 |
市場調查報告書
商品編碼
2136219
華為：在OFC和ECOC上展示的AI互連研究成果超越了英偉達。Huawei Leads Nvidia in Published AI Interconnect Research at OFC and ECOC |
||||||
報告摘要
在2025年COC和2026年OFC大會上發表的95篇人工智慧互連論文中，華為是合作論文最多的公司，共發表了7篇。相較之下，英偉達只有4篇。這種差異主要集中在一個層面：跨域規模（scale-across）。跨域規模指的是連接資料中心而非機架內部的光鏈路。儘管英偉達在2025年8月22日宣布Spectrum-XGS乙太網路是CoreWeave技術的早期採用者，並將跨域規模定位為其產品線的核心支柱，但它在這兩屆大會上都沒有發表任何關於該主題的論文。根據本研究涵蓋的論文，所有支援跨域規模的光纖通訊研究都由華為、歐洲研究機構和中國電信業者進行。
華為的論文涵蓋了本研究追蹤的人工智慧互連五個子主題中的四個。而英偉達的論文則涵蓋了三個子主題，這些子主題都與資料中心的內部運作有關。中國雲端和人工智慧平台研究人員發布了實際運行數據，包括實際運行系統的測試結果以及阿里巴巴針對102.4 Tb/s交換機（W2A.43）設定的插入損耗預算和誤碼率（BER）目標值。相較之下，用於比較的六篇美國論文報告的是實驗室評估和演示結果。
英偉達本身列出的光纖通訊生態系統公司在研究論文的影響力有限。在英偉達2025年3月宣布的11家矽光電夥伴中，只有康寧和住友電工出現在95篇論文中。另一方面，在所有受調查的雷射供應商中發表論文數量最多（5篇）的古河電工，卻並未出現在合作夥伴名單中。
中國在人工智慧互聯研究領域的比例超過了其在各學術會議整體議程中所佔的比例。該指數在2025年歐洲電腦大會（ECOC 2025）和2026年光纖通訊大會（OFC 2026）分別為1.23和1.05，均高於歐洲的0.70和0.42。
本報告涵蓋了2020年至2026年ECOC和OFC的總體趨勢、五個相互關聯的層面，並按國家和組織進行了細分。報告還包含一個附錄，依學術團體、組織、國家和次主題整理了全部95篇論文。
目錄
- 概括
- 聚合目標
- 在過去六年裡，人工智慧互聯相關論文在會議論文總數中所佔的比例從 4% 增加到 9%。
- 華為的論文涵蓋了四個互連層，而英偉達的論文涵蓋了三個互連層。
- 超大規模資料中心業者發表的研究和超大規模企業未發表的研究
- 英偉達公開宣布的光纖通訊合作夥伴在研究論文中出現的頻率有限。
- 中國在人工智慧互聯研究領域處於較高水平，而歐洲則處於較低水平。
- 本分析的局限性
- 影響
- 結論
- 附錄1：調查方法、定義與分析限制
- 附錄 2：本分析所包含的論文
- 附錄 3：報告與出版商訊息
調查對象
提及的組織：
- AIO Core
- AIST
- Alibaba
- Alibaba Cloud
- Ambition Photonics
- AttoTude
- AuthenX
- Ayar Labs
- Baidu
- Beijing University of Posts and Telecommunications
- Berxel
- Browave
- BT
- CAICT
- CEA-Leti
- China Mobile
- China Telecom
- China Unicom
- Ciena
- Coherent
- CoreWeave
- Corning
- CSA Catapult
- Deutsche Telekom
- Dublin City University
- Eindhoven University of Technology
- EPFL
- Ericsson
- Fabrinet
- FormFactor
- Foxconn
- Fraunhofer
- Fudan University
- Fujifilm
- Furukawa Electric
- GeNopsys
- GlobalFoundries
- Huawei
- Huazhong University of Science and Technology
- ICON Photonics
- imec
- Innolume
- Institute of Computing Technology, CAS
- Institute of Science Tokyo
- Kagawa University
- KDDI
- Keio University
- KPN
- Ligent
- Linfiber Technology
- Lumentum
- Max Planck Institute
- Meta
- Mitsubishi Electric
- Nanjing University
- National University of Defense Technology
- National University of Singapore
- Nokia
- NTT
- Nvidia
- Orange
- Peking University
- Politecnico di Milano
- QoreTek
- RCALS Centre for Advanced Lightwave Science
- Ruijie
- Samsung
- SENKO
- Shanghai Jiao Tong University
- Solinide Photonics
- SPIL
- Sumitomo Electric
- Suzhou Everbright Photonics
- Telefonica
- TFC Communication
- TIM
- TSMC
- TSRI
- Tyndall
- UC Davis
- UC Santa Barbara
- UESTC
- University of Bristol
- University of Cambridge
- University of Chinese Academy of Sciences
- University of Hong Kong
- University of Science and Technology of China
- YOFC
- Zhangjiang Laboratory
- Zhejiang University
- ZTE
Report Summary
Huawei co-authors 7 of the 95 AI interconnect papers at ECOC 2025 and OFC 2026, more than any other company; Nvidia co-authors 4. The gap concentrates in one layer: scale-across, the optical link between data centers rather than inside a rack. Nvidia has no paper on this topic at either event, even as it has built scale-across into a pillar of its own product line, announcing Spectrum-XGS Ethernet on 22 August 2025 with CoreWeave as an early adopter. The optical research underpinning scale-across in this sample comes entirely from Huawei, European research groups and Chinese telecom operators.
Huawei’s papers span 4 of the 5 AI interconnect sub-topics we track; Nvidia’s span 3 and stay inside the data center. China’s cloud and AI-platform authors publish real operating data, including results from a production fleet and Alibaba’s insertion-loss-budget and bit-error-rate target for a 102.4 Tb/s switch (W2A.43), while the 6 comparable American papers report lab evaluations and demonstrations.
Nvidia’s own named optics ecosystem is thin in the research base. Of the eleven silicon photonics partners it named in March 2025, only two, Corning and Sumitomo Electric, appear anywhere in the 95-paper set. Furukawa Electric, the most-published laser supplier in the sample with 5 papers, is not on that list.
China’s share of AI interconnect research outpaces its share of each conference’s overall program: an index of 1.23 at ECOC 2025 and 1.05 at OFC 2026, against 0.70 and 0.42 for Europe.
This preview covers the full 2020-2026 ECOC and OFC trend, five interconnect layers, the country and organization split behind it, and the complete 95-paper appendix by conference, organization, country and sub-topic.
Table of Contents
- Summary
- What we counted
- AI interconnect’s share of conference papers grew from 4% to 9% in six years
- Huawei papers address 4 interconnect layers, Nvidia is 3
- What the hyperscalers publish, and what they hold back
- Nvidia’s named optics partners are not in the research base
- China outperforms on AI interconnect research, Europe underperforms
- Limitations of our analysis
- Implications
- Conclusion
- Appendix 1: Method, definitions and limits
- Appendix 2: Papers in this analysis
- Appendix 3: Report and publisher information
Coverage
Organizations mentioned:
- AIO Core
- AIST
- Alibaba
- Alibaba Cloud
- Ambition Photonics
- AttoTude
- AuthenX
- Ayar Labs
- Baidu
- Beijing University of Posts and Telecommunications
- Berxel
- Browave
- BT
- CAICT
- CEA-Leti
- China Mobile
- China Telecom
- China Unicom
- Ciena
- Coherent
- CoreWeave
- Corning
- CSA Catapult
- Deutsche Telekom
- Dublin City University
- Eindhoven University of Technology
- EPFL
- Ericsson
- Fabrinet
- FormFactor
- Foxconn
- Fraunhofer
- Fudan University
- Fujifilm
- Furukawa Electric
- GeNopsys
- GlobalFoundries
- Huawei
- Huazhong University of Science and Technology
- ICON Photonics
- imec
- Innolume
- Institute of Computing Technology, CAS
- Institute of Science Tokyo
- Kagawa University
- KDDI
- Keio University
- KPN
- Ligent
- Linfiber Technology
- Lumentum
- Max Planck Institute
- Meta
- Mitsubishi Electric
- Nanjing University
- National University of Defense Technology
- National University of Singapore
- Nokia
- NTT
- Nvidia
- Orange
- Peking University
- Politecnico di Milano
- QoreTek
- RCALS Centre for Advanced Lightwave Science
- Ruijie
- Samsung
- SENKO
- Shanghai Jiao Tong University
- Solinide Photonics
- SPIL
- Sumitomo Electric
- Suzhou Everbright Photonics
- Telefonica
- TFC Communication
- TIM
- TSMC
- TSRI
- Tyndall
- UC Davis
- UC Santa Barbara
- UESTC
- University of Bristol
- University of Cambridge
- University of Chinese Academy of Sciences
- University of Hong Kong
- University of Science and Technology of China
- YOFC
- Zhangjiang Laboratory
- Zhejiang University
- ZTE
Figure & Charts
- Figure 1: AI interconnect reached 6.3% of ECOC papers in 2025 and 8.8% of OFC 2026
- Figure 2: Huawei co-authors 7 of the 95 recent AI interconnect papers, Nvidia 4
- Figure 3: Scale-up, scale-out and scale-across, and who publishes on each
- Figure 4: The two cloud papers reporting production fleet data are both Chinese
- Figure 5: Europe holds 51.6% of ECOC 2025 papers and 36.4% of its AI interconnect papers
- Figure 6: China leads or ties 4 of the five AI interconnect sub-topics, Europe leads 1, the smallest of the five at 8 papers
- Table 1: The 95 AI interconnect papers behind Figures 2 through 6, by conference, organization, country and sub-topic
