Hyperscalers back new optical interconnect standard for large-scale AI systems
AMD, NVIDIA, Microsoft and others launch an optical interconnect consortium to support scaling next-generation AI infrastructure.
Major technology companies including AMD, Broadcom, Meta, Microsoft, NVIDIA and OpenAI have formed a new industry consortium to develop an open specification for optical scale-up interconnects in AI infrastructure. The Optical Compute Interconnect (OCI) Multi-Source Agreement group aims to create a multi-vendor ecosystem designed to support the growing compute demands of large-scale AI systems.
Table Of Content
The initiative reflects mounting pressure on AI infrastructure as model sizes and training clusters expand. By aligning on a common optical interface standard, the consortium intends to provide hyperscalers and hardware vendors with a shared foundation for building future AI rack architectures.
AI scale is exposing the limits of copper interconnects
The formation of the OCI MSA highlights a structural constraint in current AI infrastructure design. As large language models increase in complexity, copper-based connectivity is increasingly constrained by physical reach and power limitations, affecting the scale-up architectures used inside AI clusters.
The OCI specification proposes a transition from copper-based connectivity toward optical interconnects that can support larger and more power-efficient compute domains. The design combines non-return to zero modulation and wavelength division multiplexing optical technologies while shifting connectivity from a module-centric model toward tighter integration between optics and compute silicon.
This architectural change is intended to increase bandwidth density and improve system scalability while maintaining aggressive power targets required by modern AI infrastructure.
Open optical interfaces aim to enable multi-vendor AI racks
A key objective of the consortium is to create an interoperable optical interface that allows different hardware components to operate within a shared infrastructure framework. The specification defines a common optical physical layer that enables hyperscalers to connect processor units and scale-up switches across vendors.
This approach enables a disaggregated architecture in which compute and networking components can be integrated across suppliers without requiring proprietary interfaces. The consortium argues that a standardised interface could reduce integration risk and shorten development cycles for companies deploying large-scale AI systems.
Dan Rabinovitsj, Vice President of Hardware Systems at Meta, said: “The appetite for technology to address the power and cost constraints impacting AI cluster design is real and imminent. We encourage adoption of this OCI protocol to decouple the need for larger scale-up domains from the limitations of electrical backplanes in high performance AI clusters.”
Roadmap targets multi-terabit optical connectivity
The OCI specification also outlines a development roadmap intended to support multiple generations of optical interconnect technologies. Early designs include high-density optical interfaces capable of supporting up to 800Gbps per fibre using OCI GEN1 and GEN2 technologies.
The roadmap extends further as AI infrastructure continues to scale. Future iterations are expected to increase wavelength counts and data rates, targeting connectivity speeds of up to 3.2Tbps per fibre and beyond while supporting multiple optical deployment models including pluggable, on-board and co-packaged optics.
Richard Ho, Head of Hardware at OpenAI, said: “The continued improvement of artificial intelligence relies on scaling of AI supercomputers with more petaflops, more memory bandwidth and, importantly, more network bandwidth across larger domains requiring further reach. The OCI MSA will be critical to allow the industry to build the AI systems that will get us to AGI.”
