OCP Global Summit 2025 | Google | OCP Optical Circuit Switching Subproject Update

Introduction

At OCP Global Summit 2025, Google presented the latest progress on the OCP Optical Circuit Switching (OCS) subproject. Led by Google engineers, the effort aims to drive open standardization of OCS so that optical circuit switching can become a new core technology for AI data centers and cloud network infrastructure. The core idea of OCS is "all-optical, zero conversion, open collaboration": replacing the bottlenecks of traditional electrical switching with light to deliver low-latency, efficient and programmable data center networks.

Content

1. What Is OCS (Optical Circuit Switching)?

At its core, an OCS is a "fiber cross-connect box" that sets up a purely optical path between two ports with no optical-electrical conversion. This means:

• Protocol-agnostic: supports Ethernet, InfiniBand and even custom protocols.

• Rate-agnostic: no hardware replacement needed when link speeds are upgraded.

• Low latency, low power: no E/O/OE conversion, and no packet queuing or buffering.

  OCS can be implemented in several ways, including robotic fiber switches, MEMS mirrors, or silicon photonics-based beam steering (SiPh-based beam steering).

2. Why Data Centers Need OCS

Traditional Ethernet switches are flexible, but they consume a lot of power, have fixed latency, and every packet has to pass through the ASIC pipeline.

OCS offers:

• Cut-through paths with zero intermediate latency

• Traffic independence (unaffected by packet type)

• Bufferless security (packets cannot be intercepted or modified)

For large AI clusters (for example, training networks with 10,000+ GPUs), this significantly reduces tail latency, saves power and simplifies the physical cabling topology.

3. Goals and Structure of the OCS Working Group

The OCS Working Group formed by Google and OCP aims to build an open OCS ecosystem, with four main workstreams:

1. Open Hardware Specs: define a common optical switching architecture and control interface.

2. Software Specs: standardize APIs for control, management, telemetry and alarms.

3. AI Cluster Optimization: develop use cases and white papers.

4. Interoperability: ensure OCS from different vendors can work together through a unified interface.

Software specs are currently the most active area. The working group has begun collaborating with Linux Foundation, SONiC, OpenConfig, IETF and other organizations to build an integrable open-source software stack.

4. Google's Implementation and Interface Design

Google is both a user and a builder of OCS.

Its strategy:

• Use SONiC as the core platform, reusing the existing Ethernet switch software base.

• Extend the OCS management interface into the UMF (Unified Management Framework), supporting three open APIs:

  • GNMI: configuration and telemetry.

  • GNOI: operations and optical path control (such as creating, deleting and replacing connections).

  • GNSI: security and key management.

• Add active monitoring and health checks to OCS to track optical connection status and report errors.

5. Security and Deployment Flow

Google designed a rigorous Secure Installation Protocol for OCS systems, covering DHCP, boot server, certificate enrollment and attestation, so that every step from factory to deployment is protected against supply chain attacks and man-in-the-middle intrusion.

6. Use Cases and Application Layers

Google's main use case comes from Project Apollo (a reconfigurable optical switching architecture). OCS is deployed mainly in:

• Scale-out layer (across racks): replacing the spine network.

• Scale-up layer (within racks): dynamically connecting GPU/CPU nodes.

• AI cluster topology reconfiguration: adjusting topology across training phases to optimize GPU bandwidth utilization.

For Google, OCS works like a "programmable optical patch panel" that can reconfigure the physical network on the fly, without recabling or downtime.

Conclusion

In this OCP 2025 update, Google clearly positioned OCS as a structural foundation of next-generation data centers. By extending the SONiC software stack into optical switching and combining it with OCP's open specifications, Google hopes to make programmable optical topology mainstream. Future data centers will be not just software-defined networks (SDN), but "Optical Defined Networks (ODN)".

Further Perspectives

1. Technology impact

   • This is an important turning point for optical interconnect. By merging the OCS control plane with SONiC, Google signals that optical networks are entering the software-defined era. Future AI clusters will no longer be statically cabled, but reconfigured in real time based on traffic and workloads.

   • SiPh (silicon photonics) and MEMS will become key foundations for OCS modules, creating strong resonance between OCS and the silicon photonics supply chain.

2. Supply chain observations

   • As Google pushes open APIs, third-party vendors (such as Lumentum, Calient, Cisco and Furukawa) have a chance to enter the hyperscale market with compatible products.

   • On the other hand, Google building its own OCS shows that hyperscalers are moving toward in-house optical equipment, which will shift the balance of power in the optical switching supply chain.

3. Market trends

   • OCS adoption marks a shift from electrical switching → optical switching → integrated optical control plane.

   • This will profoundly affect topology and cabling design in future AI-dedicated data centers (AI-DC), and could spawn a new business model: "Optical Switching as a Service".