CPO Moves Toward Open Standards! Lightmatter, Dell, Qualcomm, GUC, FIT, and 19 Other Companies Join Forces; First OCP Specifications to Launch in Q4

As AI chip computing power continues to scale, the next competitive frontier in data centers is shifting from 'how fast you can compute' to 'how fast you can move data.' With high-speed SerDes transmission rates approaching 448G, the effective transmission distance of copper wiring may shrink to just tens of centimeters. Co-Packaged Optics (CPO) is evolving from individual vendor technology races into cross-supplier open system architectures.

Silicon photonics startup Lightmatter today (14th) announced that its 'Open Silicon Photonics for AI Systems' initiative has officially become a working group under the Open Compute Project (OCP). The alliance now includes 19 companies: Celestica, Dell Technologies, Flex, Foxconn Interconnect Technology, Global Unichip Corporation (GUC), Hefei Computer, Keysight Technologies, Lightmatter, Qualcomm, and QCT.

The consortium simultaneously released a 300-page white paper titled 'Architecture Vision: Open Silicon Photonics for AI Systems,' aiming to establish a common CPO architectural blueprint. This will enable AI computing clusters compliant with Modular Hardware System (MHS) and Open Rack v3 (ORv3) standards to scale from 72 nodes to over 1,024 nodes while maintaining interoperability across different suppliers.

Notably, participants include not only U.S. chip, server, and measurement companies but also key players from Taiwan’s supply chain—ASIC design service provider GUC (3443), connector and interconnect specialist Foxconn Interconnect Technology, and server manufacturer QCT. This indicates that CPO open architecture now extends beyond individual optical components to encompass chips, interconnects, and AI server system levels.

However, while the official announcement lists participating companies, it does not disclose individual technical roles within the working group. The specific standards Taiwan-based firms will contribute to remain to be clarified in future updates.

AI Clusters Span Multiple Racks: 448G May Limit Copper to Tens of Centimeters

The joint push for open CPO architecture by 19 companies directly addresses interconnect bottlenecks in AI data centers. As AI accelerators grow more powerful, data centers can no longer focus solely on the compute performance of individual GPUs or ASICs. When hundreds or thousands of AI accelerators form scale-up computing clusters, the ability to exchange data at high speed while minimizing energy consumption for data movement is becoming a critical factor in overall AI system performance.

According to technical background disclosed in the white paper, as scale-up architectures span multiple racks and SerDes transmission rates approach 448G, copper’s effective transmission distance could shrink to just tens of centimeters, pushing the industry toward full optical interconnect architectures.

Dr. Nick Harris, founder and CEO of Lightmatter, stated: 'For AI infrastructure performance, interconnect technology has become as important as compute performance.' He believes the industry’s shift toward silicon photonics will further advance the capabilities of cutting-edge AI models.

This signals a shift in how AI hardware competitiveness is measured. While market focus was previously on GPU/ASIC compute performance and memory bandwidth, as AI clusters scale up, whether switches, SerDes, and optical interconnects enable efficient collaboration among numerous accelerators now directly impacts the actual utilization of expensive AI chips.

The Next Battle for CPO: It’s Not Just About Technology—Can Different Vendors Work Together?

Silicon photonics and CPO are not new concepts. What makes this OCP working group noteworthy is the industry’s attempt to move CPO from proprietary, company-specific solutions to a system-level common architecture that solves interoperability between different vendors.

The 300-page white paper serves as the foundational blueprint for a common CPO architecture. According to current plans, the architecture will align with MHS and Open Rack v3 (ORv3) standards, enabling AI computing clusters to scale from 72 to over 1,024 nodes, with multi-vendor interoperability built into the design.

In other words, if CPO is to be widely adopted in AI data centers, the industry must solve not just 'how close can optics get to the chip,' but how different generations of XPUs, switches, optical components, and server systems can operate together across varying data center lifecycles and suppliers.

Harris acknowledged that achieving consensus among 19 companies on a system-level architectural vision was not easy. He emphasized that the white paper is just the beginning—OCP community efforts will now focus on developing a common CPO rack architecture based on MHS and integrating it into hyperscalers’ deployment workflows.

'In truth,' he added, 'the real work has only just begun.'

Architecture Designed for 100,000-XPU Clusters, Not Tied to a Single Optical Technology

Another key design aspect is OCP’s decision not to bind the open architecture to a single optical technology. The current architectural vision adopts a 'technology-agnostic' approach, supporting not only silicon photonics but also Vertical-Cavity Surface-Emitting Lasers (VCSELs) and Micro LEDs.

More importantly, this reference architecture is not designed for a single generation of AI accelerators. It aims to span multiple SerDes generations and different switch and XPU product cycles, enhancing scalability through a multi-supplier supply chain.

In terms of scale, the alliance’s ambitions go beyond current hundreds or thousands of AI accelerators. Officially, the architecture is designed to meet the performance and power requirements of clusters with up to 100,000 XPUs.

This highlights the importance of CPO standardization: as AI cluster sizes grow, if different chips, switches, optical devices, and rack systems use closed, proprietary architectures, the complexity of upgrading, expanding, and integrating multi-vendor equipment in data centers will increase significantly. The common framework OCP is attempting to build aims to lower interoperability barriers for future large-scale deployments.

IDC: Copper’s Physical Limits Are Squeezing AI Accelerator Utilization

Beyond Lightmatter, market research firm IDC also identifies interconnects as the next major bottleneck in AI infrastructure. Jeff Janukowicz, IDC’s Vice President of Global Computing Infrastructure and Technology, stated that as demand for AI system scaling accelerates, 'optimizing AI system energy efficiency and data transfer speed has become as important as raw compute power.'

He further noted that copper’s physical limits are increasingly constraining the scalability and accelerator utilization of large AI computing clusters—and this gap will widen with each new generation of AI accelerators.

Janukowicz believes the vendor-neutral architecture established by 'Open Silicon Photonics for AI Systems' will accelerate ecosystem collaboration, helping CPO move from early deployment to mainstream AI infrastructure.

The 300-Page White Paper Is Just the Beginning—First OCP Specs Due in Q4

Thus, the release of the 300-page white paper by 19 companies marks the first step in moving CPO open architecture from 'industry consensus' to 'actual standards.' The critical milestone to watch will be in Q4 of this year. According to current plans, the working group will submit its first specifications, expected to be finalized in Q4 2026.

The key challenge ahead will be whether OCP can translate the current system architecture consensus among 19 companies into deployable technical specifications and attract more chip, switch, optical communications, and server vendors to join.

As AI data centers continue to scale toward thousands, tens of thousands, and even 100,000 XPUs, CPO competition may shift from 'who has higher bandwidth or lower power' to 'who can build a scalable, cross-vendor, cross-generational system ecosystem.' For Taiwan-based suppliers like GUC, FIT, and QCT already in the initiative, how the first OCP specifications in Q4 are defined will be a key indicator for their future entry into next-generation AI optical interconnect architectures.