Photonic Networks for AI: Challenges and Opportunities
Original Article by SemiVision Research (Oriole Network, Google , AMD , Grok , Microsoft , Nvidia ,Coherent ,Lumentum ,Huber+Suhner Polatis,iPronics,nEye Systems ,DiCon Fiberoptics,Telescent,Foxconn)
Oriole Networks is a deep-tech startup based in London, founded in 2023, and spun out from research at University College London (UCL). The company focuses on developing photonics-based networking solutions designed to enhance the performance of artificial intelligence (AI) systems, while significantly reducing data center energy consumption.
Company Background and Mission
Oriole Networks’ core technology is based on over two decades of optical networking research led by Professor George Zervas at UCL. Together with co-founders Alessandro Ottino and Joshua Benjamin, the team is working to revolutionize data movement within data centers, enabling higher energy efficiency and better infrastructure utilization through photonic technologies.
Technological Innovation and Applications
Leveraging advanced photonic switching technology, Oriole Networks is building AI chip-level networks that tightly integrate processing capabilities. Their solution claims to accelerate large language model (LLM) training by up to 100x, while consuming only a fraction of the power required by conventional hardware. This breakthrough addresses both speed and latency challenges in AI computing, while helping to manage the growing energy demandsof modern data centers.
Funding and Development
Since its inception, Oriole Networks has raised over $35 million USD. In March 2024, the company announced a €11.6 million seed round, backed by investors such as UCL Technology Fund, Clean Growth Fund, XTX Ventures, and Dorilton Ventures.
Future Outlook
Oriole Networks aims to build a photonic networking ecosystem that will transform the AI infrastructure landscape, addressing scalability and energy bottlenecks, while fostering greater competition at the GPU and accelerator system level. The company plans to deliver its early product offerings by 2025, accelerating photonics adoption in AI computing environments.
SemiVision’s Upcoming Analysis
In response to this presentation, SemiVision will provide expanded analysis on the AI networking landscape, taking Oriole Networks as a case study. The analysis will cover key concepts introduced by the company, including their perspective on training and inference chip architectures, as well as a cost comparison between in-house ASICs developed by CSPs and off-the-shelf NVIDIA chips.
Additionally, since Oriole Networks references Optical Circuit Switching (OCS)—a concept popularized by Google’s research published at IEEE—SemiVision will provide an in-depth comparison of OCS vs. Electronic Packet Switching (EPS).
If your company is involved in Optical Circuit Switching (OCS) technologies, we welcome the opportunity to connect with you.
OCS solutions today span several key technology categories, including MEMS, Liquid Crystal, Robotic Fiber Connection, Piezoelectric, and Silicon Photonics.
We invite companies with expertise in any of these areas to reach out to us and explore opportunities to exchange knowledge and collaborate on advancing OCS technologies and ecosystem development.
For readers interested in understanding next-generation networking architectures, this analysis will be a must-read.
AI Performance doubles every 9 months
Training vs Inference XPU ( Nvida , AWS ,AMD ,GroqCard LPU ,Google Ironwood)
AI Network Demands: Why Collective Communication Networks Are Becoming Critical Infrastructure
Training: Maximizing Model Completion per Megawatt
Inference: Maximizing Tokens per Kilowatt
Shared Network Challenges: Architecture Defines the Upper Bound
Network Architecture Is the Next AI Battleground
From Electronic Packet Switching (EPS) to Optical Circuit Switching (OCS)
EPS (Electronic Packet Switching): Current Role and Limitations
OCS (Optical Circuit Switching): Potential Benefits and Real-World Challenges
Key Breakthroughs Required for OCS to Replace EPS
OCS Is the Future of AI Networking, But Critical Challenges Remain