Meta Links Regional Data Centers to Build Gigawatt AI Clusters
@Scale
Meta is expanding its AI network architecture to interconnect leased and owned facilities across regional rings to overcome single-site infrastructure limits.
## Background
Published by the @Scale channel on YouTube on August 27, 2026, this keynote features Omar Baldonado, lead for Meta's global data center and AI networking groups. The video presents a high-level overview of Meta's rapidly expanding global data center networks that power its AI models and family of apps, highlighting open-source collaborations and the scale of their infrastructure.
## TL;DR
Omar Baldonado walks through the dramatic evolution of Meta's AI infrastructure, from running 1.2 million H100-equivalent GPUs at zettaflop scale to developing gigawatt-scale clusters for complex agentic AI workloads. He details Meta's dual-pronged philosophy: 'Network for AI' (scale-up via custom MTIA silicon, scale-out with Prometheus RDS sites, and scale-across regional backbones) and 'AI for Network Infrastructure' (using AI agents to automate lifecycle management). Ultimately, he frame this massive infrastructure push as the engine driving democratic AI access for Meta's 3.5 billion global users.
## Atmosphere & style
The presentation carries an energetic, forward-looking, and collaborative vibe typical of a tech leadership keynote. Baldonado speaks with grounded optimism and technical pride, balancing high-level strategic visions with authentic engineering realities like un-sanitized rack photos and civil engineering hurdles. The tone feels visionary yet pragmatic, reflecting both the relentless pace of hyper-scale AI growth and a shared mission to democratize technology for billions.
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=0s) Meta operated the compute equivalent of roughly 1.2 million NVIDIA H100 GPUs by late last year, placing its total aggregate compute capacity firmly in the zettaflop scale.
- [02:49](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=169s) Meta's AI clusters have rapidly evolved from 8K–16K GPUs in 2023 to 129K GPUs in 2024, progressing toward gigawatt-scale infrastructures in 2026.
- [02:49](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=169s) AI network engineers must account for additive traffic patterns over time, ranging from early pre-training and disaggregated inference to reinforcement learning and complex agentic AI workloads.
- [02:49](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=169s) The emergence of Agentic AI demands concurrent long-context inference, massive Key-Value (KV) cache transfers, persistent storage interaction, and intensive CPU-to-GPU communications within clusters.
- [08:22](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=502s) Meta's infrastructure architecture relies on a full-stack co-design strategy that optimizes both 'Network for AI' (scale-up, scale-out, scale-across) and 'AI for Network Infrastructure'.
- [10:08](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=608s) To support diverse workload requirements, Meta is iterating rapidly through four generations of its custom MTIA (Meta Training and Inference Accelerator) silicon in two years.
- [11:34](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=694s) Meta leverages Rapid Deployment Structures ('tents') to bring online high-density data center capacity like the Prometheus 1-GW cluster in months rather than years.
- [11:34](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=694s) Meta introduced the Metari Transport Protocol as a purpose-built, high-performance transport layer developed alongside key ecosystem partners like AMD and Keysight.
- [13:50](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=830s) Building regional scale-across backbones requires overcoming severe physical and civil engineering constraints, including regional heterogeneity across air-cooled, liquid-cooled, and temporary structures.
- [15:58](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=958s) Autonomous AI workflows are increasingly deployed across the entire network life cycle, from high-dimensional planning to automated provisioning and real-time operational diagnostics.
- [18:36](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=1116s) To rapidly scale capacity, Meta combines owned data centers, leased space, and cloud partnerships into regional ring architectures that span hundreds to thousands of kilometers.
- [21:06](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=1266s) Meta's gigawatt-scale infrastructure push is fundamentally aimed at democratizing high-performance AI access across its global user base of 3.5 billion people.
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=0s) Introduction and the Explosive Scale of AI Infrastructure
Omar Baldonado, lead for Meta’s data center and AI networking groups, kicks off the **Networking @Scale** conference by welcoming attendees to one of his favorite events of the year. He emphasizes Meta’s long-standing commitment to open knowledge sharing across the networking community, echoing earlier remarks by Mcnesh. Reflecting on past presentations, Baldonado recalls how Gaia had shared a capex chart of major hyperscalers at the previous event. Baldonado explains that while he tried updating those capex numbers over recent months, the pace of AI infrastructure buildouts is moving so fast that data becomes obsolete within weeks.
To frame the sheer scale of modern operations, he shifts focus to orders of magnitude: the industry has rapidly escalated from millions to billions of dollars, megawatt scale to gigawatt scale, and into astronomical compute totals. By the end of last year, Meta alone operated the equivalent of 1.2 million NVIDIA H100 GPUs. Given that an H100 delivers roughly one petaflop depending on numerical encoding, 1.2 million H100-equivalent accelerators push Meta’s total compute footprint solidly into the **zettaflop scale** ($10^{21}$ FLOPS).
* **Zettaflop-Scale Infrastructure:** Meta operated ~1.2 million NVIDIA H100 GPU equivalents by late last year, putting aggregate compute capacity in the zettaflops range.
* **Custom Silicon & Strategic Partnerships:** Meta continues to deepen compute partnerships with major CPU vendors while concurrently designing custom CPUs and accelerators in-house.
* **America's Workforce Academy:** Meta launched this training initiative to build specialized human expertise across manufacturing centers, warehouses, and data center sites required to physically deliver gigawatt-scale infrastructure.
[02:49](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=169s) Historical Evolution of Meta’s AI Clusters (2023–2026)
Baldonado takes the audience on a retrospective walk down memory lane, tracking the rapid growth of Meta's largest AI training clusters over consecutive Networking @Scale events:
```
+-------------------------------------------------------------------------------+
| META AI CLUSTER SCALE EVOLUTION |
+-------------------------------------------------------------------------------+
| 2023: 8K - 16K GPUs --> Ethernet for AI Training (Post-COVID return) |
| 2024: 24K - 129K GPUs --> Llama 3 Training on twin 24K clusters |
| 2025: 100K+ GPUs --> Gigawatt fabrics, backbone scale-across |
| 2026: Gigawatt Scale --> Heterogeneous workloads & Agentic AI shift |
+-------------------------------------------------------------------------------+
```
The driving force behind this relentless infrastructure growth is the rapid evolution of the AI workloads themselves. Networking engineers can no longer view the cluster as a static transport pipe; they must intimately understand workload traffic patterns, which accumulate additively year over year:
* **2022:** Focused primarily on large-scale pre-training workloads, forcing network teams to master parallelization strategies like **tensor parallelism**, **pipeline parallelism**, and **data parallelism**.
* **2023–2024:** Marked the rise of **disaggregated inference** (separating prefill and decode phases) to handle models too large for single host systems, alongside distinct traffic requirements for Large Language Models (LLMs) versus Recommendation/Ads systems.
* **2025:** Witnessed an explosion in architectural diversity, notably **Reinforcement Learning (RL)** and **Mixture of Experts (MoE)** models.
* **2026:** Defined by the rapid surge of **Agentic AI** (triggered by platforms like Open Claw, following early 2025 "DeepSeek moments"). Agentic AI demands long-context inference, massive Key-Value (KV) cache transfers, persistent storage interaction, and intensive CPU-to-GPU communications—all running concurrently with existing training and inference jobs inside gigawatt clusters.
> "It's not like I can say, 'Well, I learned about large pre-scale training in 2022, I don't need to use that in 2026.' You do. All these workloads are running simultaneously in large gigawatt-scale clusters."
When consulting with AI researchers and model developers, Baldonado notes that the classic engineering tradeoff—*"Good, fast, cheap: pick two"*—no longer applies. Modern AI teams demand all four core vectors simultaneously: **Performance**, **Reliability**, **Fast Cadence**, and **Flexibility**.
To satisfy these aggressive goals, Meta structures its operational approach around two complementary domains:
1. **Network for AI (NI for AI / NA for AI):** A full-stack co-design effort spanning scale-up, scale-out, and scale-across domains to deliver high-performance hardware and software transport.
2. **AI for Network Infrastructure (AI for NI):** Leveraging internal AI models and autonomous agentic workflows to optimize the entire network life cycle (planning, design, delivery, and ongoing operations).
[10:08](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=608s) Scale-Up: Meta Training and Inference Accelerator (MTIA)
In the scale-up domain (within the rack), Baldonado highlights Meta’s in-house **MTIA** (Meta Training and Inference Accelerator) program. Demonstrating extreme execution speed, Meta announced plans to iterate through **four silicon generations in two years**.
While initial MTIA generations targeted recommendation and training workloads, latest research papers and engineering updates (presented at the conference by Wes and Rajie) detail how MTIA architecture has been dynamically adapted to handle inference workloads, underscoring the vital need for hardware flexibility. Baldonado lightheartedly insists on showing un-sanitized photos of production MTIA rack interiors, complete with dense cabling, emphasizing that robust physical networking is where real-world reliability is won or lost.
[11:34](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=694s) Scale-Out: Prometheus Gigawatt Site, Fabrics & Transport Innovations
Moving to scale-out infrastructure (at the building level), Meta focuses on rapid deployment and high-density networking:
* **Prometheus 1-GW Cluster (Ohio):** Meta utilizes **Rapid Deployment Structures (RDS)**—formerly referred to as "tents"—to bring online data center capacity in a matter of months rather than traditional multi-year construction timelines. Multiple RDS units are fully lit up and actively training Meta's internal **Muse** suite (including Spark and Glimmer open-weights models).
* **Fabric Diversity:** Architecture spans both Disaggregated Scheduled Fabrics (DSF) and non-scheduled fabrics. (Baldonado also notes Google’s presentation on their new AI backend fabric).
* **Metari Transport Protocol:** Highlighted in a fresh Meta technical release, **Metari** is a purpose-built transport layer designed specifically for ultra-high performance, multi-path scalability, and network reliability. Key ecosystem partners including AMD and Keysight collaborate closely on validating and deploying Metari at scale, while peers from OpenAI, Oracle, and CoreWeave participate in joint industry panels on scale-out transport.
Connecting multiple data center facilities across a region introduces severe architectural heterogeneity. Baldonado showcases a section of the Prometheus cluster footprint spanning three distinct regions within a single site:
```
+-------------------------------------------------------------------------------+
| PROMETHEUS REGIONAL HETEROGENEITY |
+-------------------------------------------------------------------------------+
| [H-Shape Buildings] --> Legacy Air-Cooled Data Centers (DSF Fabrics) |
| [Bottom Facilities] --> Modern Liquid-Cooled Data Centers (High Density) |
| [Top-Left Structures] --> Rapid Deployment Structures (RDS / Tents) |
+-------------------------------------------------------------------------------+
| * All tied together via high-bandwidth, low-latency regional optical links |
+-------------------------------------------------------------------------------+
```
Building out this high-density regional AI backbone requires solving tough physical infrastructure and civil engineering challenges. Conference sessions dive into navigating complex real-world obstacles such as trenching through frozen ground, securing environmental and municipal permits, and overcoming strict local property rights to deploy continuous optical fiber backbones.
[15:58](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=958s) AI for Network Infrastructure (AI for NI) & Life Cycle Management
In the final section of Part 1, Baldonado stresses that life cycle management of gigawatt-scale infrastructure is critical, yet frequently under-discussed. Managing thousands of racks across planning, design, deployment, and operation requires leveraging advanced AI agents:
* **Planning:** Solves high-dimensional optimization problems—matching exact hardware choices and fabric capabilities against capacity deployment timelines across hundreds of megawatts simultaneously.
* **Design & Delivery:** Automates configuration generation for heterogeneous sites and speeds up hardware commissioning. Fast delivery is vital so that network provisioning never becomes the bottleneck blocking new GPU deployments (such as NVIDIA GB200/GB300 series, AMD accelerators, or MTIA racks).
* **Operations (Run Phase):** Deploys agentic AI workflows to monitor runtime health, isolate hardware degradations, and manage real-time traffic anomalies across giant 129,000+ GPU topologies.
[18:36](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=1116s) Agentic AI and the Horizon of Scale-Across Infrastructure
Managing modern AI clusters requires orchestrating hundreds of thousands of components, spanning NICs, memory, storage, and optical components. To handle this staggering level of complexity, Meta is leveraging agentic AI to correlate data across multiple operational layers and pinpoint root causes. Operating at this scale is simply impossible without AI-driven automation and diagnostics. Conference attendees are encouraged to explore the dedicated track for "AI for Networking" alongside sessions detailing optimizations across the co-design stack.
Looking toward the future, the physical boundaries of AI infrastructure continue to expand. Drawing on a framework shared earlier in the summer by Meta infrastructure leader Sharad, the focus centers on bringing massive capacity online rapidly. Meta is achieving this by combining owned data centers, leased facilities, and partnerships with major cloud providers. Rather than turning up these disparate sites independently, Meta clusters them into regional rings that leverage shared services to accelerate capacity turnup. These interconnected rings signal the next frontier: a "scale-across" network architecture spanning hundreds to thousands of kilometers.
* **Scale of Components:** Clusters require managing hundreds of thousands of network interface cards (NICs), memory modules, storage drives, and optics.
* **Capacity Strategy:** Combines owned data center construction, third-party leasing, and cloud provider partnerships.
* **Geographic Horizon:** Regional ring configurations lay the groundwork for scale-across networks stretching thousands of kilometers.
[20:23](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=1223s) Networking at the Core of Global AI Infrastructure
The trajectory of AI infrastructure buildout follows a steep hockey-stick growth curve. While pinpointing the exact position on this curve is difficult, the sheer momentum makes network engineering one of the most exciting fields today. The network serves as the central thread holding the entire infrastructure together: it interconnects individual chips inside a rack, joins racks within a building, and links entire buildings together into massive unified clusters.
* **Hierarchy of Interconnects:** Networking spans three critical layers: chip-to-chip within a rack, rack-to-rack within a facility, and facility-to-facility across a multi-building cluster.
* **Infrastructure Momentum:** Network design remains the critical dependency for unlocking next-generation AI scaling.
[21:06](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=1266s) Democratic AI and the Mission to Connect Billions
The massive infrastructure effort is ultimately driven by real-world applications and transformative consumer experiences, such as the Meta Ray-Ban smart glasses. Beyond consumer hardware, AI holds the potential to disrupt and redefine major global industries, including manufacturing, medicine, and entertainment.
Meta's core mission centers on democratizing this technology. With 3.5 billion people using Meta’s family of applications to stay connected, the company aims to apply that same philosophy to artificial intelligence, ensuring that AI access is available to everyone for good. Providing open, high-performance access to 3.5 billion users represents both the primary motivator for building gigawatt-scale infrastructure and the underlying purpose driving Meta's network engineering teams.
> "At Meta, we have three and a half billion people using some form of one of our applications to connect and we bring that same philosophy of connecting people to AI. We want AI to be for everyone and ensure everyone has that access for good."
* **Consumer & Industry Impact:** Infrastructure powers products like Meta Ray-Ban glasses while accelerating transformation in medicine, manufacturing, and entertainment.
* **Scale Target:** Engineering systems designed to support a user base of 3.5 billion people worldwide.
* **Closing Call to Action:** Attendees are urged to engage with speakers, collaborate on shared industry challenges, and make the most of the @Scale conference.
## Notable quotes
> [02:49](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=169s) It's not like I can say, 'Well, I learned about large pre-scale training in 2022, I don't need to use that in 2026.' You do. All these workloads are running simultaneously in large gigawatt-scale clusters. — Omar Baldonado
> [21:06](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=1266s) At Meta, we have three and a half billion people using some form of one of our applications to connect and we bring that same philosophy of connecting people to AI. We want AI to be for everyone and ensure everyone has that access for good. — Omar Baldonado
Meta is expanding its AI network architecture to support gigawatt-scale, zettaflop compute footprints driven by increasingly complex workloads, according to Omar Baldonado, lead for the company's data center and AI networking groups.
Cluster Growth and Evolving Workload Demands (0:00)
Meta operated the equivalent of 1.2 million NVIDIA H100 GPUs by late last year, placing its compute footprint in the zettaflop scale. Baldonado tracks the expansion of Meta's training clusters from 8,000–16,000 GPUs in 2023 to dual 24,000-GPU clusters for Llama 3 in 2024, passing 100,000 GPUs in 2025 and moving toward gigawatt-scale fabrics. Alongside third-party partnerships, Meta is also building internal custom CPUs and accelerators, complemented by America's Workforce Academy to train personnel for physical deployments.
Additive Traffic Patterns and Dual-Domain Strategy (2:49)
Baldonado emphasizes that AI workload demands accumulate additively rather than replacing older network requirements. Cluster fabrics must simultaneously support pre-training parallelism from 2022, disaggregated inference and recommendation traffic from 2023–2024, Reinforcement Learning and Mixture of Experts architectures from 2025, and emerging Agentic AI workloads in 2026. Agentic workflows add intense CPU-to-GPU communication, persistent storage access, and massive Key-Value cache transfers to the network.
Multi-Tier Infrastructure Across Silicon and Regions (8:22)
To deliver performance, reliability, rapid cadence, and flexibility simultaneously, Meta divides network engineering into Network for AI and AI for Network Infrastructure. At the rack level, Meta's custom MTIA silicon is scheduled for four generations in two years, adapting dynamically from recommendation models to inference. At the building scale, the 1-gigawatt Prometheus site in Ohio uses Rapid Deployment Structures alongside Disaggregated Scheduled Fabrics and the Metari transport protocol, which is developed with partners including AMD and Keysight to optimize multi-path routing.
Heterogeneous Facilities and Civil Deployment Hurdles (13:50)
Expanding clusters across regions introduces severe physical and technical heterogeneity. Sites such as Prometheus link legacy air-cooled facilities, modern high-density liquid-cooled centers, and Rapid Deployment Structures over low-latency optical connections. Baldonado points out that scaling this regional backbone requires navigating difficult civil engineering realities, including trenching through frozen ground, securing environmental and municipal permits, and obtaining local property easements.
Automated Lifecycle Operations and Global Reach (15:58)
Managing hundreds of thousands of network interface cards, memory modules, storage drives, and optics requires integrating agentic AI into planning, configuration, and runtime anomaly detection. Looking outward, Meta aggregates owned sites, leased space, and third-party cloud capacity into regional rings that establish scale-across backbones stretching thousands of kilometers. Baldonado frames these massive infrastructure investments as essential to democratizing AI access for Meta's 3.5 billion global application users.
While Meta continues pushing physical infrastructure across chip, building, and regional domains, the operational feasibility of gigawatt-scale clusters increasingly relies on agentic AI to manage physical heterogeneity, civil engineering delays, and complex operational lifecycles.
## Background
该视频由 @Scale 频道于 2026 年 9 月 3 日发布在 YouTube 平台。视频内容为 Meta 数据中心与 AI 网络负责人 Omar Baldonado 的开场主题演讲，页面简介指出他统领的团队负责构建与运维支撑 Meta AI 以及旗下全系应用矩阵（Facebook、Instagram、WhatsApp、Messenger）的全球网络基建，持续研发世界顶级规模的算力集群并通过开源社区分享技术成果。
## TL;DR
演讲系统梳理了 Meta 在跨入 Zettaflop 算力时代后网络基础设施的跨越式演进，呈现了集群规模从 8K 扩展至 12.9 万卡乃至吉瓦级（Gigawatt-scale）的技术路径。面对多模态模型与智能体 AI（Agentic AI）带来的流量特征重构，Meta 确立了“以网驭智（NI for AI）”的全栈协同设计——涵盖机架级自研 MTIA 芯片快速迭代、建筑级快速部署结构（RDS）与 MetaURI 传输协议研发，以及跨区域异构超大规模机房的高速互联。同时，团队推行“以智治网（AI for NI）”，依托智能体全流程赋能网络规划、设计、交付与运维，并结合自建、租赁与云合作编织环形协同算力网络（Capacity Rings），最终践行让全球 35 亿人普惠享有先进 AI 生产力的宏大愿景。
## Atmosphere & style
整场演讲充满前沿极客风范与现代重工业美学的双重张力，Omar 的表达风趣自如且极富感染力，既有对机架内部“壮观线缆之美”的深情赞叹，也有对严寒冻土施工与千行百业赋能的硬核描摹。叙事弧线从算力狂飙的微观芯片与网络协议层层递进，直至拉升到跨广域的吉瓦级土木工程与连结数十亿人的全人类使命，听者能真切体会到处于 AI 浪潮核心位置的网络工程师所肩负的沉甸责任与技术自豪感。
## Bulletin
- [00:00](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=0s) Meta 算力储备已跨越 Zettaflop 门槛，截至去年底拥有折合 120 万张 NVIDIA H100 的等效算力，训练集群规模正从 12.9 万卡向吉瓦级（Gigawatt-scale）加速迈进。
- [08:21](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=501s) 为应对长推理链与智能体 AI（Agentic AI）带来的流量特征重构，Meta 提出“以网驭智（NI for AI）”与“以智治网（AI for NI）”双向闭环协同架构。
- [11:21](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=681s) Meta 推出专有传输协议 MetaURI 并落地快速部署结构（RDS），将原本以年为周期的数据中心机房建设压缩至数月即可投产上线。
- [10:08](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=608s) 自研芯片 MTIA 保持两年推四款的高速迭代，同时在跨园区级 Scale Across 网络中整合风冷、液冷及临时建筑，克服复杂光电与土木工程挑战。
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=0s) Meta 算力基础设施呈指数级扩张，算力总量达 120 万张 H100 当量（进入 Zettaflop 时代），供电与集群正全面迈向吉瓦（GW）级。
- [00:00](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=0s) 实体制造与运维人才成为扩张瓶颈，Meta 正式设立 America's Workforce Academy 为先进制造与布线运维输送专业技术力量。
- [02:42](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=162s) 网络架构演进始终由 AI 工作负载驱动，从 2023 年 8K-16K 以太网训练集群演变至双 24K（Llama 3）、12.9 万卡集群及未来的吉瓦级织构。
- [04:35](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=275s) AI 网络知识具备滚雪球式的累积效应：预训练并行切分、解耦推理、MoE 与强化学习等异构流量在吉瓦级集群中全天候高并发共存。
- [06:21](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=381s) 2026 年 Muse 矩阵模型与 Agentic AI 崛起，长推理链、海量 KV Cache 搬运与异构访问重塑通信特征，促使微流与突发流混合调优。
- [08:21](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=501s) 顶级模型研发团队对底层网络提出“高性能、高可靠、极速交付、极致灵活性”四项指标并重的严苛诉求，杜绝工程妥协。
- [08:21](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=501s) Meta 确立双向体系：通过全栈协同设计的“以网驭智（NI for AI）”与覆盖全生命周期自动化运维的“以智治网（AI for NI）”化解矛盾。
- [10:08](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=608s) 机架内 Scale-up 领域加速自研芯片 MTIA 创新，两年内连推四款硬件，并兼顾针对大语言模型与在线推荐系统训练和推理的灵活适配。
- [11:21](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=681s) 在 Scale Out 建筑层引入快速部署结构（RDS/Tents），将建筑周期从数年缩短至数月，已承载 Muse Spark 等生产级工作负载。
- [11:21](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=681s) 自研公布 MetaURI 专用传输协议，结合解耦调度式 Fabric（DSF）与非调度式 Fabric 解决大规模集群通信与运维难题。
- [13:50](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=830s) Scale Across 跨集群面临异构形态整合挑战，在俄亥俄 Prometheus 园区同一大区内熔接风冷、超高密液冷机房与 RDS 临时结构。
- [13:50](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=830s) 跨集群高密度互联不仅是高密光学与光纤工程的突破，还涵盖了冻土开挖、用地合规等极为庞大的现代土木工程考验。
- [15:58](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=958s) 将智能体 AI 全面嵌入网络规划、设计、交付与运行（Run）全生命周期，实现数十万张 GPU/NIC/光模块海量日志的秒级根因排查与自愈。
- [19:11](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=1151s) 算力上线走向“自建超大数据中心+租赁机房+公有云合作”多源协同，并构建跨越百至数千公里的环形集群（Capacity Rings）。
- [19:11](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=1151s) 超大规模基建的终极目标是落实“AI for Everyone”愿景，利用稳固的高性能网络支撑全球超 35 亿人平等享有先进生产力。
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=0s) 欢迎致辞与 Zettaflop 时代的算力基建飞跃
大会在热烈的掌声中拉开序幕，Meta 数据中心与 AI 网络负责人 Omar Baldonado 向全场参会者致以诚挚问候。他坦言这是自己一年中最期待的日子之一，因为 Networking @Scale 不仅是 Meta 网络团队展示成果的窗口，更是整个网络工程社区交流超大规模集群运营经验的平台。回顾去年的大会，Gaia 曾展示过全球超大规模云服务商（hyperscalers）资本支出（CapEx）的激增走势。Omar 幽默地提到，自己近几个月一直试图更新那张图表，但后来发现数据基本每隔几周就会彻底过时——AI 基础设施的扩建节奏已经快到了令人眩晕的地步。
面对这种指数级增长，Omar 用几组数量级的跨越描绘了行业的巨变：网络与基础设施规模从“百万级”迈向“数十亿级”，供电能耗从“兆瓦级”（MW）飙升至“吉瓦级”（GW），而在算力层面更是迎来了天文数字般的飞跃。到去年年底，Meta 拥有的算力规模已相当于 120 万张 NVIDIA H100 GPU。按照单张 H100 约 1 Petaflop 的计算吞吐量估算，Meta 的总算力储备已明确跨入 Zettaflop（十万亿亿次浮点运算）门槛。
这些惊人的数字最终转化为了扎实落地的工程项目。Omar 重点介绍了两项核心行动：其一是深度算力合作，Meta 不仅与各大 CPU 芯片供应商保持紧密协作，甚至深入到了自研定制 CPU 的硬件底层；其二则是对实体人才队伍的建设。在近期走访各处数据中心现场、物流中转仓和硬件制造组装中心后，Omar 感慨建设现代 AI 基础设施所需的人力规模超乎想象。为此，Meta 正式设立了“美国劳动力学院”（America's Workforce Academy），专门为制造中心和现场运营输送具备专业技能的工程技术人才。
- 算力总量跨越：截至上年底，Meta 拥有折合 120 万张 NVIDIA H100 的计算卡，总算力规模达 Zettaflops 级别。
- 实体与供应链建设：启动 America's Workforce Academy，解决先进 AI 制造、布线与现场运维的高技能技术人才缺口。
---
### [02:42](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=162s) 从 8K 到吉瓦级：Meta 训练集群的以太网演进史
Omar 带领现场观众重温了一段引人入胜的 Meta 训练集群发展史。他现场互动让 2023 年参加过疫后首次线下 Networking @Scale 的老朋友举手。在 2023 年那个当时被称为“AI 网络特辑”的会场上，团队热烈探讨的还是基于以太网（Ethernet）构建用于 AI 训练的基础设施，彼时集群的核心规模在 8K 到 16K 张 GPU 之间。如今回看，那种规模甚至显得有些“小巧古朴”。
到了 2024 年，集群规模迅速翻番。Meta 当时分享了基于两个 24K GPU 集群进行 Llama 3 训练的网络架构，并展示了一系列前沿技术攻坚。而到了去年，讨论的焦点直接扩展至 2024 年建成的 12.9 万张（129,000）GPU 超大集群，网络团队详细解析了如何在这一规模上稳定支撑生成式 AI 模型的训练。不仅如此，大会还开始密集前瞻面向未来吉瓦级（Gigawatt-scale）集群所必需的网络织构（fabrics）以及用于跨集群互联的骨干网（backbone）技术。整个演进路径极其清晰：是背后日新月异的 AI 工作负载，一直在直接推动着底层网络架构的重构与扩容。
```mermaid
flowchart LR
accTitle: Meta AI 训练集群规模演进
accDescr: 展示 Meta 从 2023 年的 8K-16K 集群演进至 129K GPU 以及未来吉瓦级集群的历史历程。
subgraph S1["2023 年"]
A["8K - 16K GPU 集群<br/>以太网 AI 训练探索"]
end
subgraph S2["2024 年"]
B["双 24K GPU 集群<br/>支撑 Llama 3 预训练"]
end
subgraph S3["去年 (2025 年)"]
C["129,000 GPU 集群<br/>大规模生成式 AI 训练"]
end
subgraph S4["未来展望"]
D["吉瓦级 (GW) 集群<br/>全新织构与骨干互联"]
end
A --> B --> C --> D
```
---
### [04:35](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=275s) 知识的滚雪球效应：多重 AI 工作负载的复合叠加
回顾这段历程，令 Omar 与广大网络从业者倍感振奋的一点在于，网络团队必须与模型算法团队同频共振，深入理解各种复杂的 AI 作业模型。2022 年前后，大型训练集群主要跑大规模预训练任务，网络工程师必须熟稔各种分布式并行切分模式，包括张量并行（Tensor Parallelism）、流水线并行（Pipeline Parallelism）以及数据并行（Data Parallelism）。
随后，工作负载的形态迅速演进。随着推理模型体量超出单机承载极限，2023 至 2024 年间，团队开始攻坚解耦式推理（Disaggregated Inference），深入将 Prefill（预填充）阶段与 Decode（解码）阶段在物理节点上解耦，同时兼顾大语言模型与在线广告推荐系统截然不同的通信模型。再往后，混合专家系统（MoE）与强化学习（RL）任务又迎来了集中爆发。
Omar 特别向网络工程师强调，这套知识体系是“累积式”（cumulative）的，绝对不存在“2022 年学会了分布式训练网络技术，到 2026 年就可以抛掉”的情况。在现如今的大型吉瓦级集群中，高密集度的数据并行预训练、解耦推理服务、MoE 以及强化学习是7×24小时全并发共存的。这就要求工程师必须把历年沉淀的各种流量工程策略与调度技巧悉数带入集群设计中。例如就在本月早些时候，Meta 广告工程团队分享了与网络团队协同的成功战果：他们将大语言模型训练中的通信优化经验成功迁移并定制到了广告模型训练中，使业务效率获得了质的突破。
---
### [06:21](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=381s) 2026 新常态：Muse 模型矩阵与智能体（Agentic AI）的流量重构
进入 2026 年，创新的步伐丝毫未减。在模型层面，Meta 先后推出了 Muse 系列前沿模型——涵盖面向图像与视频处理的 Spark、Glimmer，以及能在单卡 GPU 上独立部署运行的开源端侧模型。但在系统底层，2026 年又给网络工程师出了全新的复合命题：如果说 2025 年初的震撼是 DeepSeek 带来的架构冲击，那么 2026 年初最引人注目的莫过于 Open Claw 的突破以及 Agentic AI（智能体 AI）的全面崛起。
Agentic AI 极其依赖长周期的多步推理链（long-term inference）。这种计算范式彻底颠覆了传统的网络通信特征：网络侧不仅需要承载算子间交互，还必须实时应对海量的 KV Cache 搬运、超高频的远端持久化存储访问，以及 CPU 与 GPU 之间密集的跨总线与跨网络协同交换。
集群网络不仅要吞吐 2022 年那种高爆发的集中式训练同步流量，还必须同时处理极其碎片的异步微流，网络工程师以往在各种不同分布式场景中用过的“工具箱里的所有工具与杠杆”，现在必须每天在同一个 AI 算力集群里全盘调动。
---
### [08:21](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=501s) 模型团队的四重严苛需求与全栈协同设计
当 Omar 深入一线与 Meta 的 AI 研究员及模型开发者交流时，听到了对底层基础设施最严苛的呼声。这让人不禁想起工程界那句经典老话：“多快好省，任选其二”（"Good, fast, cheap — pick two"）。但在当下的顶级 AI 研发团队面前，妥协是不存在的。研究人员开门见山地要求必须同时满足四项核心指标：“高性能（Performance）、高可靠（Reliability）、极速迭代交付（Fast cadence）以及极致灵活性（Flexibility）——四项我们全都要”。
```mermaid
flowchart TD
accTitle: AI 基础设施全栈协同设计框架
accDescr: 展示模型研究员的四项核心诉求，以及由此驱动的以网驭智（NI for AI）与以智治网（AI for NI）双飞轮体系。
subgraph Requirements["模型团队的核心需求 (全部兼顾)"]
R1["高性能 (Performance)"]
R2["高可靠 (Reliability)"]
R3["极速迭代 (Fast Cadence)"]
R4["高灵活性 (Flexibility)"]
end
subgraph Solutions["系统应对架构"]
direction TB
subgraph NI["以网驭智 (NI for AI: 全栈协同)"]
ScaleUp["Scale-up: 机架内加速芯片级互连"]
ScaleOut["Scale-out: 集群级 Fabric 织构"]
ScaleAcross["Scale-across: 跨集群与数据中心骨干"]
end
subgraph AI["以智治网 (AI for NI: 闭环运维)"]
Ops["全生命周期智能化运维<br/>自愈与流量调度"]
end
end
Requirements --> NI
Requirements --> AI
```
为了从根本上化解这一矛盾，Meta 确立了贯穿全生命周期的双向体系：
1. **以网驭智（NI for AI / Network Infrastructure for AI）**：将底层芯片、整机机架、网络织构与编排软件打通，实施跨层级的全栈协同设计（Full Stack Co-design）。网络覆盖了机架内的 Scale-up、集群内的 Scale-out 以及跨地域集群互通的 Scale-across。
2. **以智治网（AI for NI / AI for Network Infrastructure）**：在网络自身的建设、监控与运维生命周期中，深度使用自研 AI 模型，驱动超大规模网络的自动化管控与自愈。
---
### [10:08](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=608s) Scale-up 领域的自研实践：MTIA 的急速迭代与线缆美学
在具体拆解全栈协同设计时，Omar 首先将目光聚焦在机架内部互联的 Scale-up 域。Meta 研制的自研芯片 MTIA（Meta Training and Inference Accelerator）正是满足“极速迭代交付”的最佳样本：Meta 在今年早些时候正式宣布，计划在短短两年时间内连推四款定制芯片。这种激进的迭代节奏在定制硅片历史上极为罕见，目的就是在最短时间内为算法研究员递上最称手的算力武器。
除了高算力与高可靠性，MTIA 的另一个核心特征在于“极高的灵活性”。第一代 MTIA 芯片在立项时重点针对训练场景设计，但在落地部署的过程中，团队敏锐捕捉到模型架构的突变，在随后的技术论文和技术博客中披露了如何通过架构调整，让这款硬件在兼顾训练任务的同时，高效支撑起大规模推理场景。
在展示 MTIA 实际落地部署的整机机架系统时，Omar 笑着分享了一个圈内观察：不少硬件厂商在对外发布 AI 机架宣传照时，习惯为了所谓的整洁美感而刻意拔除所有线缆；但他坚信在网络工程师的眼中，“那些密布于机架内外、纵横交错的线缆才是整套系统最为壮观的工程灵魂所在”（"all the glorious networking"），没有这些高速物理互连，任何算力都无法真正运转起来。
### [11:21](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=681s) 横向扩展（Scale Out）：建筑级网络与快速部署架构
打造兼具可靠性与极致性能的网络极为不易，但这正是前沿研究人员提出的硬性要求。在完成机架内互联的探讨后，视野扩展至“Scale Out”（横向扩展），即建筑级别的网络层（Building-level Network）。建筑层网络与物理建筑本身深度绑定，整个技术栈都迎来了密集的技术革新。
```mermaid
flowchart TD
accTitle: Scale Out Building Level Infrastructure
accDescr: 展示Meta建筑级网络中快速部署结构与网络Fabric及传输协议的构成
subgraph Facility["物理设施层 (Facility Level)"]
RDS["快速部署结构 RDS (Tents)<br/>月级交付上线"]
TradBldg["传统数据中心建筑<br/>年级建设周期"]
end
subgraph Fabric["数据中心 Fabric 网络"]
DSF["DSF 解耦调度式 Fabric"]
NonSched["非调度式 Fabric (Non-Scheduled)"]
end
subgraph Transport["传输层革新 (Transport Layer)"]
MetaURI["MetaURI 专有传输协议<br/>大规模性能、可靠性与运维"]
end
RDS --> Fabric
TradBldg --> Fabric
Fabric --> Transport
```
在建筑物理形态上，Meta 展示了位于俄亥俄州 Prometheus 1GW（吉瓦）集群的实景。为了以极快节奏跟上算力激增的需求，Meta 采用了名为“快速部署结构”（Rapid Deployment Structures，缩写 RDS，此前常被称为“帐篷” Tents）的建设方案。过去建设传统数据中心大楼通常需要数年时间（虽然传统建设仍在推进），但为了以最高速度提供算力，RDS 将建筑交付周期缩短至数月。目前已有许多 RDS 结构全面投入生产运行，支撑着 Muse Spark 以及集群内的多种核心工作负载。
- **快速部署结构（RDS）**：将传统以年计的机房建设周期压缩至以月计，全面承载一线生产算力。
- **多元化网络 Fabric**：深入到建筑内部，Meta 持续演进其解耦调度式结构（Disaggregated Scheduled Fabric, DSF）与非调度式结构（Non-scheduled Fabric）；同期 Google 也在探索面向 AI 后端的新型 Fabric。
- **传输层创新（MetaURI）**：Meta 正式公布了针对大规模性能、大规模可靠性与运维量身打造的传输协议 MetaURI，并与 AMD、Keysight 等产业伙伴共同推进软硬件落地。
- **行业协同探讨**：针对 AI 的大规模 Scale Out 网络与传输问题，OpenAI 亦将分享其传输设计，后续更设立了汇聚 CoreWeave、OpenAI、Oracle 和 Meta 的专属圆桌研讨。
---
### [13:50](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=830s) 跨集群扩展（Scale Across）：异构形态与土木工程挑战
当网络进一步超越单一建筑乃至单一大区时，便进入了“Scale Across”阶段——将各大超大规模集群紧密互联。许多人对跨集群的第一反应仅仅是“距离是多少公里”，但现实情况远比单纯的距离度量更加复杂和异构。
以 Meta 俄亥俄 Prometheus 园区的部分区域（含数百兆瓦）为例，航拍视角展现了三种不同时代的设施与异构网络在同一大区内的汇聚：
1. 顶部呈 H 型的建筑属于传统的风冷数据中心（Air-cooled Data Centers），内部搭载 DSF 架构；
2. 底部相对更新的建筑是高密度的液冷数据中心（Liquid-cooled Data Centers），其占地面积虽小，容纳的 GPU 规模却几乎等同于上方三栋 H 型建筑的总和；
3. 左上角则是快速搭建并投入运营的 RDS 帐篷结构。
```mermaid
flowchart LR
accTitle: Heterogeneous Scale Across Architecture
accDescr: 展示俄亥俄 Prometheus 园区中多种异构建筑形态跨距离高速互联的结构
subgraph RegionCluster["Prometheus 区域互联架构"]
direction TB
subgraph AirCooled["风冷机房 (H型建筑)"]
AC["DSF 架构数据中心"]
end
subgraph LiquidCooled["液冷机房 (高密新建筑)"]
LC["高密 GPU 算力集群"]
end
subgraph TentSite["RDS 临时结构 (Tents)"]
RDS_Node["快速上线集群"]
end
end
AirCooled <-->|"高密度高带宽 AI 骨干网<br/>(MetaURI + 定制光模块/光纤)"| LiquidCooled
LiquidCooled <-->|"大区内部跨距融合"| TentSite
```
将这数百兆瓦算力整合在一起，必须解决高密度光学收发模块（Optics Transceivers）、各异传输距离和差异化硬件拓扑所带来的极大异构性（MetaURI 在此类长距异构场景中优势显著）。除了光电软硬件，底层土木工程面临的物理挑战同样巨大：在高密度、高带宽 AI 区域骨干网的施工中，团队必须应对严寒冻土开挖、繁琐的施工许可办理以及复杂的地产土地所有权法律合规，“这完全是一场不可思议的现代土木工程壮举”。
---
### [15:58](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=958s) AI 赋能网络全生命周期（AI for Network Infrastructure）
构建吉瓦级集群不仅关乎网络技术本身，更关乎全生命周期的精细化治理。吉瓦级算力池是一笔极其宝贵的资产，不能对其管理掉以轻心。Meta 正在全流程深度引入智能体 AI（Agentic AI），赋能涵盖规划（Plan）、设计（Design）、交付（Deliver）与运维（Run）的网络全生命周期。
```mermaid
flowchart LR
accTitle: Agentic AI for Network Lifecycle
accDescr: 展示智能体AI在网络规划、设计、交付和运维四个阶段的赋能流程
Plan["1. 规划 (Plan)<br/>优化百兆瓦级容量分配与技术落地窗口"] --> Design["2. 设计 (Design)<br/>针对加速器选型、拓扑距离与 Fabric 生成专属方案"]
Design --> Deliver["3. 交付 (Deliver)<br/>边施工边投产，确保网络不阻塞 GPU 机架上线"]
Deliver --> Run["4. 运维 (Run)<br/>智能体AI关联分析数十万GPU/网卡/光模块，秒级定位根因"]
```
- **规划阶段（Plan）**：解决复杂的组合优化难题——评估谁在什么时间节点需要多少容量、何种加速器技术，每一次优化排期都涉及数百兆瓦算力的平稳落地。
- **设计阶段（Design）**：根据不同地域与建筑的异构特征（加速器形态、光纤跨度、特定网络 Fabric），利用智能体 AI 进行端到端架构的自动化推演与精细化设计。
- **交付阶段（Deliver）**：交付是“理论与现实碰撞的关口”。团队核心目标是绝不能让网络成为瓶颈，“最糟糕的局面莫过于大批 NVIDIA GB200、B300，或者 AMD、MTIA 芯片机架已经就位，却都在空转等待网络打通”。为此交付流程被极致压缩优化，甚至在紧邻施工工地的环境下即可并行完成生产上线。
- **运维阶段（Run）**：面对单集群 129,000 颗 GPU 及其配套的数十万张网卡（NICs）、海量内存、存储节点与高密光模块，传统人工排查难以为继，必须依靠智能体 AI 在错综复杂的多网络层与系统日志间实时关联异常、秒级锁定故障根因。
---
### [19:11](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=1151s) 全球算力版图、网络核心地位与“AI for Everyone”愿景
展望基础设施的下一步演进，算力上线正迈向多源协同与集群化部署的新范式。正如 Meta 基础设施团队此前展示的蓝图，为了以最快速度获取算力，Meta 正在多管齐下：自建超级数据中心、租赁成熟数据中心机房，并与头部公有云厂商建立深度算力合作。
在推进这些混合算力节点上线时，行业经验表明绝不能将各站点孤立开辟，而应将它们编织成相互协同的“环形集群”（Rings），共享底层关键服务以提速算力交付。这种跨越数百至数千公里的多中心环网，正是未来跨广域“Scale Across”的雏形。
```mermaid
flowchart TD
accTitle: Future Distributed Capacity Rings
accDescr: 自建数据中心、租赁机房与云厂商合作节点共同组成环形协同网络
subgraph CapacitySources["多维算力来源"]
Own["自建超大数据中心"]
Lease["外部租赁数据中心"]
Cloud["头部云服务商算力资源"]
end
subgraph RingTopology["协同服务环网 (Capacity Rings)"]
Own --- Lease
Lease --- Cloud
Cloud --- Own
end
RingTopology --> ScaleAcrossFuture["未来长距 Scale Across<br/>跨越数百至数千公里的大协同"]
```
虽然整个 AI 算力底座正处于陡峭的曲棍球棒（Hockey Stick）指数级攀升阶段，难以预测其最终峰值，但网络工程师们深知自己正处于技术变革的最核心位置：从机架内的芯片互连、建筑内的机架互连，到跨建筑、跨区域将海量设施熔铸为统一的超大集群，网络贯穿了现代 AI 系统的每一层血脉。
这一宏大基建的核心动力，最终来自于真实的应用场景与 Meta 连结世界的长远使命：
- **场景渗透**：从 Meta Ray-Ban 智能眼镜等前沿消费硬件，到加速变革制造业、医学诊疗、泛娱乐内容等实体工业，AI 正在重构千行百业。
- **普惠愿景**：全球有超过 35 亿人每天使用 Meta 旗下的应用矩阵（Facebook、Instagram、WhatsApp、Messenger 等）建立连接。Meta 坚信应当将连结人与人的哲学延续至 AI 时代——“让 AI 惠及每一个人，确保人人都能向善享有这一先进生产力”（AI for everyone and ensure everyone has that access for good）。正是为了支撑这 35 亿人平等拥抱 AI 的宏伟未来，构筑如此庞大规模的高性能网络基础设施才具备了终极意义。
### [22:16](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=1336s) 总结致谢与社区呼吁：共赴大规模网络挑战
Omar Baldonado 在演讲的尾声动情地谈到，正是这种直面未知的探索精神与技术突破，持续驱动着他自己以及整个 Meta 团队不断前行。在表达由衷谢意之余，他再次呼应了 Vignesh 此前提到的“社区精神”（Community），热切呼吁现场所有参会者充分利用大会的交流平台，主动与演讲嘉宾互动、彼此切磋探讨，共同正视并思考未来整个行业即将面对的各项技术挑战。
在他看来，超大规模网络（Networking @Scale）的前路依然充满未知与机遇，仅凭单一团队的力量无法彻底解决前方的重重难关，唯有依靠整个开发者与技术社区的紧密协同和开放交流。最后，他真诚地祝愿大家尽情享受接下来一整天充实丰富的议程，并在热烈的掌声中为本场主题演讲画上了句号。
- 核心呼吁："engage with the speakers, talk to each other, think about the problems that we have ahead of us."（积极与演讲者交流，彼此沟通，深入思考我们前方所面临的技术挑战。）
- 行业展望："There's a lot of great networking at scale ahead."（超大规模网络领域前方仍有无数精彩的挑战与机遇等待探索。）
## Notable quotes
> [00:00](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=0s) 网络与基础设施规模从“百万级”迈向“数十亿级”，供电能耗从“兆瓦级”（MW）飙升至“吉瓦级”（GW），而在算力层面更是迎来了天文数字般的飞跃。 — Omar Baldonado
> [04:35](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=275s) 这套知识体系是“累积式”（cumulative）的，绝对不存在“2022 年学会了分布式训练网络技术，到 2026 年就可以抛掉”的情况。 — Omar Baldonado
> [08:21](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=501s) 多快好省，任选其二（Good, fast, cheap — pick two）。 — Omar Baldonado
> [08:21](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=501s) 高性能（Performance）、高可靠（Reliability）、极速迭代交付（Fast cadence）以及极致灵活性（Flexibility）——四项我们全都要。 — Omar Baldonado
> [10:08](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=608s) 那些密布于机架内外、纵横交错的线缆才是整套系统最为壮观的工程灵魂所在（all the glorious networking）。 — Omar Baldonado
> [13:50](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=830s) 这完全是一场不可思议的现代土木工程壮举。 — Omar Baldonado
> [15:58](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=958s) 最糟糕的局面莫过于大批 NVIDIA GB200、B300，或者 AMD、MTIA 芯片机架已经就位，却都在空转等待网络打通。 — Omar Baldonado
> [19:11](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=1151s) 让 AI 惠及每一个人，确保人人都能向善享有这一先进生产力（AI for everyone and ensure everyone has that access for good）。 — Omar Baldonado
> [22:16](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=1336s) engage with the speakers, talk to each other, think about the problems that we have ahead of us. — Omar Baldonado
> [22:16](https://www.youtube.com/watch?v=fgMtBpgnQ0E&t=1336s) There's a lot of great networking at scale ahead. — Omar Baldonado
## Mentioned
Omar Baldonado, Meta, Gaia, Vignesh, Networking @Scale, Open Compute Project, TorchComms, FBOSS, PyTorch, America's Workforce Academy, NVIDIA H100, NVIDIA GB200, NVIDIA B300, AMD, Keysight, Google, OpenAI, CoreWeave, Oracle, Llama 3, Muse (Spark, Glimmer), DeepSeek, Open Claw, Agentic AI, MTIA (Meta Training and Inference Accelerator), Prometheus 1GW 集群, RDS (Rapid Deployment Structures / Tents), DSF (Disaggregated Scheduled Fabric), MetaURI, Meta Ray-Ban 智能眼镜, Facebook, Instagram, WhatsApp, Messenger
## Video description
Omar Baldonado leads the groups that develop/operate Meta's global data center networks. These networks support all of Meta’s AI models and the Meta family of apps (Meta AI, Facebook, Instagram, WhatsApp, Messenger).
These groups have developed some of the largest AI clusters in the world (with gigawatt-scale clusters on the way), and they continually share their work through open-source libraries (e.g., TorchComms for PyTorch, FBOSS for switches) and in communities like the Open Compute Project. Omar has been in networking since the early 1990s.
Learn more about the @Scale conferences here: https://atscaleconference.com/
[Source](https://www.youtube.com/watch?v=fgMtBpgnQ0E)
Source quotes
[0:00] Omar Baldonado: 网络与基础设施规模从“百万级”迈向“数十亿级”，供电能耗从“兆瓦级”（MW）飙升至“吉瓦级”（GW），而在算力层面更是迎来了天文数字般的飞跃。
[4:35] Omar Baldonado: 这套知识体系是“累积式”（cumulative）的，绝对不存在“2022 年学会了分布式训练网络技术，到 2026 年就可以抛掉”的情况。
[8:21] Omar Baldonado: 多快好省，任选其二（Good, fast, cheap — pick two）。
[8:21] Omar Baldonado: 高性能（Performance）、高可靠（Reliability）、极速迭代交付（Fast cadence）以及极致灵活性（Flexibility）——四项我们全都要。
[10:08] Omar Baldonado: 那些密布于机架内外、纵横交错的线缆才是整套系统最为壮观的工程灵魂所在（all the glorious networking）。
[13:50] Omar Baldonado: 这完全是一场不可思议的现代土木工程壮举。
[15:58] Omar Baldonado: 最糟糕的局面莫过于大批 NVIDIA GB200、B300，或者 AMD、MTIA 芯片机架已经就位，却都在空转等待网络打通。
[19:11] Omar Baldonado: 让 AI 惠及每一个人，确保人人都能向善享有这一先进生产力（AI for everyone and ensure everyone has that access for good）。
[22:16] Omar Baldonado: engage with the speakers, talk to each other, think about the problems that we have ahead of us.
[22:16] Omar Baldonado: There's a lot of great networking at scale ahead.
Source quotes
[2:49] Omar Baldonado: It's not like I can say, 'Well, I learned about large pre-scale training in 2022, I don't need to use that in 2026.' You do. All these workloads are running simultaneously in large gigawatt-scale clusters.
[21:06] Omar Baldonado: At Meta, we have three and a half billion people using some form of one of our applications to connect and we bring that same philosophy of connecting people to AI. We want AI to be for everyone and ensure everyone has that access for good.
Original transcript
[0:00] Here
[0:03] we are. Welcome to Networking at Scale.
[0:05] Thank you all for coming. This is one of
[0:08] my favorite days of the year. I love
[0:10] having everyone come together in the
[0:12] community, learning from the great
[0:15] groups at Meta Networking as well as a
[0:17] bunch of other companies that are
[0:19] operating networks at scale. My name is
[0:22] Omar Baldonado, part of the data center
[0:23] and AI networking group. We started
[0:26] networking at scale many years ago
[0:28] because it was our way, as Mcnesh said,
[0:30] to make sure that we're sharing with the
[0:32] community, make sure that we're sharing
[0:33] the lessons of running large-scale
[0:36] networks.
[0:37] So last year, Gaia presented a chart
[0:42] that showed the capex spend of some of
[0:44] the largest hyperscalers.
[0:46] And I was going to try to update that.
[0:48] I've been trying to update that over the
[0:49] months, and I find that that usually
[0:51] falls out of date within weeks. It's
[0:54] really hard to keep track of the
[0:55] buildout of AI infrastructure. So I
[0:58] thought I'd just talk about the the
[1:01] orders of magnitude that we have and
[1:03] then talk a little bit about our own
[1:06] scale. I think many of you here who work
[1:09] on atscale networks and atscale
[1:11] infrastructure, you see this that we've
[1:13] gone from millions to billions. We've
[1:16] gone from megawws to gigawatts. But this
[1:19] last one I wanted to talk about was the
[1:21] flops, right?
[1:23] End of last year, we had Meta had the
[1:27] equivalent of 1.2 million Nvidia H100s.
[1:32] And if you think about that, an H100
[1:34] capable of somewhere around a a pedlop
[1:36] or so depending upon the encoding. So a
[1:39] pedlop times 1.2 million and we're
[1:42] clearly in the zoflop scale, right? This
[1:44] is unbelievable the size of the
[1:46] infrastructure that that many of the
[1:48] hyperscalers including men are driving.
[1:51] So those numbers then translate into
[1:55] concrete programs and initiatives. Here
[1:58] are two that we've shared recently. One
[2:00] is around the compute partnerships that
[2:02] we've had with many of the CPU
[2:05] providers. In fact, even building custom
[2:07] CPUs for our infrastructure.
[2:10] Then the other one is around people. We
[2:13] build all this infrastructure. But I've
[2:14] gone and recently with some of the team
[2:16] to go visit different sites, different
[2:18] warehouses, different manufacturing
[2:20] centers. It's unbelievable how many
[2:24] people it takes to to build this
[2:25] infrastructure. So very proud that Meta
[2:28] announced America's Workforce Academy to
[2:30] make sure that we have people with the
[2:33] skills and expertise to work in these
[2:36] sites, to work in our manufacturing
[2:38] centers and warehouses to deliver this
[2:40] AI infrastructure.
[2:42] All right, so that's a little bit about
[2:44] the numbers and the scale that we were
[2:46] were talking about. Let's talk
[2:47] networking.
[2:49] I've shared this before. This is the
[2:51] size of Meta's largest training cluster
[2:53] over the years. And it was actually
[2:55] pretty fun from a walk down history lane
[2:57] perspective. Okay, so raise your hand
[3:00] here. Who was here? Can kind of see you
[3:02] all. Who was here at networking at scale
[3:04] 2023? You know, raise your hand up high,
[3:07] right? This was this was right after co
[3:10] this was our first one back in person.
[3:13] It was fun. But you can see already in
[3:16] 2023. Thank you.
[3:18] We were talking about Ethernet networks
[3:21] for AI training and at the time 22 2023
[3:25] we were talking about 8K 16k size GPU
[3:28] clusters.
[3:30] 2024
[3:33] you can then see same thing but now
[3:34] bigger. We talked about training for
[3:37] Llama 3 on two 24K clusters, right?
[3:41] Along with a number of other advances
[3:44] within AI. That was our second one that
[3:46] actually the one in 2023 was called
[3:49] networking at scale AI networking
[3:51] edition, right? Seems kind of quaint now
[3:53] because now after many years that is all
[3:56] about, you know, AI driving us. So
[3:59] that that's really been the theme across
[4:01] these last four years. Now
[4:04] last year right last year it really
[4:08] started to broaden out. You saw now
[4:10] we're talking about 129,000 GPU cluster
[4:14] that we had built in 2024 and how we
[4:16] train generative AI on that. But then we
[4:18] were also hinting at future the
[4:21] networking technologies that would be
[4:23] used for the gigawatt scale cluster the
[4:25] fabrics in the gigawatt scale and even
[4:28] the backbone and the scale across that
[4:30] would be used for the gigawatt scale
[4:32] cluster. We talked about that last year.
[4:35] Now what's clear from this walk down
[4:37] history lane is what drives the AI
[4:39] infrastructure, what drives the network
[4:41] infrastructure is the AI workloads. you
[4:44] could see the talks year over year and
[4:47] that's really been that's actually been
[4:49] very motivating I think for myself and
[4:51] for many of you in the networking
[4:53] community over the last few years we
[4:56] have had to learn about all the
[4:58] different sorts of AI jobs and AI
[5:00] workloads. So look back to 2022 right we
[5:04] were talking we were learning about what
[5:06] was the cluster the cluster was mostly
[5:08] for running training and we were
[5:10] learning about parallelisms tensor
[5:13] parallel pipeline parallel data parallel
[5:16] then we started looking at okay training
[5:18] runs on large clusters but now inference
[5:21] can't fit on one machine so in 23 24 we
[5:25] started talking about disagregated
[5:27] inference prefill and decode we also
[5:29] started talking about training for
[5:31] different types of use cases, large
[5:34] language models versus ads.
[5:37] Last year we talked about and then the
[5:38] explosion of use cases, reinforcement
[5:41] learning, mixture of experts. And as
[5:45] networking enthusiasts, again I say
[5:46] networking, we have to know these
[5:50] applications. We have to know the
[5:51] traffic patterns. And importantly, it's
[5:54] cumulative.
[5:55] It's not like I can say, well, I learned
[5:57] about large pre-scale training in 2022.
[6:00] I don't need to use that in 2026. You
[6:02] do. All these workloads are running
[6:05] simultaneously in large gigawatt scale
[6:08] clusters. So you've this is the good
[6:10] news for all of us in networking. All of
[6:12] this you've been learning year over
[6:14] year. You've got to keep on bringing it
[6:16] bring all that knowledge into the future
[6:19] clusters that we build.
[6:21] As an example, just earlier this month,
[6:25] Meta Ads shared some of the wins that
[6:27] we've had working with the networking
[6:30] team to apply large language model
[6:33] learnings and training, adapting that,
[6:36] customizing it toward our ads model
[6:39] training. The use cases keep expanding,
[6:41] the type of workloads keep expanding.
[6:44] Now, 2026, it's been another busy year
[6:47] for all of us. Particularly at Meta, we
[6:50] announced our Muse sets of models,
[6:52] Spark, Glimmer, and Open Model runs on
[6:55] one GPU, image, and video. But 2026 has
[6:59] also seen yet another change in the
[7:02] workloads that we've had to add on to
[7:04] our cumulative set of experiences.
[7:06] Just like 2025 had its early 25 deepseek
[7:11] moment, early 2026 saw Open Claw and the
[7:15] rise of agentic AI.
[7:18] Now
[7:20] that brought another set of traffic
[7:24] patterns and network demands that we had
[7:28] to kind of understand again as
[7:30] networking experts that needed to
[7:32] understand the workloads. So agentic AI
[7:36] is dependent upon long-term inference.
[7:39] So now we're having to think about KV
[7:42] cache and storage and communication
[7:44] between CPUs and GPUs all happening at
[7:48] once in the clusters even as all the
[7:50] other workloads are happening training
[7:52] and disagregated inference and all. So
[7:55] at this point I would say networking
[7:57] you've it's pretty much all traffic
[7:59] patterns. It's no longer just the the
[8:02] highly intense training that we started
[8:05] in 2022.
[8:07] It's it's all of it. It it really feels
[8:09] like you have to bring all the levers,
[8:12] all the tools in your toolbox that
[8:14] you've looked at applications over the
[8:16] years and bring them every day into the
[8:19] into the AI cluster.
[8:22] Now, when I talk to those AI researchers
[8:24] and model developers,
[8:27] this is what I I hear, right? And I'm
[8:29] sure many of you talk to to your model
[8:31] developers or your customers about what
[8:33] are their workloads. This reminds me of
[8:36] the the old adage you know good fast
[8:38] cheap pick two know that now the AI
[8:42] researchers say this we need performance
[8:45] reliability fast cadence and flexibility
[8:47] all four so there's no sort of like hey
[8:50] just pick pick two I can only you know
[8:51] they want all of it right that's what
[8:53] drives our infrastructure builds
[8:57] so
[8:59] keeping in theme of the conference this
[9:01] is this is how we're thinking about the
[9:04] problem of how do we achieve all these
[9:06] requirements
[9:08] there's both network infrastructure for
[9:10] AI NI for NI and or NA for AI and that's
[9:15] a full stack codeesign problem
[9:18] then we'll also talk about life cycle
[9:21] where we are actually using the AI
[9:23] ourselves in running our large scale
[9:26] infrastructure and that's the AI for NI
[9:30] let me just go quickly touch on a few
[9:32] areas of this full stack problem that we
[9:34] have to deal with. This is familiar to
[9:37] many of you in networking, right?
[9:39] Probably you you are working on at least
[9:41] one of these if not multiple scale up,
[9:45] scale out, scale across.
[9:47] This is not even the complete picture of
[9:49] the networking stack that we have
[9:51] because we have the chips inside the the
[9:54] racks, the AI accelerators. There's a
[9:56] lot of software that runs on top that
[9:58] orchestrates all this. So let's delve in
[10:01] to see in each of these domains how are
[10:04] we meeting all those cluster
[10:06] requirements.
[10:08] So first one let's go into let's take
[10:11] MTIA. MTIA we announced this meta
[10:14] training and inference accelerator. We
[10:16] announced earlier this year we're
[10:19] working on four chips in two years. If
[10:23] that's not an example of fast cadence I
[10:25] don't know what is. That is an example
[10:26] of getting out the silicon that our AI
[10:30] researchers need as fast as possible. Of
[10:33] course, it's high performance and
[10:34] reliable. It's why we we're building it.
[10:36] But then also flexibility, right?
[10:39] We initially designed the first
[10:41] generation of these chips for training,
[10:43] but you'll see in in a paper that we
[10:46] just released in the blog post, you'll
[10:48] also see in talks from Wes and Rajie
[10:49] later today that we've had to adapt that
[10:52] even as we targeted training, we needed
[10:55] to target and see how it can also be
[10:57] used for inference. So flexibility,
[11:01] these chips reside in in a scaleup AI
[11:04] system rack. And so this is actually one
[11:06] of our MTIA systems. You can see all the
[11:09] glorious networking that's inside that
[11:11] MTI system. I've seen pictures of
[11:13] different AI systems. Sometimes they
[11:15] don't put all the cables in. I find that
[11:17] personally like you have to show the
[11:19] cables. That's the whole point of of
[11:21] network. It's tough. It's tough to make
[11:24] these reliable. It's tough to get these
[11:26] to be performant. But it's what
[11:29] the researchers demand.
[11:32] All right, let's now look at scale out.
[11:34] Scale out is the building level network
[11:39] and it's very tied to the building and
[11:40] there's a lot of innovations in the full
[11:42] stack here as well. Let's look first at
[11:45] the building level. This is a picture of
[11:49] some of our Prometheus 1 gawatt cluster
[11:52] in Ohio. These are the rapid deployment
[11:55] structures which you may have heard
[11:57] about in formerly known as tents. And
[12:00] this is again an example of how we've
[12:01] had to be flexible and have a rapid
[12:04] cadence. We used to build buildings in
[12:06] the years time frames and we still do
[12:10] but when we need to bring capacity fast
[12:12] we're building these buildings in the
[12:14] months time frames. That's in response
[12:17] to get all the capacity again that we
[12:18] need. Happy to share that these rapid
[12:21] deployment structures are we have many
[12:23] of these in production. They're all lit
[12:24] up and running. They're running
[12:26] supporting Muse Spark and and many other
[12:28] workloads in our fleet. So that's at the
[12:31] building layer. Let's go inside the
[12:33] building. The fabrics we've been talking
[12:35] about different fabrics over many many
[12:38] years. This one shows our disagregated
[12:40] scheduled fabric as well as our
[12:42] non-scheduled fabric. And today actually
[12:44] Google is going to be talking about a
[12:46] new fabric that they've been developing
[12:47] for their AI backends
[12:50] building fabric transport.
[12:53] very excited to talk about transport.
[12:56] There's been a lot of innovation here
[12:58] and yesterday we released a blog around
[13:02] Metari and our transport that we are
[13:05] that we've developed specifically for
[13:07] performance at scale, reliability at
[13:09] scale and operations at scale. Again,
[13:11] this is a very hot area. We're excited
[13:14] to share this and years of work that the
[13:16] teams have been doing. people from Meta
[13:19] as well as AMD and Ksite will be sharing
[13:22] lessons around this.
[13:24] It's so it's so active right that right
[13:26] after this you'll hear from open AI
[13:28] they'll talk about some of their
[13:30] approaches to transport and then we'll
[13:32] also have a panel right after lunch
[13:35] where we'll have Coreweave
[13:38] Open AAI Oracle and Meta all talking
[13:40] about scale out networking and transport
[13:44] for AI hot area you're not going to want
[13:47] to miss that
[13:50] okay let's let's look at scale across
[13:53] this is where we move beyond the
[13:54] building beyond the region and connect
[13:57] up connect up large clusters to each
[14:00] other. Now when we think about scale
[14:04] across the first thing that most people
[14:06] think about is distance
[14:08] how many kilometers right this is again
[14:11] part of our Prometheus cluster this has
[14:12] several hundred megawatts pictured it's
[14:14] not the full gigawatt cluster but even
[14:17] within this picture there are three
[14:19] separate regions pictured they're of
[14:21] differing distances but they're also
[14:23] very heterogeneous in nature
[14:26] the buildings in the top the H shapes
[14:29] those are some of our air cooled data
[14:31] centers. Those have the DSF fabric
[14:34] inside of them.
[14:36] But then the buildings at the bottom,
[14:37] slightly newer, are our liquid cooled
[14:40] data centers
[14:42] holding just about as many GPUs as the
[14:44] top three H's, but much more dense. And
[14:48] then in the top left, you see those
[14:50] structures are the rapid deployment
[14:52] structures or tents, right? That we that
[14:54] we quickly brought up. I showed those in
[14:55] the earlier pictures. So all of this
[14:57] several hundred megawatts brought
[14:58] together in a region. Of course there
[15:00] are varying distances but look at the
[15:03] technologies that we have to use the the
[15:05] density the the optics transceivers the
[15:08] there's a lot of heterogeneity in here
[15:10] that has to come together for a scale
[15:11] across network. I'm not going to go over
[15:14] all the technologies that are needed for
[15:16] scale across. We've covered some of that
[15:17] before and in fact well Metari is a
[15:20] great example of handling long distance
[15:21] heterogeneous. It's well suited for that
[15:24] later today. What you're going to hear
[15:25] about though is about the physical
[15:27] infrastructure, more about the fiber,
[15:30] the fiber design and planning and
[15:32] actually the civil engineering that it
[15:34] takes to build this. We have a talk on
[15:38] how do you deal with frozen ground and
[15:40] permits and working around property law.
[15:42] It's incredible what a civil engineering
[15:44] job this is to build effectively what is
[15:48] an an high density high bandwidth AI
[15:52] backbone right within this region. So
[15:56] more on that.
[15:58] Okay. So now let's move on to the the
[16:01] life cycle part or the all of that was
[16:04] about the network infrastructure we're
[16:06] building for AI. This is now how we're
[16:08] leveraging AI for the network life
[16:11] cycle. This is an area that I think is
[16:14] not I always say it's not paid attention
[16:16] to enough but this is a huge lesson from
[16:19] gigawatt scale clusters. You have to
[16:22] think about your life cycle because it
[16:25] is a tremendous it's a tremendous asset
[16:27] that we have a gigawatt scale fleet and
[16:30] how do you carefully manage that through
[16:32] all of its life cycle. We are leveraging
[16:34] AI and in fact a lot of agentic AI to be
[16:37] doing this. Many of you who work at at
[16:41] hypers scale networks you will have a
[16:44] similar sort of life cycle. Maybe the
[16:46] names of the boxes or the number of
[16:47] boxes may be different but that whole
[16:51] life cycle represents
[16:54] different phases of how we need to
[16:56] either plan, design, deliver or run the
[16:58] network. If I look into planning a
[17:01] little bit, there's an optimization
[17:03] problem. Who needs the capacity with
[17:05] what technology by what time? And this
[17:08] is really now at the we're we're we're
[17:11] planning out and optimizing hundreds of
[17:13] megawatts at a time that need to land.
[17:16] Now, once we've decided what we need to
[17:18] build, what exactly is going to get
[17:21] instantiated in any given region? So,
[17:24] you saw the heterogeneity in that
[17:25] picture. You need to take into account
[17:28] what type of accelerator, what are the
[17:30] specific distances, what's the fabric
[17:32] that we're going to build. All of that
[17:33] is again being used and being optimized
[17:36] with agentic AI delivery. This is where
[17:40] rubber meets the road. This is where all
[17:41] that planning and design now it's going
[17:44] to land in a data center and we need to
[17:46] turn that up as quickly as possible. The
[17:48] delivery speeds are something we've been
[17:51] spending a lot of time optimizing
[17:53] because
[17:54] you don't want to block. The network
[17:56] never wants to block AI racks of GPUs
[18:00] that are coming up. It's the worst thing
[18:01] you want to have is like I've got all
[18:03] these Nvidia GP200s, 300's or AMDs or
[18:07] MTS waiting for the network. No one
[18:09] wants that. So the delivery is is a
[18:13] place where we're optimizing bringing
[18:15] stuff into production even as we're
[18:18] doing construction right next to it.
[18:20] Right? So all of that has to happen in
[18:22] in lock step to turn this capacity up.
[18:26] Finally run or operations. This is
[18:29] something that's probably the most
[18:30] obvious use of agentic AI. Let's just
[18:33] take the example 129,000 GPUs goes along
[18:36] with so many hundreds of thousands of
[18:39] nicks and memory and storage and optics.
[18:42] We are using agentic AI to help make
[18:45] sense of all these different layers to
[18:47] correlate and find root cause and that's
[18:50] the only way you can scale handling
[18:52] this.
[18:55] So again that's that's kind of where we
[18:58] are. You're going to hear a lot of talks
[19:00] around all of this AI for NI. We have a
[19:04] whole track for that as well as the the
[19:06] different parts of the codeesign stack
[19:08] that we're optimizing.
[19:11] So, let's step back. Let's step back
[19:12] with where we started. What's coming
[19:14] next?
[19:16] I'll I'll tease this a little bit,
[19:18] right? But let's start with those
[19:20] concentric those those circles that were
[19:22] getting bigger and bigger. Earlier this
[19:24] summer, my colleague in infrastructure
[19:26] at Meta, Serupa, shared this picture
[19:29] where Meta, we're just trying to bring
[19:31] as much capacity online as quickly as
[19:33] possible. We're building our own data
[19:35] centers. We're leasing data centers and
[19:38] we're actually partnering with cloud
[19:39] providers to get all of this capacity.
[19:42] Now, all these locations, how do you
[19:45] turn up that capacity as quickly as
[19:47] possible? And what we learned was that
[19:50] you don't want to bring up these sites
[19:51] independently. You want to sort of
[19:54] cluster them together into these rings
[19:56] where we can say, hey, they can leverage
[19:58] each other for services and turnup to
[20:01] bring up capacity faster. Now, this is
[20:03] from a a capacity turnup perspective,
[20:05] but it's not hard to see, you know, is
[20:07] that the scale across of the future that
[20:09] we're targeting, right? That is hundreds
[20:11] of kilometers, thousands of kilometers
[20:13] in distance. I can see that coming. So,
[20:16] so that's
[20:18] that's in the in the in the horizon.
[20:23] the numbers
[20:25] we are clearly at at at this hockey
[20:28] stick. You know where exactly we are in
[20:30] the AI infrastructure build. Hard to
[20:32] say. So I don't not going to put a graph
[20:33] up there again. But this is what drives
[20:36] all of us. This is what has made
[20:38] networking and infrastructure build so
[20:40] exciting over the last couple of years.
[20:42] We are in the middle of all this as
[20:44] networking. We network the chips in a
[20:47] rack.
[20:49] We network the racks in a building and
[20:51] we network the buildings together into
[20:55] one large cluster if needed. So network
[20:58] is at the is is really at the middle of
[21:00] all of this infrastructure build. It's
[21:03] exciting exciting.
[21:06] And now stepping all the way back what
[21:07] motivates this infrastructure it's for
[21:09] the use cases. It's for the consumerf
[21:12] facing use cases like the meta rayband
[21:14] glasses. But so many industries are
[21:16] being transformed by AI. Whether it's
[21:18] manufacturing, medicine, entertainment,
[21:23] AI is has disrupted or is has the
[21:25] potential to change and evolve so many
[21:28] of our industries.
[21:31] Let me close here with something I find
[21:33] personally very motivating and it it
[21:35] comes down to our mission. At Meadow, we
[21:38] have three and a half billion people
[21:40] using some form of one of our
[21:42] applications to connect and we bring
[21:44] that same philosophy of connecting
[21:46] people to AI. We want AI to be for
[21:50] everyone and ensure everyone has that
[21:52] access
[21:54] for good.
[21:56] I think if you think then about 300 or
[21:59] 3.5 billion people
[22:02] getting access to AI that is a clear
[22:05] motivator of the size of infrastructure
[22:07] that we need to build. It's a clear
[22:09] driver of the importance of the network
[22:12] that we are providing for these
[22:14] applications and it's what it's what
[22:16] really drives me and a lot of the team
[22:18] today.
[22:19] So thank you very much. Please, please,
[22:23] again, Vignesh talked about the
[22:25] community,
[22:27] engage with the speakers,
[22:29] talk to each other, think about the
[22:32] problems that we have ahead of us.
[22:34] There's a lot of great networking at
[22:35] scale ahead. So, thank you very much and
[22:37] enjoy the rest of the day.
