AI 的基础设施正日益由 AI 自身来交付”，这是应对超大规模工程复杂度的必然路径。 - [00:05](https://www.youtube.com/watch?v=nZgru_53Ask&t=5s) Meta 网络电力足迹从 1.3 GW 暴增到 7 GW，但宏观数字掩盖了工程异构性所带来的混乱度，无法按线性扩展模型应对。 - [00:05](https://www.youtube.com/watch?v=nZgru_53Ask&t=5s) 容量管理的本质是一整套端到端流水线，涵盖从最初的需求信号输入到最终交付在线、健康且具备业务承载能力的网络容量全流程。 - [01:47](https://www.youtube.com/watch?v=nZgru_53Ask&t=107s) 数据中心内部 $N$、$N+1$ 与 $N+2$ 代计算硬件（GPU/XPU/NIC）长期并存，迁移窗口永久重叠，硬件无法相互替代。 - [01:47](https://www.youtube.com/watch?v=nZgru_53Ask&t=107s) 供电密度、空间布局以及散热制冷（液冷/风冷）的多样性与硬件异构性耦合，导致网络交付蓝图发生组合爆炸。 - [01:47](https://www.youtube.com/watch?v=nZgru_53Ask&t=107s) “分歧鸿沟”（Divergence Gap）体现为系统复杂度呈指数级爆发，而工程团队人数保持持平，传统脚本与抽象工具已达极限。 - [04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) 规划阶段（Plan）需要提前数年做下注决策，将粗糙的产品预期和电力供应信号转化为特定园区和节点的容量承诺。 - [04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) 设计阶段（Design）确立了“默认每批容量都具备异构性”的防御性假设，必须定制兼顾物理拓扑与逻辑路由的工程蓝图。 - [04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) 四阶段之间以“乘号（$ imes$）”级联，前置阶段遗留的单个未知假设会在后续交付和运营阶段引发配置漂移与工具链失效的雪崩。 - [07:35](https://www.youtube.com/watch?v=nZgru_53Ask&t=455s) 构建与运营流程中的 AI Agent 充当了力量放大器（Force Multiplier），形成了业务需求与网络交付相互促进的正向工程飞轮。 - [07:35](https://www.youtube.com/watch?v=nZgru_53Ask&t=455s) 流水线中部署的 AI Agent 均由具备一线实战经验的资深网络工程师打造，而非孤立的中心化算法团队，确保了深层工程经验的代码化。 ## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=nZgru_53Ask&t=5s) 核心命题：支撑 AI 的基础设施正由 AI 亲自交付 Mohab Gawish 开宗明义地拉开了 Meta 系列闪电演讲的序幕，探讨 Meta 如何在网络基础设施（Network Infrastructure, NI）的大规模建设中深度引入人工智能。本次系列的核心主题被凝练为一句掷地有声的断言：“支撑 AI 的基础设施，正日益由 AI 自身来交付”（"The infrastructure that powers AI is increasingly delivered by AI"）。Gawish 强调，这绝非一句动听的宣传口号，而是工程现实逼迫下的必然选择。在这里，所谓的“容量管理”（Capacity Management）并非抽象的资源调度，而是一整套端到端的基础设施流水线——它必须能够接收最原始的业务需求信号，并将其扎扎实实地转化为在线、健康且具备完整运营能力的网络容量。在过去短短几年间，Meta 的网络基础设施电力足迹从 1.3 吉瓦（GW）飙升至 7 吉瓦（GW），实现了约 5 倍的庞大跨越。对传统网络工程师而言，听到“5 倍增长”往往会本能地将其理解为可预测的线性水平扩展（scale-out）：无非是机架数量增加 5 倍、网络架构（fabrics）扩大 5 倍、交换机端口与光纤数量同比例扩容。然而在生成式 AI 狂飙突进的技术背景下，这个宏观的数字掩盖了极其严酷的工程现实——实际面对的复杂度和工程混乱度，远远超出了任何线性扩展模型的预测范围。 - **电力规模跃迁**：Meta 网络基础设施电力足迹在几年内从 1.3 GW 扩展到 7 GW，增长近 5 倍。 - **容量管理定义**：从初始需求信号输入，到最终交付在线、健康且具备实际业务承载能力的网络容量全流程。 --- ### [01:47](https://www.youtube.com/watch?v=nZgru_53Ask&t=107s) 异构性与复合变数：被宏观数字掩盖的“分歧鸿沟” 表面上的吉瓦增长并不等同于标准模块的简单复制，因为这些新建的吉瓦容量彼此之间“根本无法相互替代”。在高速演进的 AI 世界中，每一个物理和逻辑维度都在同时发生剧烈变化。以作为核心算力载体的 GPU 为例，数据中心内部并非运行着单一制式的硬件，而是 $N$ 代、$N+1$ 代乃至 $N+2$ 代多代 GPU 同时并存。它们在计算速度、显存容量、网络互联接口标准以及网卡（NIC）处理能力上各不相同，且新旧硬件在生产环境中同时混部——工程师们“甚至还没来得及完成上一代架构的迁移，下一代架构的演进就已经全面铺开”。硬件本身的异构性又进一步与底层物理数据中心环境发生耦合与碰撞。不同代际的硬件对供电密度、空间布局以及散热制冷技术（如液冷与风冷）的要求千差万别。每一次芯片层面的变动，叠加机房物理约束的多样性，都会直接导致网络架构蓝图产生复合式的衍生变化。如果将这些要素绘制在同一张图表上，就会清晰地呈现出致命的“分歧鸿沟”（Divergence Gap）： ```mermaid flowchart LR accTitle: Divergence Gap in AI Network Buildout accDescr: 展现电力容量、系统复杂性与团队人力随时间推移的背离趋势图 subgraph Trends["趋势背离 (Divergence Gap)"] direction TB C["系统复杂度：指数级爆发 ↗↗"] P["电力容量需求：线性高速攀升 ↗"] H["工程团队人力：基本保持持平 ➔"] end Trends --> Sol["破局点：AI 原生自主代理 (AI-Native Agents)"] ``` 电力容量以线性高速攀升，系统交织的复杂度呈指数级爆发，而负责规划、设计与落地的工程团队总人数却基本保持平缓。Gawish 直言，人力不可能也不应该随着吉瓦数的增加而同步线性扩招。传统依赖编写更多脚本工具、增加软件抽象层或建立数据模型的工程手段，已经彻底触碰到了吞吐极限。弥合这一巨大鸿沟的唯一出路，就是引入 AI 原生的自主化能力，让 AI 直接参与到容量建设的核心链路中。 - **多代硬件并存**：$N$、$N+1$ 与 $N+2$ 代 GPU、XPU 与智能网卡（NIC）在生产环境中高度混合，迁移窗口永久重叠。 - **环境约束裂变**：机房空间、电力密度与冷却方案的组合差异，导致交付蓝图产生组合爆炸。 - **分歧鸿沟（Divergence Gap）**：工程任务量的指数级复杂度增长与基本持平的工程团队规模之间的结构性失衡。 --- ### [04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) 容量交付的四阶段乘法模型：从 Plan 到 Run 的未知数裂变 Meta 将整个 AI 网络基础设施的构建流程严格划分为四个高度依赖的序贯阶段：规划（Plan）、设计（Design）、交付（Deliver）与运营（Run）。每一个阶段的输出，都是下一阶段不可或缺的输入依据： 1. **规划（Plan）**：将极其粗糙的前端需求信号（如自研 XPU/第三方 GPU 路线图、业务产品预期、变电站与区域可用电力指标）转化为具有法律与工程约束力的投资决策。团队必须提前数年下注，精确押注在哪个季度的哪个具体园区与建筑空间内，部署多少兆瓦（MW）乃至吉瓦（GW）的特定容量。 2. **设计（Design）**：将规划阶段的宏观信号转化为兼顾物理与逻辑层面的工程蓝图。这涵盖了具体的网络拓扑架构、机架空间排布、跨机架布线路由、端口映射矩阵，以及如何与空间、功耗与散热等物理制约精确匹配。在 AI 时代，团队设立了一条基本工程防御假设：“任何容量组合都必须默认是异构的，因而可能需要完全定制的蓝图设计”。 3. **交付（Deliver）**：负责物料采购、设备物理安装、固件引导、配置下发与联调验收（commissioning），将设计蓝图从图纸变为真正承载在线业务流量、通过严苛健康检查并交付生产系统的网络容量。 4. **运营（Run）**：涵盖保持整个超大规模机群在跨越多代硬件、跨越多版固件及混合网络架构下长期高可用运转所需的全部监控、自愈与运维工作。 ```mermaid flowchart LR accTitle: Capacity Buildout Multiplicative Pipeline accDescr: 展示 Plan、Design、Deliver 与 Run 四个阶段及其乘法级复杂性传递机制 subgraph Pipeline["容量构建四阶段 (乘法裂变效应)"] direction LR P["规划<br>(Plan)"] -->|"✖ 需求不确定性"| D["设计<br>(Design)"] D -->|"✖ 异构蓝图分叉"| DL["交付<br>(Deliver)"] DL -->|"✖ 配置漂移与定制失效"| R["运营<br>(Run)"] end ``` 特别耐人寻味的是，在展示这套流水线时，Meta 刻意没有使用单向导通的箭头，而是在每个阶段之间醒目地打上了“乘号（$\times$）”。Gawish 强调，AI 基础设施的复杂性与变异性从来不是逐级累加的，而是“呈乘法效应级联扩散”。前置阶段遗留的每一个未知数或特例假设，都会在后续阶段迅速分叉为多个相互纠缠的次生未知数。例如，设计阶段为一个新代际集群量身定制的特殊布线与端口映射蓝图，传递到交付阶段就会直接引发配置漂移（configuration drift）、自动化部署流水线兼容性断裂以及海量专有工具链适配失败。 - **提前数年的规划押注**：在实际工作负载运行前数年，就必须锁定特定地理位置与时间节点的电力与容量落地。 - **设计防御性假设**：摒弃“一套标准打天下”的旧思维，默认每批交付的硬件基础设施都具有异构性。 - **乘法级复杂度裂变**：阶段间的变异度以乘法形式（$\times$）传导，局部特异性会迅速蔓延为全局工具链和配置策略的雪崩。 --- ### [07:35](https://www.youtube.com/watch?v=nZgru_53Ask&t=455s) AI 飞轮效应与业务一线专家打造的 Agent 矩阵面对跨阶段的异构挑战与复合级复杂度爆炸，Meta 的解法是在流水线的每一个环节全面注入 AI 原生能力，从而催生出一个正向循环的工程“飞轮效应”（Virtuous Cycle）：AI 业务的爆发直接催生了庞大的网络容量诉求；海量的容量诉求迫使底层网络必须以超前规模极速构建并稳固运行；而在这个极限承压的构建与运营流程中，AI Agent 反过来充当工程生产力的力量放大器（Force Multiplier）。这一正向闭环有效破解了前述的“分歧鸿沟”。 Gawish 特别澄清了一个行业内常见的认知误区：流水线中所应用的这些 AI Agent 绝非出自某个远离一线生产、闭门造车的专门 AI 算法团队。恰恰相反，在 Plan、Design、Deliver、Run 四个阶段大显身手的每一个 AI Agent，都是由常年扎根在对应业务一线、拥有多年实战排障经验的资深网络工程师主导构建的。这些 Agent 的内核沉淀了最懂该领域系统痛点的工程师认知，是人类专业经验的直接代码化与智能实体化。演讲最后给出了本次闪电演讲系列的完整议题导航，后续由四位核心一线领域专家分别展开全链路拆解： - 由 Ben Vallas 深入剖析规划阶段（Plan）的 AI 实践； - 由 Maz Mohammadi 带来面向异构蓝图的设计阶段（Design）深度分享； - 由演讲者本人 Mohab Gawish 详解交付阶段（Deliver）的自动化配置、联调与交付攻坚； - 由 Yuling Chen 呈现超大规模集群在运营阶段（Run）的 AI 运维与健康自愈保障。 ## Notable quotes > [00:05](https://www.youtube.com/watch?v=nZgru_53Ask&t=5s) 支撑 AI 的基础设施，正日益由 AI 自身来交付 — Mohab Gawish > [00:05](https://www.youtube.com/watch?v=nZgru_53Ask&t=5s) The infrastructure that powers AI is increasingly delivered by AI — Mohab Gawish > [01:47](https://www.youtube.com/watch?v=nZgru_53Ask&t=107s) 根本无法相互替代 — Mohab Gawish > [01:47](https://www.youtube.com/watch?v=nZgru_53Ask&t=107s) 甚至还没来得及完成上一代架构的迁移，下一代架构的演进就已经全面铺开 — Mohab Gawish > [01:47](https://www.youtube.com/watch?v=nZgru_53Ask&t=107s) 分歧鸿沟 — Mohab Gawish > [04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) 任何容量组合都必须默认是异构的，因而可能需要完全定制的蓝图设计 — Mohab Gawish > [04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) 呈乘法效应级联扩散 — Mohab Gawish > [04:12](https://www.youtube.com/watch?v=nZgru_53Ask&t=252s) 一套标准打天下 — Mohab Gawish > [07:35](https://www.youtube.com/watch?v=nZgru_53Ask&t=455s) 飞轮效应 — Mohab Gawish > [07:35](https://www.youtube.com/watch?v=nZgru_53Ask&t=455s) 力量放大器 — Mohab Gawish ## Mentioned Mohab Gawish, Meta, @Scale, Ben Vallas, Maz Mohammadi, Yuling Chen, GPU, XPU, 智能网卡（NIC）, 网络基础设施（Network Infrastructure, NI）, AI 原生自主代理（AI-Native Agents）, Plan-Design-Deliver-Run 容量交付流水线 ## Video description Learn more about the @Scale conferences here: https://atscaleconference.com/ [Source](https://www.youtube.com/watch?v=nZgru_53Ask)

Source quotes

[0:05] Mohab Gawish: 支撑 AI 的基础设施，正日益由 AI 自身来交付 [0:05] Mohab Gawish: The infrastructure that powers AI is increasingly delivered by AI [1:47] Mohab Gawish: 根本无法相互替代 [1:47] Mohab Gawish: 甚至还没来得及完成上一代架构的迁移，下一代架构的演进就已经全面铺开 [1:47] Mohab Gawish: 分歧鸿沟 [4:12] Mohab Gawish: 任何容量组合都必须默认是异构的，因而可能需要完全定制的蓝图设计 [4:12] Mohab Gawish: 呈乘法效应级联扩散 [4:12] Mohab Gawish: 一套标准打天下 [7:35] Mohab Gawish: 飞轮效应 [7:35] Mohab Gawish: 力量放大器

Source quotes

[0:05] Mohab Gawish: The infrastructure that powers AI is increasingly delivered by AI. [3:03] Mohab Gawish: divergence gap

Original transcript

[0:05] Hello everyone. This is a lightning talk [0:08] series where we'll be walking you [0:10] through how we have been leveraging AI [0:14] in network infrastructure buildout here [0:16] at Meta. In the next few minutes, I'll [0:20] be setting the frame for this series. [0:23] Why network capacity buildout is the [0:25] problem that we have picked to go over [0:28] in this lightning series and why we are [0:31] restructuring it around AI agents. Now [0:35] the theme of this series can be [0:36] summarized in one sentence. The [0:39] infrastructure that powers AI is [0:42] increasingly delivered by AI. And now [0:45] while this makes for a nice slogan, it [0:48] isn't just that. It isn't just a slogan. [0:51] It's actually a necessity. And to be [0:54] clear in this context, what we mean by [0:56] capacity management is the endto-end [0:59] pipeline that takes a demand signal and [1:02] turns it into a live, healthy, and [1:05] operational network capacity. [1:09] In the past few years, we have increased [1:11] our network infra power footprint from [1:14] 1.3 GW to 7 GW here at Meta. That's [1:18] roughly 5x growth, more or less. Now [1:21] when you hear a number like that as a [1:24] network engineer you may assume it's [1:26] just linear growth more of the same [1:29] scale out as we call it 5x more racks 5x [1:32] network fabrics 5x the number of ports [1:34] and all of these other parameters but in [1:37] the context of AI this number actually [1:39] hides much more complexity [1:42] than mere linear growth or scale out. [1:47] What that headline number actually hides [1:50] is that those gigawatts that we've [1:52] built, they aren't interchangeable. They [1:55] aren't more of the same. In the fast [1:59] evolving AI world, every one of these [2:02] axis is varying at the same time. When [2:05] we pick GPUs, for example, which are [2:08] kind of the primary host when we're [2:10] talking, you know, about AI network [2:14] buildout, we're actually talking about [2:16] multiple generations of GPUs, N, N [2:18] plus1, N plus2, varying in their [2:21] capabilities, and their speed, in their [2:24] memory, in their interface to the [2:25] network, in the capabilities of the [2:27] neck, and many other parameters. And [2:29] they all coexist in production [2:31] simultaneously. we don't get to finish a [2:34] migration before the next one starts. [2:36] There's also infra constraints around [2:39] those GPUs, the environment that [2:41] incubates the network itself and its [2:44] hosts. There are varying parameters [2:47] there as well. We're talking about [2:48] cooling, we're talking about power [2:50] density, we're talking about space, [2:53] and each variation of those combined, [2:56] they create compounding variations of [2:58] the buildout blueprint. [3:03] Now putting these two things on the same [3:06] chart, we find that power footprint [3:09] grows linearly. Complexity, however, [3:12] grows exponentially [3:14] and the team that builds all this out [3:16] remains roughly flat. Headcount doesn't [3:20] necessarily scale in lock step with [3:22] gigawatt and it shouldn't be. And this [3:24] is what we call the divergence gap. [3:27] That's the gap that emerged from putting [3:30] these charts on top of each other. And [3:33] this divergence gap between the quantity [3:36] of work that we need to do and the [3:38] exponential complexity that it entails [3:42] and the fact that the team that handles [3:44] it all out remains flat. This is the key [3:49] motivation for exploring what we can do [3:51] with autonomous AI native capabilities [3:55] in this capacity buildout. We're talking [3:58] here about leveraging AI directly in [4:00] building this infrastructure and not [4:02] merely relying on more tools, [4:04] automation, abstraction, modeling and [4:07] the standard softwarebased work. [4:12] When we talk about AI network [4:14] infrastructure buildout, we're actually [4:18] talking about a full pipeline. [4:21] More specifically, we're talking about [4:23] four sequential phases. Each phase [4:25] provides the necessary input to the [4:27] next. The first phase is plan. This is [4:31] where we turn raw demand signal like XPU [4:34] road maps, product forecasts, power [4:37] availability into a committed bet on how [4:40] many megawatts or gigawatts we want to [4:43] deploy. It also entails deciding what [4:46] kind of capacity we want to land in [4:49] which area, in what space, and at which [4:51] time, which quarter. It's basically a [4:54] bet that we place years ahead of the [4:58] workload that we'll actually run on. [5:01] The subsequent phase is design and this [5:03] is the phase where we translate that [5:07] plan signal into a physical and a [5:10] logical blueprint. This is where we make [5:13] decisions around the network topology [5:16] form the rack layout the cabling the [5:18] port maps how this all fits in the infra [5:22] constraints that we that we just covered [5:24] like space power density and cooling. [5:28] It basically covers pretty much all the [5:31] high level and low-level details that an [5:34] installer can use to build the network. [5:38] And in the fast evolving evolving world [5:40] of AI, we make a defensive assumption [5:43] that every capacity mix is heterogeneous [5:47] and may require a different blueprint. [5:50] I'll come back to this point later [5:52] because it's one of the most fundamental [5:56] pieces that motivated this whole effort [5:59] around leveraging AI in capacity [6:02] buildout. The next phase is deliver. [6:05] This is where we procure, install [6:08] everything that is related to the [6:10] network and it also entails the [6:12] provisioning work, the configuration [6:13] work, the commissioning, basically any [6:16] activity that contributes to turning a [6:19] blueprint that comes from design into [6:22] live traffic carrying capacity that's [6:25] been validated and handed to production. [6:29] The next phase is run and that one [6:31] entails all activities necessary to keep [6:33] that fleet in operation [6:36] and that happens across multiple [6:38] generations of software, hardware and [6:41] architecture. While this is meant to be [6:43] a pipeline, we deliberately didn't draw [6:46] directional arrows in between them. As [6:48] you might have noticed, we chose instead [6:50] to use a multiply sign between these [6:53] phases. And this is to signify a key [6:56] point with this pipeline. Variability [6:59] which has been at the heart of the AI [7:01] infra buildout doesn't add up across [7:04] these phases. It actually multiplies. [7:07] Each phase inherits a novel set of [7:09] unknowns and constraints from the one [7:12] before it. And each one of those [7:14] unknowns can fork into multiple unknowns [7:18] in the subsequent phase. For example, [7:21] every new design blueprint creates a new [7:23] vector of configurations [7:26] to drift and break customizations in the [7:29] tooling pipeline and many other [7:31] variations. [7:35] Because of all of this and the challenge [7:36] of heterogeneity [7:38] and compounding complexity across [7:41] phases, we wanted to explore how we can [7:44] use AI in every phase of those. And this [7:47] creates an interesting virtual cycle. AI [7:49] drives capacity demand and that demand [7:52] means networks have to be built and run [7:54] at scale and AI agents can help build [7:58] and run them. So in this context AI [8:01] becomes a force multiplier that makes [8:03] engineering expertise scale and I think [8:06] that to a great extent addresses that [8:09] the key challenge that we've presented [8:12] compounding complexity met by AI [8:15] capability that is a force multiplier [8:21] none of these agents were built in a [8:23] vacuum or by some specialized AI team [8:27] that hasn't been actually executed. [8:28] computing this work. AI agents that we [8:31] have been leveraging at every phase of [8:34] this pipeline were built by an AI team [8:37] off to the side. Each of them were [8:39] actually built by experts in their [8:43] respective domain. People who have been [8:45] running these activities for years and [8:48] in that sense those AI agents actually [8:51] embody and represent the expertise of [8:54] their authors. So we have a few [8:58] lightning talks covering a deep dive [9:01] into how we have been leveraging [9:03] AI agents in every phase of that [9:06] pipeline. We have Ben Valis covering [9:09] clan. We have Maz Mohein covering [9:12] design. We have myself Mohab Gawish [9:15] covering deliver. And we have Yuling [9:18] Chen covering run. Enjoy. Thank you.