Meta Rewrites Fiber Playbook for Ohio Supercluster

Meta为俄亥俄超级集群重塑光纤建设方案 建设Meta的Prometheus AI超级集群需从轻量电信开槽转向重型土木工程，以部署10万芯光纤。 为了支持Meta的首个AI训练超级集群Project Prometheus，网络团队必须将俄亥俄州新奥尔巴尼的光纤容量从864芯提升100倍至10万芯。Meta与AFL联合开发了1728芯高密度光缆，并从标准电信开槽转向重型土木工程，建造了单间公寓大小的地下井室，并在161号高速公路下挖掘了5英尺宽的微型隧道。通过在北卡罗来纳州的现场测试以及井下气吹敷设过程中的持续调整，团队成功应对了紧迫的工期、当地交通拥堵以及市政严苛的“一次开挖”施工限制。 Meta 为其首个 AI 训练超级集群 Prometheus 大幅升级地面网络，将数据中心间的光纤需求从以往的 10 至 864 芯激增 100 倍至 100,000 芯，重点落地于俄亥俄州新奥尔巴尼。 (0:03) 通过“光纤改善（Fiber Kaizen）”测试，Meta 与厂商 AFL 联合研发出 1,728 芯高密度光缆，在直径仅增加 1.4 毫米的同时实现容量翻倍，并在 5 公里半径内铺设了 210 万光纤公里的线缆。 (3:51) 项目从传统轻型通信施工转向重型市政土木工程，挖掘了深达 20 英尺、宽达 9 英尺的管沟，并在 161 号高速公路下方开凿了 5 英尺宽的微型隧道，配合单间公寓大小的地下机房完成管线汇聚。 (6:46) 面对周边 Google、微软、亚马逊数据中心以及英特尔芯片工厂建设带来的严重交通拥堵，Meta 依托地下管网全要素测绘贯彻“只挖一次”策略，并在地下机房内级联多台气吹机高效推进光缆铺设。 (8:59)

To support its Prometheus AI training supercluster in New Albany, Ohio, Meta had to scale its fiber connectivity from historical baselines of 10 to 864 strands up to 100,000 strands. According to Meta's Fabrice Ouandji and Sebastian Gault, this hundredfold increase forced the company to transition from standard telecom installations to heavy civil engineering and customized high-density cabling. Exponential Fiber Scaling and High-Density Cable Innovation (0:03) In 2018, Meta's initial data center connectivity required approximately 10 fiber strands across two diverse paths, which evolved into Middle Mile Infrastructure deployments using single 864-fiber cables inside 1.25-inch to 1.5-inch conduits. By 2025, the Prometheus supercluster required 100,000 fiber strands within a 5-kilometer radius, totaling 2.1 million fiber-kilometers—a distance equivalent to three round trips to the Moon. To identify cabling suited for long-distance pneumatic jetting, Meta conducted a 'Fiber Kaizen' testing run in Asheville, North Carolina. While micro-cables were rejected as too fragile and prone to kinking, Meta initially deployed 140 standard 15-millimeter, 864-fiber cables. To conserve conduit capacity for future expansions, Meta collaborated with AFL to develop a 1,728-fiber cable that doubled strand capacity while increasing diameter by only 1.4 millimeters to 16.4 millimeters. Heavy Civil Engineering and Underground Infrastructure (6:46) Standard telecom projects typically rely on trenches 6 inches wide and 3 to 4 feet deep. To house the 140 cables required for Phase 1, Meta had to construct 'mega duct banks' up to 9 feet wide and 20 feet deep—large enough to stack three SUVs—shifting the project into heavy civil engineering comparable to municipal drainage or water mains. Because existing margins along public rights-of-way were congested with utilities, crews had to excavate directly through active roadways. Major obstacles required specialized subterranean structures. To cross Highway 161 without halting surface traffic, Meta bored a 5-foot-wide micro-tunnel large enough for a machine operator to excavate soil from within. At critical junctions, the company bypassed standard handholes in favor of precast concrete underground vaults the size of studio apartments to route and house the massive conduit bundles. Regional Congestion and the 'Dig Once' Strategy (8:59) Project requirements expanded rapidly from connecting a single data center to interconnecting multiple facilities across a 5-kilometer radius, even as the exact locations and fiber counts for later phases remained uncertain. To prevent continuous multi-year road disruptions, Meta adopted a strict 'dig once' approach, conducting comprehensive underground utility mapping across every road in the area to secure conduit capacity ahead of construction. Execution was complicated by intense regional traffic. The 5-kilometer corridor housed active facilities and expansions for Meta, Google, Microsoft, Amazon, third-party operators, and a multi-billion-dollar Intel semiconductor foundry. This created an influx of 5,000 to 10,000 construction and trade vehicles daily, requiring continuous permitting coordination with municipal authorities and careful scheduling among competing developers. Sub-Surface Jetting Adaptations and Cable Weight Limits (13:32) To prevent surface lane closures, Meta installed jetting machines and personnel directly inside the underground vaults. Initial plans to jet 864-fiber cables across spans of 3,700 to 3,800 feet failed because the cable's physical weight exceeded the lifting capability of compressed air. Engineers adjusted the operational distance to a range of 2,500 to 2,600 feet, linking five to six jetting machines in series ('cascading') to install five to six continuous cables per day. Meta enforced a zero-slack policy, eliminating inline splice enclosures and coils inside vaults to maximize density, which means any future damage will require a full cable replacement. When Phase 2 introduced the heavier 1,728-fiber cable, the doubled physical weight further limited jetting runs to approximately 1,800 feet, necessitating additional inline jetting stations. Ouandji and Gault conclude that delivering physical infrastructure for AI superclusters requires agility when field conditions invalidate plans, cross-industry human coordination among competitors and civic authorities, and partnering with heavy civil contractors rather than relying solely on conventional telecom practices.

Meta网络投资团队的Fabrice Ouandji与Sebastian Gault指出，为支撑首个AI训练超集群Prometheus的算力互联，Meta在俄亥俄州新奥尔巴尼将数据中心间的光纤需求从以往的数百芯激增至10万芯；这种百倍级的容量跃升迫使网络建设跳出传统电信工程模式，全面转向高密度定制光缆、市政级重型土木开挖以及地下级联气吹等系统性创新。 需求激增与高密度光缆技术突破 (0:03) 根据Meta团队的回顾，2018年其数据中心初期的连通需求仅为两条物理路由上的约10芯光纤，主要依赖Zayo和Lumen等电信运营商交付。随着中程基础设施（MMI）项目推进，Meta开始自建包含单根864芯光缆的管道系统。然而，2025年启动的Prometheus超集群使光纤需求激增至10万芯，形成传统规模100倍的容量跃升。在新奥尔巴尼该集群方圆5公里范围内，Meta铺设的光缆总量达到210万光纤公里，体量相当于往返月球3次。 为了在有限的管道空间内装入激增的光纤，Meta在北卡罗来纳州阿什维尔的3700英尺测试管道中组织了厂商参与的“光纤改善”（Fiber Kaizen）现场竞标。微型光缆因结构脆弱、易扭结且不利于长距离气吹而被放弃；团队最终选定刚性更优的15毫米直径标准864芯光缆用于第一阶段的140根敷设任务。随后，Meta与AFL联合研发了1728芯高密度光缆，在直径仅增加1.4毫米（达到16.4毫米）的情况下实现了容量翻倍，为北美骨干网后续扩张保留了宝贵的管道冗余。 市政级重型土木开挖与地下空间重构 (6:46) 将第一阶段的140根光缆埋入地下，直接推动Meta的施工范畴从轻量电信工程转变为市政重型土木基础设施。传统的电信管沟通常仅为6英寸宽、3至4英尺深，但为了容纳Prometheus的巨型管道组，Meta开掘了深达20英尺、宽达9英尺的深槽，其空间足以纵向叠放三辆标准SUV且车顶勉强触碰地面。由于公共通行带绿化区已被既有市政管网占满，Meta不得不申请特别许可切开通行道路，并与专门从事重型土方开挖的工程承包商合作。 在道路跨越与地下接续节点上，工程规格同样超越常规。为穿越161号公路且不扰民阻断交通，工程团队在公路下方顶进开凿了一条宽5英尺的微型隧道，其宽度足以让机械操作员直接入内挖掘排土。而在管线交汇处，Meta摒弃了传统电信施工的街头小型手孔井，预制安装了空间尺度堪比单身公寓的地下混凝土储藏井（vaults），用以容纳多向交汇的庞大管道组并作为地下作业的核心工作井。 “单次开挖”策略与拥堵走廊的协调博弈 (8:59) 工程实施中面临着动态变化的需求与高度拥挤的外部施工走廊双重制约。第一阶段最初仅计划连通单个数据中心，但数月内范围便扩展至5公里半径内的多个园区。由于前期挖掘时二期数据中心的精确落点和芯数尚未最终敲定，而市政管理部门绝不可能容忍长达三年的连续破路，Meta确立了“单次开挖”（dig once）原则，通过对区域内全部道路进行地下管线全量测绘，精准规划管道组走廊以最大化单次铺设容量。 与此同时，该5公里区域也是全美工业建设极为密集的区域之一。Meta不仅要面对谷歌、微软、亚马逊数据中心及第三方多租户机房的扩张竞争，其东北角还有英特尔在建的数十亿美元芯片制造基地。每天有5000至10000辆施工技术人员车辆涌入狭窄路网，与光纤掘路封道严重冲突。工程的推进不仅考验地下施工，更依赖于Meta内部建设团队与市政监管部门、行业竞争对手之间的持续协调。 地下级联气吹技术与线缆物理极限突破 (13:32) 在光纤入管阶段，核心挑战落在利用压缩空气推进光缆的气吹（jetting）工艺上。为避免占用地面车道导致交通瘫痪，Meta决定不在路面设置作业区，而是将操作人员和气吹设备全数转入地下公寓大小的混凝土井内作业。最初设计的3700至3800英尺气吹间距在现场测试中遭遇失败，原因是864芯光缆的物理自重过大，导致压缩空气无法将其顺利托起并远距离输送。团队随后将气吹间距调整至2500至2600英尺的平衡区间，并采用5至6台气吹机串联的地下“级联”（cascading）方案。 地下级联方案实现了整缆直通穿透，井内不预留任何松弛盘留光纤，亦不安装在线接头盒；一旦光缆损坏必须全段抽出替换。该方法使团队达到了每天完成5至6根整缆敷设的施工进度。而在第二阶段引入1728芯超重线缆后，翻倍的线缆重量使气吹最大极限距离进一步萎缩至约1800英尺，迫使工程团队在地下部署更多级联站点以应对缆重挑战。 Meta团队强调，AI超大规模网络基础设施的核心在于现场应对突发限制的敏捷应变，以及打破传统电信界限、与市政及重工业伙伴建立深度协同。尽管Prometheus证明了高密光缆与重型土木结合应对AI海量互联的可行性，但随着光缆重量对气吹物理极限的压缩以及井内零盘留策略的采用，超大容量网络后续的维护更换成本与长期地下空间容量，仍是该模式向更大规模扩展时需要权衡的工程难题。

## Background 本演讲由 YouTube 频道 @Scale 于 2026 年 9 月 3 日发布，记录了 Meta 的 Fabrice Ouandji 与 Sebastian Gault 在圣克拉拉会议中心现场发表的演讲。视频简介指出，Meta 的 Prometheus 超级集群对互联网络提出了前所未有的超高光纤芯数、动态选址以及极度压缩的交付工期要求，演讲详细揭示了团队为满足跨数据中心互联需求所实现的诸多网络创新。 ## TL;DR 为了支撑 Meta 首个 AI 训练超级集群 Prometheus，跨数据中心的光纤芯数需求从过去的 864 芯直接暴增百倍至 100,000 芯，促使团队重新发明了物理层连接方案。Meta 联合 AFL 研发出外径仅 16.4 毫米的 1728 芯超高密度光缆，并在俄亥俄州新奥尔巴尼实施了等同于市政重型防洪排污标准的深坑地下排管群工程。面对无法反复开挖的公共路权限制、密集的科技巨头施工冲突以及光缆超重带来的气吹停滞危机，团队通过将气吹间距极限压缩至 1800 英尺、转入地下人手孔作业及零余长策略，最终在死线前如期打通了互联通道，保障了海量 GPU 算力的顺利点亮。 ## Atmosphere & style 整场演讲兼具硬核工程的震撼感与技术老兵的幽默自嘲。两位讲者配合默契，Fabrice 擅长以“深达 20 英尺能叠放三辆 SUV”、“人手孔大如单身公寓”等极具视觉张力的比喻调动全场气氛，并诚恳向俄亥俄当地居民道歉；Sebastian 则展现出严谨求实的工程师底色，直言白板推演在现实泥潭中的屡屡受挫。从最初雄心勃勃到遭遇实地重拳，再到敏捷调优最终成功交付，整个叙事宛如一部跌宕起伏的地下土木史诗，让观众真切感受到冰冷庞大的 AI 算力背后，是人类工程智慧与跨界协作的极限拉扯。 ## Bulletin - [00:03](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=3s) 为支撑 Meta 首个 AI 训练超级集群 Prometheus，跨数据中心光纤需求直接从过去的 864 芯跃升至 100,000 芯，催生了百倍规模的超高密度地下排管群。 - [03:51](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=231s) Meta 与 AFL 联合研发了外径仅 16.4 毫米的 1728 芯超高密度光缆，仅比 864 芯线缆增粗 1.4 毫米即实现容量翻倍，在核心 5 公里半径内敷设光纤总长达 210 万公里。 - [06:46](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=406s) 项目开挖深度达 20 英尺、沟宽 9 英尺，体量等同于市政重型给排水工程，迫使 Meta 跨界引入重型深坑挖掘工程承包商进行作业。 - [13:32](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=812s) 由于光缆重量超预期，团队推翻 3800 英尺的白板推演，将气吹机布设在单身公寓大小的地下人手孔内，摸索出 864 芯缆 2500 英尺与 1728 芯缆 1800 英尺的极限推进间距。 ## Key takeaways - [00:03](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=3s) 从 2018 年依赖运营商采购双路由 10 芯光缆，到 MMI 自建 864 芯光缆，再到 2025 年 Prometheus 的 10 万芯需求，Meta 光纤网络经历了百倍的容量跃迁。 - [03:51](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=231s) 在北卡阿什维尔的真实管道竞标测试中，微缆方案因脆弱易扭结被淘汰，结构更坚固的硬质结构设计成为超长距离气吹敷设的首选。 - [03:51](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=231s) 与 AFL 定制研发的 1728 芯光缆以微小的外径增幅（增粗 1.4 毫米）换取了成倍的纤芯密度，极大节省了地下极其有限的管道空间资源。 - [06:46](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=406s) 通信光缆管沟尺寸从传统的 6 英寸宽、3 至 4 英尺深，剧增至 9 英尺宽、20 英尺深，空间足以垂直叠放三辆标准全尺寸 SUV。 - [08:59](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=539s) 面对未来二期数据中心选址和规模的不确定性，Meta 严格落实公共道路通行权下“只挖一次”的原则，提前全量勘测地下现有市政管线并预留管群。 - [08:59](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=539s) 新奥尔巴尼 5 公里半径内汇聚了 Meta、Google、微软、亚马逊及英特尔晶圆厂，每天多达 5000 至 10000 名工程人员通勤，导致道路占道开挖的内外部协调难度倍增。 - [12:02](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=722s) 在穿越俄亥俄州 161 号公路时，团队开挖了一条 5 英尺宽的地下隧道，允许工程机械直接驶入内部掘进掏土，并建造了单身公寓大小的地下人手孔。 - [13:32](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=812s) 一期 864 芯气吹因线缆重量无法实现原定 3800 英尺间距，团队将设备移入地下手孔级联作业以减小路面交通干扰，并将间距调整为 2500 至 2600 英尺黄金区间。 - [15:44](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=944s) 地下人手孔采取极端的“零余长”直抽无接头策略，不设盘留与熔接包，虽换取了空间并确保每天气吹 5 至 6 根缆的工期进度，但一旦线缆受损必须全线抽换。 - [15:44](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=944s) 二期 1728 芯光缆因重量接近翻倍导致气浮难度剧增，气吹机间距被迫进一步压缩至 1800 英尺并增加接力设备才攻克敷设难题。 - [17:40](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=1060s) AI 超级集群建设的核心启示在于工程敏捷应变、人际与多方利益相关者的高效协同，以及从传统电信运营商合作向重型市政土木工程的跨界重构。 ## Detailed walkthrough ### [00:03](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=3s) 面临前所未有的挑战：从传统骨干到 Prometheus 的百倍跃迁 在演讲开场，Meta 网络投资团队的合作伙伴经理 Fabrice Ouandji 向观众提出了一个极具挑战性的问题：“如果公司要求你去完成一项前所未有的任务，去负责一个极其庞大复杂的基础设施项目，以支撑全公司的 AI 加速推进，你会怎么做？”Fabrice 的团队主要负责 Meta 全球数据中心与网络接入点（PoP）之间的互联互通，涵盖陆地光缆与海底光缆。而这次，他们面临的全新任务是为 Meta 的首个 AI 训练超级集群——Prometheus 提供网络互联支持。 为了说明 Prometheus 所带来的颠覆性变化，Fabrice 带领大家回顾了 Meta 光纤需求的发展历程。早在 2018 年左右，一个数据中心或 PoP 接入网络时，首日连接需求（Day 1 connectivity）通常只需要极少的纤芯——在两条互为备份的物理路由上各部署约 10 芯光纤即可。当时 Meta 的解决方式主要依靠外部电信运营商伙伴（如 Zayo 和 Lumen 等），通过采购其建设和交付的线路来实现互联。 随着业务扩张，数据中心与节点数量激增，Meta 随后启动了雄心勃勃的中继基础设施（MMI, Middle Mile Infrastructure）建设规划。在这一时期，Meta 与建设公司及电信运营商展开大规模合作，在公共道路两旁开挖铺设几根 1.25 英寸或 1.5 英寸规格的导管，并在其中敷设一条包含 864 芯的光缆。这套方案在当时完全能够满足首日乃至未来的扩容需求，供应商甚至感叹 Meta 几乎是在独自开创全球规模最大的私有光纤网络。然而到了 2025 年，Prometheus 的出现彻底颠覆了既有的建设模型：跨数据中心的光纤需求直接从 864 芯飙升至整整 100,000 芯。面对这种百倍（100x）的容量增长，Meta 必须在俄亥俄州新奥尔巴尼（New Albany）的公共道路通行权（public right-of-way）内，打造出前所未见的超大规模地下排管群（mega duct banks）。 ```mermaid flowchart TD accTitle: Meta 数据中心光纤需求演进历程 accDescr: 展示从 2018 年传统模式到 MMI 建设，再到 2025 年 Prometheus 带来的百倍光纤需求跃迁。 subgraph Phase1["2018 年前后：传统运营商合作"] A["需求：双路由仅需约 10 芯"] --> B["方式：向 Zayo、Lumen 等运营商采购"] end subgraph Phase2["MMI 时期：自建中继骨干"] C["需求：快速增长，覆盖更多数据中心"] --> D["方式：开挖道路，敷设 1 条 864 芯光缆"] end subgraph Phase3["2025 年：Prometheus 超级集群"] E["需求：100,000 芯光纤（100x 暴增）"] --> F["方式：重型土木工程，打造超级地下排管群"] end Phase1 --> Phase2 Phase2 --> Phase3 ``` - **2018 年传统阶段**：单个数据中心首日需求仅约 10 芯光纤，采用两条物理独立路径，依赖传统运营商交付。 - **MMI 阶段**：敷设 1.25 至 1.5 英寸导管，单条光缆容纳 864 芯光纤。 - **Prometheus 阶段**：芯数跃迁至 10 万芯，实现 100 倍容量飞跃，拉开超高密度光纤网络创新的序幕。 --- ### [03:51](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=231s) 光缆创新与改善（Kaizen）：高密度 1728 芯光缆的诞生 接替发言的 Sebastian Gault 聚焦于光纤层面的材料与结构创新。为了在严苛的时间与物理空间限制下找到最合适的光缆，团队引入了源自日本的“改善”（Kaizen，意为持续改进）机制，邀请了三家主流线缆制造厂商展开竞标测试。他们将试验场地定在北卡罗来纳州阿什维尔（Asheville）的导管供应商测试跑道上，并在长达 3,700 英尺（约 1,128 米）的真实测试管道中，对各家光缆进行气吹敷设（jetting）评估。 虽然三家厂商的光缆都能吹进相同的距离，但机械特性的差异显露无遗。其中两家厂商带来了微型光缆（micro cable），这种光缆直径非常小，但线体极度脆弱且容易发生打折扭结（kink）；而另一家厂商 AFL 则提供了标准硬度的 864 芯光缆，外径为 15 毫米。出于对长距离施工稳定性和可靠性的考量，团队最终选择了结构更为坚固的 AFL 标准方案。按照第一阶段的部署规模，他们需要在地下敷设多达 140 根 864 芯光缆。 确定方案后，Meta 并未止步于此，而是进一步与 AFL 展开联合定制研发，推出了全新的 1728 芯超高密度光缆。令人惊叹的是，这款 1728 芯光缆的外径仅为 16.4 毫米，相比原先 15 毫米的 864 芯光缆，外径仅增加了微不可察的 1.4 毫米，却实现了纤芯容量的整整翻倍。这一跨越式突破不仅节省了极其宝贵的地下管孔资源，为未来预留了巨大的弹性扩展空间，还能直接推广至 Meta 在全美范围内的其他网络工程。Sebastian 给出了一个震撼全场的数据：在 Prometheus 仅 5 公里的核心半径范围内，他们总共将铺设 210 万光纤公里（fiber kilometers）的光缆，这一长度足足可以在地球与月球之间往返三次。 - **测试场地**：北卡罗来纳州阿什维尔，3,700 英尺真实导管测试轨道。 - **方案权衡**：放弃易扭结脆弱的微缆方案，选用机械强度更高的硬质结构设计。 - **尺寸突破**：AFL 研发的 1728 芯光缆外径仅 16.4 毫米，仅比 864 芯光缆（15 毫米）增粗 1.4 毫米即可达成容量翻倍。 - **部署体量**：5 公里半径内敷设光纤总长达 210 万公里，相当于“地球到月球往返三次（three times to the moon and back）”。 --- ### [06:46](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=406s) 重型土木工程重构：公用路权下的超级地下管群 随后，Fabrice 重新接管讲台，分享了将这海量光纤埋入地下所经历的工程奇迹。在第一阶段，仅接入首个数据中心就需要一次性敷设 140 条大芯数光缆，而这一切施工都必须在俄亥俄州新奥尔巴尼的公用道路通行权范围内完成。Fabrice 半开玩笑地对当地居民致歉：“如果现场有来自俄亥俄州的朋友，对于我们在当地制造的所有麻烦和拥堵，我深表歉意。” 为了容纳百倍于以往的光缆管群，Meta 彻底抛弃了传统的通信光缆施工标准。过去的常规通信管沟通常只有 6 英寸（约 15 厘米）宽、3 到 4 英尺（约 0.9 至 1.2 米）深，刚好能埋下几根细导管。而 Prometheus 的超大排管群沟槽开挖直接占据了一整条机动车道，深度下切至地表以下 20 英尺（约 6.1 米），沟宽最宽处达到 9 英尺（约 2.7 米）。Fabrice 形容道，如果你站在施工沟槽边缘往下看，这个巨大的地下空间足以把三辆标准尺寸的 SUV 垂直叠放进去，最顶上一辆的车顶刚好与路面持平。 这种量级的开挖已经完全脱离了传统电信工程的范畴，直接跨入了大口径下水道、给水主管道和大型雨水防洪系统的市政重型土木工程水准。通信工程团队不得不跨界寻找全新的合作伙伴。最终，他们与一家名称中自带“挖掘（Excavating）”字样、专门从事大型深坑地基开挖的重型工程公司合作，凭借对方在极端地下工况方面的深厚积累，才顺利推进了这一庞大的地下管网工程。 | 对比维度 | 传统光纤网络开挖标准 | Prometheus 超级排管群标准 | |:--- |:--- |:--- | | **开挖宽度** | 约 6 英寸（约 15 厘米） | 宽达 9 英尺（约 2.7 米） | | **开挖深度** | 约 3 至 4 英尺（约 0.9–1.2 米） | 深达 20 英尺（约 6.1 米） | | **空间参照** | 容纳数根细导管 | 空间足以垂直叠放 3 辆标准全尺寸 SUV | | **工程性质** | 传统弱电与轻型管线敷设 | 等同于重型市政排污、供水与防洪工程 | | **首日线缆量** | 1 根 864 芯光缆 | 140 根大芯数光缆 | --- ### [08:59](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=539s) 动态需求与施工交织：在超大规模集群核心区“只挖一次” 项目正式动工后，团队遭遇了两大严峻考验。首要挑战来自于业务需求的极度动态变化与高度不确定性。在工程启动初期，规划仅仅要求连接一所数据中心；但短短几个月内，业务侧便明确 Prometheus 将会是覆盖方圆 5 公里半径内多个数据中心的超级互联集群。棘手的是，在推进一期工程建设时，二期数据中心的具体地理坐标尚未确定，所需的光纤芯数也是未知数，但可以确定的是，未来的管网必然要复用一期所经过的同一批市政主干道。 由于公共道路的开挖许可极其严格，市民和市政部门绝不可能容忍一条路连续三年反复开挖施工。团队必须贯彻“只挖一次（dig once）”的严苛原则。为了化解“需求未知、规模海量、仅能开挖一次”的矛盾，Meta 团队提前对 5 公里半径内的每一条现有道路展开了详尽的先期土木工程勘测。他们彻底摸排了路面下所有既有市政管线（电力、水气、通信等）的分布情况，精准测算在路边绿化带中挤出排管空间的可能性，并在绿化带空间不足时，提前研判向市政申请占用道路破土开挖的可行性方案。 第二个巨大挑战则是极其复杂的现场交通与多方主体冲突管理。新奥尔巴尼的这处 5 公里半径区域是全球算力巨头最密集的区域之一：Meta、Google、微软、亚马逊均在此处建有已投运或在建的大型设施，此外还有多家第三方数据中心服务商在加紧赶工；更夸张的是，在该区域的东北角，英特尔正在推进其投资数百亿美元的半导体代工厂（Foundry）。在这片产业高地中，每天有多达 5,000 到 10,000 名建筑施工与技工人员需要驱车往返通行。与此同时，Meta 却需要为了开挖 20 英尺深的超级管沟而封闭部分交通车道。这带来了密集的内外部冲突：对内，道路施工团队与数据中心园区现场建设团队之间存在作业面重叠；对外，必须不断调整工程设计方案以避让市政自身的基础设施更新项目，这迫使团队投入了大量精力与市政当局及各方建设主体进行高强度的协调联动。 - **需求动态性**：从单数据中心连接扩展至 5 公里半径内多中心互联，面对二期选址与规模未知的情况，强制实施“只挖一次”策略。 - **地下管线勘探**：对 5 公里半径内所有干道进行全量管线勘测与土木预工程化设计，最大限度优先使用路旁绿带。 - **高密度施工集群**：5 公里核心区内汇集 Meta、Google、微软、亚马逊数据中心、第三方托管商以及英特尔超级晶圆厂项目。 - **交通与管群冲突**：每天 5,000 至 10,000 名各类施工人员进出，道路封锁与深沟开挖面临极其复杂的内外部工程协调。 ### [12:02](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=722s) 穿越161号公路：单身公寓大小的地下人手孔与重型土建 在完成了前期的路线规划与跨团队沟通后，工程进入了极具挑战性的实体施工阶段。Fabrice 展示了施工现场的震撼照片。为了打通两座数据中心之间的连接通道，团队必须横穿位于俄亥俄州新奥尔巴尼（New Albany）的 161 号公路（Highway 161）。因为穿越主干高速公路是一次性工程，“这种机会你一生只有一次”，必须在施工阶段预留出应对未来极端增长的充裕空间。 为此，工程团队在 161 号公路下方开挖了一条宽达 5 英尺（约 1.5 米）的专用隧道。这个尺度庞大到操作员可以直接坐在挖掘工程机械上驶入地下通道，一边向前钻进挖掘，一边将掏出的泥土运送出来。在这条巨大的地下隧道贯通后，团队在管廊内铺设了尽可能多的管道（conduits），不仅满足了项目“第一天（Day 1）”的高密度连接需求，也为未来集群规模的进一步扩容锁定了扩展弹性。 随后展示的是被称为地下接线井/手孔（vaults / handholes / manholes）的地下空间。Fabrice 笑着指出，这些地下结构根本不是普通市政道路上随处可见的小井盖，它们的物理体积堪比一套“单身公寓（studio apartment）”。之所以建造得如此庞大，是因为唯有这种空间规模才能容纳汇入其中的巨量管网，并在管线敷设完毕后充当工程师维护和接入系统的核心作业空间。 ```mermaid flowchart TD accTitle: 土建与管道部署流程 accDescr: 展示从穿越161号公路开挖隧道到建造公寓级人手孔以及敷设光纤管道的土建流程。 subgraph Tunneling["穿越161号公路土建"] A["5英尺宽地下隧道挖掘<br/>(操作员开机进洞排土)"] --> B["预埋高密度管道群<br/>(满足Day 1并预留未来冗余)"] end subgraph UndergroundVaults["地下机房级人手孔"] B --> C["公寓级人手孔/地下井建造<br/>(单身公寓大小空间)"] C --> D["光纤气吹敷设接入点<br/>(无盘留直通级联作业)"] end ``` - 隧道规格：宽 5 英尺（约 1.5 米），位于新奥尔巴尼 161 号公路下方，允许工程机械开入内部挖掘排土。 - 管道网络：在隧道中一次性密排大量管道，锁定长远演进的管网弹性。 - 地下人手孔（Vaults）：体积相当于单身公寓大小，作为海量管道汇聚及后续网络接驳的关键节点。 --- ### [13:32](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=812s) 一期 864 芯气吹试水：白板方案遇挫与 2500 英尺黄金间距 完成重型土建并埋设管道后，光纤敷设的工作交由 Sebastian（Seb）负责。由于前期施工导致新奥尔巴尼市内多条道路封闭，当地居民已经颇有微词。对 Sebastian 而言，在项目的所有风险登记簿（risk register）中，光纤气吹（fiber jetting）与敷设是他内部多次强调的“最大未知数与最大风险项”。因为第一阶段（Phase 1）的目标，是要在既定时间内敷设多达 140 根 864 芯的光缆。 团队最初在办公室做了极为周密的白板推演，设定了气吹机（jetting machine）的放置点，胸有成竹地认为“方案很完美，一定能成”。然而实际一到现场进行实地测试，计划立刻碰壁。最初规划的气吹机间距约为 3,800 英尺（约 1,158 米），但因为光缆重量超出预期，气吹机产生的气流根本无法在如此长距离上重新将光缆托举浮起。 为了不因施工长期封闭车道而进一步惹恼居民，团队做出了关键调整：将原本架在路面上的气吹机全部搬入地下人手孔内作业，采用级联（cascading）直通方式推进，使路面受影响的交通监测范围从整条路缩减到仅 100 至 200 英尺。同时，团队将气吹机数量从原定的 3 至 4 台增加到 5 至 6 台，将布设间距逐步收紧，最终摸索出了约 2,500 至 2,600 英尺（约 762 至 792 米）的“黄金间距（sweet spot）”。 - 一期敷设规模：140 根 864 芯高密度光纤缆线。 - 交通最小化影响：气吹设备全部沉入地下人手孔作业，路面受干扰区缩减至 100 至 200 英尺。 - 机器调优：原计划 3,800 英尺间距失败，通过增加至 5-6 台机器，将间距缩短为 2,500-2,600 英尺。 --- ### [15:44](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=944s) 进度决战与二期 1728 芯挑战：1800 英尺极限间距与零余长策略 解决了气吹间距后，项目最大的瓶颈落在了交付速度上。因为如果进度拖延，两端已经就绪的数据中心就无法互联，“数以万计的 GPU 就只能闲置在那里，连指示灯都无法闪烁”。最终，团队通过精细化的施工节拍，达成了每天气吹推进 5 到 6 根光缆的稳定速度，精准赶上了交付死线。 在地下人手孔的布线管理上，团队采取了极端的“零余长（no slack）”直抽贯通策略。人手孔内完全不留任何盘留余线（slack coils），也没有空间放置熔接盒（splice enclosures）。这意味着整条链路是完全一体化贯通的，但代价也是残酷的——未来如果任意一根光缆发生受损破坏，团队无法就地剥线补熔，必须将整根光缆全线抽出并重新穿入一根新缆。 到了第二阶段（Phase 2），挑战进一步升级。团队需要铺设 76 根全新的 1,728 芯超高密度光缆。尽管该光缆的外径仅增加了 1.4 至 1.6 毫米，但内部容纳的光纤数量直接翻倍，导致整缆重量几乎翻了一倍。团队最初试图套用一期 2,500 英尺的间距经验，结果再次失败——沉重的光缆在管道中根本无法被气流悬浮。经过又一轮实地试错，团队最终被迫将气吹机间距进一步压缩到 1,800 英尺（约 549 米），并部署了更多台气吹机进行接力推送，才最终攻克了这一前所未有的敷设难题。 - 施工速率：稳定达到每天气吹 5 至 6 根光缆，确保数据中心 GPU 算力如期点亮。 - 维护模式：人手孔内零盘留、无熔接包；单缆受损必须全线抽换更新。 - 二期规格与调校：铺设 76 根 1,728 芯光缆，重量近乎翻倍，气吹机间距被迫缩短至约 1,800 英尺。 --- ### [17:40](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=1060s) 创新复盘：现场应变、人际联结与重型工程的跨界融合 Fabrice 在总结时提出了一个核心问题：“当你做一件前人从未做过的事情时，你能学到什么？”他将整个项目的经验凝练为三大核心心得： 第一，是“多度量，但关键时刻敢于落剪（measure twice, but do not be afraid to cut）”。虽然团队在前期做了极其扎实的规划、田野试验以及筛选缆线的改善（Kaizen）流程，但正如迈克·泰森的名言所说：“每个人都有一个计划，直到嘴上挨了一记重拳（Everyone has a plan until they get punched in the mouth）。”真正的工程创新要求团队必须敢于在实践中迈出切除和改动的那一步，根据工地的实际地质与物理反馈保持敏捷、迅速调整。 第二，“AI 的底色是人与人的连接（AI is really about human connections）”。虽然项目涉及的是冰冷的千芯光缆、管道与芯片，但交付大型 AI 基础设施的本质是处理复杂的人际协作。面对道路封锁引发的居民情绪、错综复杂的工期冲突与多方协调，成败的关键在于所有人能够坐在一起坦诚沟通优先级，共同寻找解决问题的平衡点。 第三，是彻底拥抱全新的协作模式。Meta 以往的骨干网建设主要是与传统电信基础设施运营商（telco carriers）合作，而此次 Prometheus 项目要求网络团队直接与重型土木工程承包商（heavy civil engineering firms）并肩作战，并且以前所未有的深度与市政许可审批部门（permitting authorities）进行高频联动。这彻底重塑了团队的作业方式。Fabrice 坚信，这绝不会是 AI 算力驱动下的最后一个复杂基础设施工程，未来的挑战只会更多，而团队已经为应对下一轮极限创新做好了准备。 ## Notable quotes > [00:03](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=3s) 如果公司要求你去完成一项前所未有的任务，去负责一个极其庞大复杂的基础设施项目，以支撑全公司的 AI 加速推进，你会怎么做？ — Fabrice Ouandji > [03:51](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=231s) 在 Prometheus 仅 5 公里的核心半径范围内，他们总共将铺设 210 万光纤公里（fiber kilometers）的光缆，这一长度足足可以在地球与月球之间往返三次。 — Sebastian Gault > [06:46](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=406s) 如果现场有来自俄亥俄州的朋友，对于我们在当地制造的所有麻烦和拥堵，我深表歉意。 — Fabrice Ouandji > [06:46](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=406s) 如果你站在施工沟槽边缘往下看，这个巨大的地下空间足以把三辆标准尺寸的 SUV 垂直叠放进去，最顶上一辆的车顶刚好与路面持平。 — Fabrice Ouandji > [08:59](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=539s) 只挖一次（dig once） — Fabrice Ouandji > [12:02](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=722s) 这种机会你一生只有一次 — Fabrice Ouandji > [12:02](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=722s) 根本不是普通市政道路上随处可见的小井盖，它们的物理体积堪比一套“单身公寓（studio apartment）”。 — Fabrice Ouandji > [13:32](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=812s) 方案很完美，一定能成 — Sebastian Gault > [15:44](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=944s) 数以万计的 GPU 就只能闲置在那里，连指示灯都无法闪烁 — Sebastian Gault > [17:40](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=1060s) 当你做一件前人从未做过的事情时，你能学到什么？ — Fabrice Ouandji > [17:40](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=1060s) 多度量，但关键时刻敢于落剪（measure twice, but do not be afraid to cut） — Fabrice Ouandji > [17:40](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=1060s) 每个人都有一个计划，直到嘴上挨了一记重拳（Everyone has a plan until they get punched in the mouth） — Fabrice Ouandji > [17:40](https://www.youtube.com/watch?v=N7G3yBolmsQ&t=1060s) AI 的底色是人与人的连接（AI is really about human connections） — Fabrice Ouandji ## Mentioned Fabrice Ouandji, Sebastian Gault, Meta, Prometheus, Zayo, Lumen, AFL, Google, 微软 (Microsoft), 亚马逊 (Amazon), 英特尔 (Intel), 新奥尔巴尼 (New Albany), 阿什维尔 (Asheville), 161 号公路 (Highway 161), 迈克·泰森 (Mike Tyson) ## Video description Fabrice Ouandji and Sebastian Gault from Meta present "Fiber Networks Innovations for AI at Scale" live at the Santa Clara Convention Center. Prometheus introduced significant complexity for connectivity. We were faced with significantly higher fiber count requirements, highly dynamic demand location, and compressed timelines for delivery. Tune in to find out the innovations required to meet the demand across multiple data centers in the cluster. Learn more about @Scale here: atscaleconference.com [Source](https://www.youtube.com/watch?v=N7G3yBolmsQ)

[0:03] Fabrice Ouandji: 如果公司要求你去完成一项前所未有的任务，去负责一个极其庞大复杂的基础设施项目，以支撑全公司的 AI 加速推进，你会怎么做？ [3:51] Sebastian Gault: 在 Prometheus 仅 5 公里的核心半径范围内，他们总共将铺设 210 万光纤公里（fiber kilometers）的光缆，这一长度足足可以在地球与月球之间往返三次。 [6:46] Fabrice Ouandji: 如果现场有来自俄亥俄州的朋友，对于我们在当地制造的所有麻烦和拥堵，我深表歉意。 [6:46] Fabrice Ouandji: 如果你站在施工沟槽边缘往下看，这个巨大的地下空间足以把三辆标准尺寸的 SUV 垂直叠放进去，最顶上一辆的车顶刚好与路面持平。 [8:59] Fabrice Ouandji: 只挖一次（dig once） [12:02] Fabrice Ouandji: 这种机会你一生只有一次 [12:02] Fabrice Ouandji: 根本不是普通市政道路上随处可见的小井盖，它们的物理体积堪比一套“单身公寓（studio apartment）”。 [13:32] Sebastian Gault: 方案很完美，一定能成 [15:44] Sebastian Gault: 数以万计的 GPU 就只能闲置在那里，连指示灯都无法闪烁 [17:40] Fabrice Ouandji: 当你做一件前人从未做过的事情时，你能学到什么？ [17:40] Fabrice Ouandji: 多度量，但关键时刻敢于落剪（measure twice, but do not be afraid to cut） [17:40] Fabrice Ouandji: 每个人都有一个计划，直到嘴上挨了一记重拳（Everyone has a plan until they get punched in the mouth） [17:40] Fabrice Ouandji: AI 的底色是人与人的连接（AI is really about human connections）

[0:03] Fabrice Ouandji: What if you were asked to do something that’s never been done before in support of a massive AI push? [3:51] Sebastian Gault: Fiber Kaizen [6:46] Fabrice Ouandji: mega duct bank [8:59] Fabrice Ouandji: dig once [17:40] Fabrice Ouandji: Measure twice, but don't be afraid to cut [17:40] Fabrice Ouandji: Everyone has a plan until they get punched in the mouth [17:40] Fabrice Ouandji: AI infrastructure relies on human connections [17:40] Fabrice Ouandji: Embrace new ways of working

[0:03] What if you were asked to do something [0:05] that's never been done before? [0:08] What if you were asked to work on a [0:10] complex and ambitious infrastructure [0:13] project in support of the company's [0:16] uh push for AI acceleration. [0:20] Hi everyone, my name is Fabris Wanji. [0:22] I'm a partner manager within the network [0:25] investments team at Meta. My team is [0:28] responsible for delivering connectivity [0:30] to interconnect our data centers and our [0:33] point of presence at POPS globally. So [0:36] that requires terrestrial fiber in some [0:38] cases but also submarine cables. Today [0:42] my colleague Sebastian and I are excited [0:44] to come and talk to you about some of [0:46] the innovations that we had to drive in [0:49] order to support Prometheus which is a [0:52] company's first AI training [0:54] supercluster. [0:57] So before we start talking about [0:59] Prometheus, it'd be interesting to kind [1:01] of talk about where we started from as [1:03] far as fiber demand is is concerned. So [1:07] if I take you back to about 2018 [1:11] when a data center or a pop required [1:14] connectivity, we needed to deliver about [1:17] 10 very small number of fiber strands on [1:20] two diverse path to connect that data [1:23] center for day one connectivity. The way [1:25] we would solve it is we'll go work with [1:27] carriers who are partners like Zo and [1:30] Lumen and a few other in the industry [1:32] and we'll contract them to go build and [1:35] deliver that connectivity. [1:39] Starting in about 2018 or so, we started [1:42] seeing a bit of growth in the demand, [1:44] right? So probably connecting more data [1:46] centers, we probably had more pops [1:49] coming up. And anytime we needed to [1:51] deliver a new route, we started to see [1:53] the demand grow up. And what we started [1:56] to do is we embarked on this huge build [1:59] plan called MMI, middle mile [2:01] infrastructure, and we'd go and partner [2:03] again with construction companies, [2:06] telecom carriers to build fiber [2:09] infrastructure. So they go and apply for [2:12] permits, open up the roads, place a [2:15] couple of conduits, each one of them [2:16] about an inch and a half or inch and a [2:18] quarter, and then we'll install one [2:20] cable with about 864 strengths for and [2:24] this was enough to really cover our day [2:26] one needs and probably a bit more. And [2:29] we're doing pretty well. The vendors [2:31] tell us told us that we're probably the [2:33] only ones doing that in the industry. [2:35] We're building probably the largest [2:37] fiber network um globally. [2:41] Fast forward to 2025, Prometheus came [2:44] about and completely [2:47] shifted our model. Again, we went from [2:50] needing about 864 fibers to 100,000. [2:55] Let me repeat that. We were required to [2:58] deliver 100,000 fiber strands between [3:01] data centers in support of Prometheus. [3:04] That's a 100x in capacity in terms of [3:08] growth. So how did we do that? We went [3:11] about to build some of the largest duck [3:14] banks that we've ever seen in the public [3:16] right away in Wabin, Ohio. [3:19] So now that we've grounded in Prometheus [3:21] and kind of what it means, the rest of [3:23] the talk would talk would cover a few [3:25] topics. Seth would take you through some [3:28] of the work around innovation for the [3:30] fiber cable that was required for [3:32] Prometheus. uh come back and cover some [3:35] of the civil engineering work. We talk [3:37] about construction that was involved in [3:39] this. Uh Seb will come back and talk [3:41] about placing the fiber in conduits and [3:44] some of the work that had to happen. And [3:46] then we'll close with some uh lessons [3:48] learned. With that, I'll turn it over to [3:49] Seb. [3:51] Thank you, Fris. [3:54] As you heard for Breeze, he said we were [3:56] tasked with something that was never [3:57] been done before. [3:59] So I'm going to walk you through the [4:01] fiber innovation and the fiber kaizen [4:04] we've done to find the best cable to use [4:08] in this deployment to ensure a smooth [4:12] deployment as well. [4:15] We did a fiber kaizen kaizen being the [4:18] Japanese word for continuous innovation. [4:22] We invited three of our cable [4:24] manufacturers to bring what they think [4:26] is the best cable for this. So, [4:31] we met in Asheville, North Carolina at [4:33] our [4:37] conduit vendors test track. We've asked [4:40] them to place the same conduit that [4:42] we're placing [4:44] for Prometheus. [4:47] They all jet at the same distance. The [4:48] track was 3,700 ft long. They all jet at [4:51] the same distance, but we had to make a [4:54] decision on what cable to take. Two of [4:56] the manufacturers brought a micro cable. [4:58] Micro cable being very small, very [5:01] fragile, prone to kink. [5:04] And one vendor brought a standard 864 [5:08] fiber cable 15 mm in um in diameter. We [5:13] chose that cable because of the more [5:15] rigid design. And I'm going to go [5:17] through at the fiber placement is we're [5:20] gonna we're going to be placing in phase [5:21] one 14864 fibers cables. [5:28] After we chose this cable, we start [5:30] working with AFL [5:32] who we chose for the 864 to come up with [5:36] a new design and develop a 1728 cable. [5:41] This innovation allowed us to double our [5:44] fiber capacity and hence preserve [5:46] conduit for optionality. [5:49] It opens up way more for future [5:52] scalability in the rest of the network [5:54] as we can use it in all of our other [5:56] deployments across the US as well. [5:59] So as you can see on the slide, the 864 [6:03] was a 15 millimeter, the 1728 is a 16.4. [6:07] So we only added 1.4 4 mm of diameter, [6:13] but we're doubling the fiber capacity. [6:19] In our 5 km radius in Prometheus, we're [6:22] going to be placing 2.1 million fiber [6:26] kilometers. So, let let that sink in. [6:30] 2.1 million fiber kilometers. That is [6:34] three times to the moon and back. I'm [6:38] going to be handing back over to Fbreze [6:40] and he's going to talk you through [6:42] everything we had to do at civil [6:44] engineering. [6:46] >> Thank you, Seb. So, phase one, you heard [6:50] from Seb, we needed a connect one data [6:52] center and we needed about 140 cables on [6:56] day one. Again, contrast that to pre2024 [7:00] where we're placing one cable and this [7:01] is all in the public right away. So, how [7:04] do you create capacity? That's 100x what [7:07] you used to do. You build a mega duck [7:09] bank. And that's what we did in New [7:11] Albany. If anybody lives in Ohio, we [7:13] apologize for all the trouble that we [7:15] created there. But if you were standing [7:20] in front of this trench, right, which [7:22] basically took up a lane of traffic, and [7:24] you were looking down, you'd be looking [7:26] about 20 ft deep cuz that's how far down [7:29] we had to go. In some cases, that trench [7:32] was about 9 ft wide, which meant in some [7:35] cases the rightway, which is the grass [7:38] where you typically trench, was full of [7:40] utilities and would not accommodate a 9 [7:42] ft wide trench. We had to go take lanes [7:44] of traffic, cut open the road, and drop [7:48] this 20 ft deep. How big is that? [7:52] That's basically you could basically fit [7:54] three standardsiz SUVs and they will [7:57] rise to the top of the surface. That's [7:58] how wide it was. And again, taking you [8:00] back to what we used to build. [8:02] Typically, we build a trench that's [8:04] about 6 in wide and we probably go about [8:08] 3 to 4 ft deep. And that was enough to [8:10] place a couple of candidates and on [8:12] cable. We went much wider than that. [8:18] How do you do this? You really bring [8:20] civil engineering expertise. This is no [8:22] longer telecom infrastructure. [8:24] This is a type of work that's typically [8:27] required for building sewers to build [8:30] water lines to to build storm systems. [8:34] And we basically had to go to a [8:36] different type of partners. So we found [8:38] a company that has deep pun intended [8:41] expertise in trenching massive [8:44] holes in the ground. It so happened that [8:48] um their name actually includes the word [8:49] excavating which is exactly what was [8:51] required here. we're able to partner up [8:53] with them again to be able to deliver [8:56] this type of conduit bank. [8:59] I think once we got started we faced a [9:02] couple of challenges. I think the first [9:04] one was around really the volatility of [9:06] the demand. Seb talked about phase one [9:09] phase two. Phase one was one data center [9:12] that we were asked to connect. [9:15] Within a few months we realized that [9:17] there was going to be multiple data [9:18] centers, right? So Prometheus is really [9:20] about interconnecting multiple data [9:22] centers that are within a 5 kilometer [9:25] radius which is the the picture on the [9:27] right side of the the slide there. The [9:29] challenge is that as we're working on [9:31] phase one, we did not know where the [9:34] phase 2 data centers would be. We did [9:36] not know how many fiber strands they [9:38] would require, but we're pretty sure [9:41] that they were probably going to require [9:44] using the same roads that we're working [9:46] on for phase one. So, you don't know [9:50] where the demand is. You don't know how [9:51] big the demand is going to be. You just [9:53] know it's going to be huge in the [9:54] hundreds of thousands. At the same time, [9:56] you probably get to open up those roads [9:58] once. No one wants to have, you know, [10:02] construction for three years with no [10:04] end. So we had to kind of manage that [10:06] and the way we did that was by going and [10:08] engineering every road in that 5 [10:11] kilometer radius that you see on the [10:12] right side of the slide. So what that [10:14] meant is doing a deep study of existing [10:18] utilities along each of those roads so [10:20] we could start to understand where we [10:22] could place a duck bank on the side of [10:24] the road which is a preference or [10:27] potentially going to the city and asking [10:29] for permits to trench the road which is [10:32] very complicated. So that was kind of [10:35] the first challenge dynamic demand dig [10:38] once approach only and we did it by [10:40] engineering. [10:43] The second part of the man the second [10:45] challenge was around just management [10:49] traffic. If I talk about that 5 [10:52] kilometer radius, you have Meta, Google, [10:56] Microsoft, Amazon all have existing [10:59] facilities or facilities under [11:01] construction. [11:03] On top of that, you have a couple of [11:05] data center providers who also have [11:07] facilities under construction. [11:10] As if this wasn't enough, in the [11:12] northeast quadrant of that radius, Intel [11:17] has a multi-billion dollar investment [11:19] constructing a a foundry. [11:22] So when you put all of that together, [11:24] you're talking about upwards of 5,000 to [11:28] 10,000 trades that are trying to get to [11:31] somewhere within that quadrant every [11:34] day. [11:35] This is in the midst of us potentially [11:38] shutting down some roads. So, as you can [11:41] imagine, there's a lot of conflicts, [11:43] right? Sometimes the conflicts are [11:44] internal. [11:46] I want to work on a road. The data [11:48] center team wants to work on their site. [11:50] How do you resolve that? In some cases, [11:53] external conflicts. Maybe we have to [11:55] change our design to accommodate another [11:58] project that the city had. So that [12:00] required a lot of engagement both with [12:02] internal team but also with external [12:05] teams. [12:08] So again more pictures of construction [12:12] we had to cross a highway and how to [12:15] build basically what is a tunnel on a [12:18] highway. The reason to do this is [12:20] because you only get to do this once. [12:22] So, what you see on the left side of the [12:24] photo is a five feet wide uh tunnel that [12:28] we built under Highway 161 in New [12:30] Albany. That was wide enough for the [12:33] operator to sit on the machine, drive [12:36] through it, and pull the the the dirt [12:38] out to dig through the tunnel. Huge [12:41] assive. And once you can have that, you [12:43] can pack as many conduits as you can to [12:46] serve your day one needs, but also build [12:49] optionality for the future. [12:51] What you see on the right is what we [12:53] call vaults. Actually, we we tend to [12:56] call them hand holes and manholes, but [12:59] they're not the typical hand holes that [13:00] you drive on in your street. These are [13:03] the size of a studio apartment, and you [13:06] need that many of them to house the the [13:09] huge number of conduits that we had. And [13:11] these allow those are basically access [13:13] points for you to connect to the system [13:16] after it's installed. [13:18] So that was the the civil engineering [13:21] that was required to create the conduit. [13:24] Once that was done, we had to go and [13:25] start placing cables in the conduit. And [13:29] Seb is going to talk to you about that. [13:32] >> Thanks, Free. [13:34] As you guys saw that we had to close [13:36] down a lot of roads in the city of New [13:38] Albany, and as you can imagine that a [13:41] lot of residents that still live there [13:42] are not very happy about that. [13:45] When we started with the fiber jetting [13:47] and the fiber placement, [13:51] this was my biggest risk register on the [13:54] whole project. And I've mentioned this [13:56] multiple times to our internal teams [13:58] that we do not know how this is going to [14:00] go. I mentioned it before. We're placing [14:02] 140 864 cables day one for phase one. [14:08] So I coordinated a design and a field [14:11] trial. We did a whiteboard session, [14:13] everything nice and tidy, [14:16] drew everything out and be like, "Okay, [14:18] good. We know where we're going to place [14:19] our jetting machines. [14:22] This is going to work." Was our thought. [14:26] It didn't. [14:28] So, the first tests have showed that we [14:32] needed to move the jetting machines [14:33] closer. We had them set at around 3,800 [14:37] ft apart. And we also placed the jetting [14:40] machines into the vaults instead of [14:43] having outside because we didn't want to [14:45] shut down lanes. We if you have to shut [14:48] down a whole lane for the whole route, [14:51] we would have incurred more traffic and [14:54] more stuff on the residents again. So we [14:58] placed them into the vaults and it's [14:59] called cascading that we don't leave any [15:01] slack either in the vaults [15:05] which made the monitoring of traffic [15:08] only about 100 to 200 ft. [15:13] So [15:15] going back to the jetting part is like I [15:19] said we place him around 37 3800 ft. We [15:23] found out that the cable was actually [15:24] too heavy to lift it back up. So, we had [15:28] to move jetting machines. We had planned [15:30] between three and four jetting machines. [15:32] Now, we had five and six because we had [15:34] to place them closer together. And we [15:36] found the sweet spot [15:38] to be about 25 2600 ft. [15:44] Mentioned this a couple times now. 140 [15:46] 864s. [15:49] We went through the jetting, how we [15:52] placed the machines. [15:54] Now, our risk, biggest risk was how many [15:58] cables can we jet per day because we had [16:01] a deadline that we had to meet cuz you [16:03] don't want two data centers sitting [16:05] there not connected and no GPUs [16:08] blinking. [16:11] We found out that we can jet about five [16:13] to six cables a day, which was actually [16:15] perfect to get through our to our [16:18] deadline. [16:21] Then we also had the cam uh cable [16:23] management inside of the vaults which I [16:25] already talked about. We didn't we [16:26] didn't leave any slack coils. We just [16:28] pulled it straight through. No slack. So [16:31] if we have to if we have a damage on a [16:34] cable, we have to rip it out and put a [16:36] new one in cuz there's no slack. No [16:39] place for any splice enclosures. [16:42] On phase two, we've placed 76 1728s. [16:46] This is the newly developed cable I [16:48] talked about all the way at the [16:49] beginning. [16:52] We had the same game plan. We're going [16:54] to place the machines in the same [16:55] distances. [16:57] Everything's going to work out. Of [16:59] course, it didn't because this cable is [17:01] almost double the weight even though [17:03] it's only 1.6 1.4 mm thicker, but it has [17:08] double the fibers. [17:11] So, it was a trial and error again. We [17:13] then found the sweet the sweet spot to [17:16] be around 1,800 ft. So, we needed even [17:19] more jetting machines because the cable [17:21] is so heavy to lift it back up with the [17:23] air. [17:26] So, you can imagine how complicated all [17:29] this was to get to this point. So, I'm [17:32] going to hand it back over to Fris and [17:34] he's going to take you through the [17:36] takeaways and what the future brings. [17:38] >> Thank you. [17:40] So what do you learn when you do [17:43] something that's never been done before? [17:45] Um in our case I think we had three [17:47] takeaways. The first one is really about [17:50] measuring twice but at some point to not [17:52] be afraid to cut right. So you heard [17:55] from Sarah a lot of planning went into [17:57] this and it was appropriate because it [17:58] was the first time this was being done [18:01] field trial right before the fiber [18:03] placement kaizen to select the fiber [18:06] cable. Uh but at some point you have to [18:08] jump in. I think Mike Tyson said [18:10] everyone has a plan until they get [18:11] punched in the mouth. I think what was [18:13] important for us was get to the cutting [18:16] and be flexible and adapt to the [18:19] conditions in the field. That was kind [18:20] of the first learning. The second one [18:22] which might be a bit of a surprise and I [18:24] think the way we sum it up is that at [18:26] the end of the day it's really about [18:27] human connections talked about all the [18:30] coordinations that had to the [18:31] coordination that had to happen to help [18:34] minimize impact of traffic and help make [18:37] sure that schedules were aligned and you [18:41] know conflicts arose. But really what [18:43] was important is human beings coming [18:45] together and talking about their [18:47] priorities and finding a a way to get [18:51] things done together. So that's the [18:53] second piece. AI is really about human [18:55] connections. Delivering the [18:57] infrastructure for AI is really about [18:59] human connections. And the third one [19:01] which seems obvious but bears repeating [19:04] is that we really need to embrace new [19:06] ways of working. We had never worked [19:09] with a deep heavy civil engineering [19:12] firm. We worked with telecom [19:13] infrastructure carriers. That was a [19:15] change and that was a learning for us. [19:18] We had never worked as closely as we did [19:20] with uh permitting authorities. We had [19:23] to do that. We had to build those [19:25] relationships and constantly engage them [19:27] so that we can so that the project could [19:29] be successful. [19:31] This is probably not the last of the [19:33] type of complex projects that AI [19:35] requires. I'm sure there'll be plenty of [19:37] challenges, but u we're always excited [19:39] to solve them. With that, we thank you [19:41] for your time.