每機架PB級：AI流量正在重塑網路—OFC 2026揭示網路架構、測量差距和供應鏈風險。

Petabits Per Rack - How AI Traffic is Reshaping Networks: What OFC 2026 Reveals about Network Architecture, Measurement Gaps, and Supply Chain Risk

出版日期: 2026年04月27日 | 出版商:

MTN Consulting, LLC | 英文 31 Pages | 訂單完成後即時交付

價格

簡介目錄圖表

人工智慧流量與現有網路傳統上依賴的雲端和企業流量截然不同。訓練叢集以每機架PB級頻寬運作，採用零丟包協議，並實現微秒同步，這些都對基礎設施提出了新的挑戰。光是今年，超大規模資料中心業者就已投資超過6,000億美元來應對這些挑戰。

本報告基於 Meta、KDDI、中國行動、三星和台積電等公司在 OFC 2026 上發表的 30 篇論文，從技術角度揭示了人工智慧網路建設的現狀、架構的未來以及供應鏈中風險最高的領域。

主要結論：

人工智慧流量與傳統流量有著本質差異。訓練叢集對每個機架的頻寬要求高達PB級，採用零丟包協議，並具有微秒級的同步限制，這與現有網路旨在傳輸的機率性、容錯性流量幾乎沒有通用。
從某種意義上說，人工智慧產業就像「蒙著眼睛」摸索前進。目前還沒有關於人工智慧流量的規模、模式或成長的全面公開研究。雖然諾基亞和愛立信等公司已經進行了一些初步分析，但超大規模資料超大規模資料中心業者並未共用流量資料。這對於今年資本支出超過6,000億美元的人工智慧產業來說，無疑是一個巨大的規劃挑戰。
共封裝光元件 (CPO) 的可靠性已突破閾值。 Meta 公司對 CPO 交換器進行了 3,600 萬裝置小時的現場評估，有力地證明了該市場的可行性。在規模化擴充層中替換支援重定時功能的插件模組，不再是「是否會發生」的問題，而是「何時發生」的問題。
「跨域擴展」是下一個前沿領域。由於電力限制，GPU叢集不得不分佈在多個設施和園區內。 KDDI已證實，即使分散式訓練跨越30公里，其AI處理完成時間也能與單一站點叢集相當。微軟已宣布部署15,000公里中空光纖用於AI連線。
光纖通訊價值鏈正在經歷一場變革。隨著超大規模資料中心業者利用台積電和三星的平台實現矽光電設計的內部化，互連晶片的價值正從分離模組供應商轉移到晶圓代工廠、晶片供應商和超大規模資料中心業者。那些沒有晶圓代工廠製程設計套件（PDK）策略的供應商，將在其總產量達到頂峰之際，面臨市場萎縮的困境。

提及的組織

1FINITY Inc.
Alcatel Submarine Networks
Alibaba Cloud
Alphabet (Google Cloud Platform)
Amazon (Amazon Web Services, AWS)
AMD
Ampere
Anthropic (Claude)
ARM
AttoTude Inc.
深圳博塞爾光電科技有限公司（中國深圳）
Broadcom
ByteDance
加泰隆尼亞電信技術中心 (CTTC-CERCA)（西班牙）
中國移動研究院（北京，中國）
Chinese University of Hong Kong
Ciena
Cornell University
Corning Inc.
DeepSeek
Ericsson
Flexcompute Inc.
Furukawa Electric Co., Ltd.
華中科技大學（中國武漢）
Hubei Optical Fundamental Research Center
II-VI/Coherent
Innolight
Intel
iPronics
Jinyinhu Laboratory
KDDI Research, Inc.
Lumentum
Lumiphase AG
McGill University
Meta Platforms
Microsoft (Azure)
Mistral
名古屋大學（日本名古屋）
日本產業技術綜合研究所（AIST）
Nokia Bell Labs
Nokia Corporation
NVIDIA
NYSERNet
OpenAI (ChatGPT)
光子電子技術研究協會（PETRA）（日本東京）
光子學-電子學整合研究中心（日本筑波）
都靈理工大學（義大利都靈）
銳捷網路（中國福州）
Samsung Electronics Co., Ltd.
北京郵電大學資訊光子學與光通訊國家重點實驗室（中國北京）
北京大學光子學與通訊國家重點實驗室（中國北京）
台灣積體電路製造股份有限公司（TSMC）（台灣新竹）
Toyota Technological Institute
Tsinghua-Berkeley Shenzhen Institute
University of California, Santa Barbara
Wuhan Changjin Photonics Technology Co.
Wuhan Research Institute of Posts and Telecommunications
延世大學（韓國）

概括
人工智慧流量基礎知識
- 在超大規模投資激增的背景下，了解人工智慧流量至關重要。
- 人工智慧流量類型：概述
- 衡量和預測人工智慧流量面臨的挑戰
流量方向與分析結果：向上擴展/向外擴展/橫向擴展
- 擴展：叢集內部流量
- 橫向擴展：資料中心架構內的流量
- 跨資料中心規模化：資料中心之間的人工智慧流量
對超大規模市場的影響
- 擴大規模
- 橫向擴展
- 規模跨越
- Transocean
對進入市場的企業的建議
- 適用於光元件和收發器供應商
- 對於連貫通訊領域的供應商而言
- 適用於 InfiniBand 和乙太網路交換器/網卡供應商
- 對於部署人工智慧基礎設施的資料中心營運商而言
- 對於計劃建構分散式培訓基礎設施的企業而言
- 適用於在區域和城市網路中處理人工智慧流量的通訊業者。
- 對於擁有海底電纜資產的通訊業者
結論
附錄 1：OFC 試卷詳情
附錄 2：報告與出版商訊息

簡介目錄圖表

Product Code: DCPC-27042026-1

AI traffic is vastly different than the cloud and enterprise traffic that existing networks were built to carry. Training clusters running at petabit-per-rack bandwidth, zero-loss protocols, and microsecond synchronization are a new infrastructure challenge. Hyperscalers are spending over $600 billion in capex this year addressing the challenge. This report draws on 30 papers from OFC 2026, from Meta, KDDI, China Mobile, Samsung, TSMC, and others, to establish what the engineering record shows about how AI networks are built today, where the architecture is heading, and where the supply chain risk is highest.

Our key conclusions are as follows:

AI traffic is a different animal. Training clusters running at Petabit-per-rack bandwidth requirements, zero-loss protocols, and microsecond synchronization constraints have little in common with the stochastic, jitter-tolerant traffic that existing networks were built to carry.
The industry is flying partially blind. No comprehensive public study of AI traffic volumes, patterns, or growth exists. Nokia, Ericsson, and a handful of others have made partial contributions, but hyperscalers don’t share traffic data. For an industry spending over $600B in capex this year, this is a significant planning liability.
Co-packaged optics has crossed the reliability threshold. Meta’s 36 million device-hour field evaluation of CPO switches provides strong support for this market’s viability. The displacement of retimed pluggable modules at the scale-up tier is no longer a question of if, but when.
Scale-across is the next frontier. Power constraints are forcing GPU clusters across multiple facilities and campuses. KDDI confirms that distributed training across 30 km produces AI job completion times on par with single-site clusters. Microsoft has announced 15,000 km of hollow-core fiber deployment for AI connectivity.
The optical supply chain faces a shift. As hyperscalers bring silicon photonics design in-house using TSMC and Samsung platforms, the value of interconnects is moving from discrete module vendors toward foundries, chiplet suppliers, and hyperscalers. Vendors without a foundry process design kit (PDK) strategy face a narrowing addressable market at precisely the moment overall volumes are peaking.

Organizations mentioned

1FINITY Inc.
Alcatel Submarine Networks
Alibaba Cloud
Alphabet (Google Cloud Platform)
Amazon (Amazon Web Services, AWS)
AMD
Ampere
Anthropic (Claude)
ARM
AttoTude Inc.
Berxel Photonics Co. Ltd. (Shenzhen, China)
Broadcom
ByteDance
Centre Tecnologic de Telecomunicacions de Catalunya (CTTC-CERCA) (Spain)
China Mobile Research Institute (Beijing, China)
Chinese University of Hong Kong
Ciena
Cornell University
Corning Inc.
DeepSeek
Ericsson
Flexcompute Inc.
Furukawa Electric Co., Ltd.
Huazhong Univ. of Science and Technology (Wuhan, China)
Hubei Optical Fundamental Research Center
II-VI/Coherent
Innolight
Intel
iPronics
Jinyinhu Laboratory
KDDI Research, Inc.
Lumentum
Lumiphase AG
McGill University
Meta Platforms
Microsoft (Azure)
Mistral
Nagoya University (Nagoya, Japan)
National Institute of Advanced Industrial Science and Technology (AIST) (Japan)
Nokia Bell Labs
Nokia Corporation
NVIDIA
NYSERNet
OpenAI (ChatGPT)
Photonics Electronics Technology Research Association (PETRA) (Tokyo, Japan)
Photonics-Electronics Integration Research Center (Tsukuba, Japan)
Politecnico di Torino (Turin, Italy)
Ruijie (Fuzhou, China)
Samsung Electronics Co., Ltd.
State Key Lab of Information Photonics and Optical Communications, BUPT (Beijing, China)
State Key Laboratory of Photonics and Communications, Peking University (Beijing, China)
Taiwan Semiconductor Manufacturing Company (TSMC) (Hsinchu, Taiwan)
Toyota Technological Institute
Tsinghua-Berkeley Shenzhen Institute
University of California, Santa Barbara
Wuhan Changjin Photonics Technology Co.
Wuhan Research Institute of Posts and Telecommunications
Yonsei University (South Korea)

Summary
AI Traffic 101
- Understanding AI traffic is high-stakes as hyperscale capex explodes
- AI traffic types: A primer
- The challenge of measuring and forecasting AI traffic
Traffic directions & findings for scale up, out and across
- Scale-up: Traffic within the cluster
- Scale-out and the data center fabric
- Scale-across: Inter-datacenter AI traffic
Implications for the hyperscale market
- Scale up
- Scale out
- Scale across
- Transoceanic
Recommendations for industry players
- For optical component and transceiver vendors
- For vendors in the coherent space
- For InfiniBand and Ethernet switch and NIC vendors
- For data center operators deploying AI infrastructure
- For operators planning distributed training infrastructure
- For operators carrying AI traffic in regional and metro networks
- For telcos with subsea cable assets
Conclusion
- Using the appendix
Appendix 1: OFC paper details
- Paper metadata
- Paper summaries
Appendix 2: Report and publisher information

簡介目錄圖表

List of Figures and Tables

Figure 1: Hyperscale capex history and near-term outlook ($B)
Figure 2: Training, inference, and agentic AI traffic patterns are very different
Figure 3: AI training cluster size progression since 2016
Figure 4: Scale up, out, and across bandwidth requirements all growing fast
Figure 5: Three versions of AI cluster GPU racks at Meta