ECOC 2025 Tech Spotlight: The World's First Pure Photonic AI Network

Introduction

As AI factories drive soaring demand for compute and networking, today's data center network architectures face the challenges of high power, high latency and insufficient reliability. Conventional designs rely on multiple tiers of electrical packet switching, requiring O-E-O conversion at every tier, which brings enormous energy consumption and latency.

At ECOC 2025, Oriole Networks proposed an entirely new network architecture: a Pure Photonic Network. This design abandons electrical packet switching altogether and instead uses nanosecond photonic switching to build a flat, all-optical network, fundamentally changing how data centers operate.

Contents

1. Bottlenecks in Today's Data Centers

• Multi-tier switching architecture: server-to-server traffic must traverse multiple tiers of switches - ToR (Top-of-Rack), Leaf, Spine and Core.

• O-E-O conversion: every tier converts optical signals to electrical and back to optical, resulting in high power and high latency.

• Reliability issues: hundreds of thousands of transceivers and active components become sources of failure, frequently causing interruptions when running AI workloads.

2. The Core Idea of a Pure Photonic Network

• Flat architecture:

  • Servers connect to each other directly over optical links, no longer relying on electrical packet switches.

  • The network core is purely passive - "just a piece of glass" - and consumes no power.

• Nanosecond photonic switching:

  • Avoids the limitation of OCS (optical circuit switching), which can only reconfigure the network "a few times a day."

  • Switches at high speed with fine granularity, enabling contention-free communication.

• Fewer devices:

  • Take a 32,000-node network as an example: a conventional architecture needs 2,500 switches and more than 160,000 transceivers.

  • A pure photonic architecture needs transceivers only at the nodes, with nothing but fiber in between.

3. Key Advantages

1. Dramatically lower power:

   • Conventional switches consume power at every tier; the pure photonic core consumes zero power, with energy used only at the nodes.

2. Significantly lower latency:

   • Removing multiple tiers of O-E-O conversion delivers low and predictable latency.

3. Higher reliability:

   • With fewer active components, the failure blast radius shrinks dramatically.

   • A single transceiver failure affects only part of one GPU's capacity.

4. Higher compute efficiency:

   • GPUs no longer wait on network transfers and can keep running efficiently.

   • In AI inference workloads, the pure photonic network maintains 99% efficiency even at 99% load, far better than packet networks.

4. Comparison with OCS and Electrical Packet Switching

• NVIDIA Quantum-X: electrical packet switching + CPO, still a hybrid architecture.

• Google MEMS OCS: can only reconfigure topology "a few times per day" - too coarse-grained.

• Oriole Networks:

  • Provides fast (nanosecond-scale), fine-grained pure optical switching.

  • Supports new collective operations that map CUDA jobs directly onto the network, improving AI workload efficiency.

5. Bringing the Technology to Market

• Origin: the technology comes from a decade of research at University College London (UCL) and is being commercialized by Oriole Networks.

• Manufacturability: it does not depend on rare materials or specialty fabs, so it can scale quickly to high volume.

• In-house optical transceivers: overcome the limitations of conventional transceivers and support nanosecond switching.

Conclusion

The "pure photonic AI network" Oriole Networks presented at ECOC 2025 shows a thoroughly disruptive line of thinking:

1. Remove electrical packet switching and build a network core made only of fiber and optical switching.

2. Use nanosecond optical switching to replace slow OCS, enabling fine-grained, low-latency all-optical connectivity.

3. Dramatically cut power and failure rates while improving GPU compute efficiency.

If this architecture can be deployed successfully, it will mark a "paradigm shift" in data center networking, truly realizing AI factories driven by photons.