AMD, Broadcom, and Nvidia Team Up With Hyperscalers on Optical AI Interconnect
Key Facts
- What: AMD, Broadcom, Nvidia, Microsoft, Meta, and OpenAI have formed the Optical Compute Interconnect (OCI) Multi-Source Agreement (MSA) group to develop an open, protocol-agnostic optical scale-up interconnect specification for AI clusters.
- Target Speeds: Starts at 200 Gb/s per fiber (4 wavelengths × 50 Gb/s) using NRZ signaling and WDM, scaling to 800 Gb/s per fiber initially and eventually reaching 3.2 Tb/s per fiber and beyond.
- Supported Technologies: Pluggable optical modules, on-board optics, and co-packaged optics (CPO) integrated directly with compute silicon.
- Purpose: Replace copper interconnects with optical cables for short-reach scale-up links inside AI racks to support larger clusters at high speed and predictable power.
- Key Benefit: Enables different processors and protocols (such as AMD/Broadcom UALink and Nvidia NVLink) to operate over the same fiber infrastructure and multi-vendor switches.
Lead paragraph
A powerful alliance of chipmakers and hyperscalers has launched a new industry group to define the optical interconnect technology that will power the next generation of massive AI training clusters. AMD, Broadcom, Nvidia, Microsoft, Meta, and OpenAI announced the formation of the Optical Compute Interconnect (OCI) Multi-Source Agreement this week, aiming to create a common optical physical layer for short-reach scale-up connections within AI racks. The initiative addresses the growing limitations of copper interconnects as AI systems expand, promising to deliver dramatically higher bandwidth at manageable power levels while maintaining compatibility with competing protocols like NVLink and UALink.
The Growing Need for Optical Scale-Up
As AI clusters continue to balloon in size, traditional copper-based interconnects are reaching their practical limits for connecting accelerators and switches inside the same rack or scale-up domain. While optical interconnects have already become common for scale-out connectivity between racks, the industry is now preparing for the day when optical links will be required even for shorter, high-bandwidth scale-up connections.
The OCI MSA specifically targets these short-reach links used inside large AI systems and racks. According to the group's founding members, the technology will define a common optical physical layer (PHY) based on NRZ signaling and wavelength-division multiplexing (WDM). The initial specification starts modestly at four wavelengths times 50 Gb/s, delivering 200 Gb/s per direction per fiber, but includes a clear roadmap to 800 Gb/s per fiber in the near term.
Over time, the consortium plans to expand both the number of wavelengths and the signaling rates, with an ultimate target of 3.2 Tb/s per fiber and beyond. This massive bandwidth scaling is critical as next-generation AI accelerators demand ever-higher interconnect performance to avoid becoming bottlenecked by data movement.
Protocol-Agnostic Design Preserves Competition
One of the most notable aspects of the OCI MSA is its protocol-agnostic approach. The common optical layer will allow different AI processors and their respective interconnect protocols to operate over the same fiber infrastructure and switches from multiple suppliers.
This means Nvidia's NVLink, AMD and Broadcom's UALink, and potentially other future protocols can all leverage the same standardized optical interconnects. The design preserves the competitive differentiation of each company's proprietary protocols while providing hyperscalers with greater flexibility in building their AI infrastructure.
"The growing need for optical scale-up interconnect to support large AI systems later this decade is clear," said Brian Amick, Senior Vice President of Technology & Engineering at AMD, in a statement. "AMD is a founding member and strong supporter of the OCI MSA as it establishes an open specification for the industry to foster a robust, multi-vendor optical scale-up interconnect ecosystem."
Broadcom echoed this sentiment, highlighting its experience with co-packaged optics. "Broadcom is proud to draw upon our multi-generational CPO platform and industry partnerships to drive the OCI specification forward," said Near Margalit, Vice President & General Manager of the Optical Systems Division at Broadcom. "The OCI-MSA allows for seamless integration with existing electrical SerDes-based ASICs while providing a clear path to direct ASIC integration, ensuring the ecosystem remains flexible and high-performing."
Hyperscaler-Driven Initiative
Unlike many traditional industry consortia that are primarily organized and led by independent hardware vendors, the OCI MSA is notably hyperscaler-driven. Microsoft, Meta, and OpenAI are key participants alongside the silicon and networking suppliers.
This structure reflects the increasing influence of the largest cloud providers and AI companies in defining the future of AI infrastructure. These organizations are the primary customers for massive GPU clusters and have the clearest view of the bottlenecks that will emerge as systems scale to tens or hundreds of thousands of accelerators.
The MSA will support multiple form factors, including pluggable optical modules, on-board optics, and co-packaged optics (CPO) that integrate optical engines directly with compute silicon. This flexibility should help accelerate adoption across different system architectures and deployment timelines.
Impact
This collaboration represents a significant step toward solving what many experts see as the next major bottleneck in AI scaling: high-speed, power-efficient data movement within large clusters. By standardizing the optical physical layer, the group aims to simplify system integration, reduce development risk, and shorten deployment cycles for new generations of AI hardware.
"The common optical layer will enable different processors and interconnect protocols to operate over the same fiber infrastructure and switches from different suppliers, ensuring flexibility for hyperscalers while retaining the competitive advantages of the protocols used by developers of AI accelerators."
For developers and AI companies, this means greater choice in building hybrid clusters that might combine accelerators from multiple vendors without being constrained by proprietary interconnect limitations. For hyperscalers like Microsoft, Meta, and OpenAI, it provides a clearer path to constructing even larger training systems while managing power and cooling constraints.
The initiative also highlights the intensifying competition and collaboration in the AI infrastructure space. While Nvidia currently dominates the AI GPU market, companies like AMD and Broadcom are pushing hard with alternative solutions, and hyperscalers are actively supporting open standards that prevent vendor lock-in.
What's Next
The OCI MSA group is expected to begin detailed technical work immediately to define the initial specification. While no specific timeline for the first optical components or systems has been announced, the group has indicated that optical scale-up interconnects will be needed to support large AI systems later this decade.
The roadmap calls for progressive increases in both wavelength count and per-wavelength data rates, ultimately targeting 3.2 Tb/s per fiber. As the specification matures, it should enable broader adoption of co-packaged optics and other advanced photonic technologies directly integrated with AI accelerators and switches.
Success of the OCI MSA could accelerate the shift from copper to optical interconnects at ever-shorter distances, fundamentally changing how future AI supercomputers are architected. The involvement of all major AI accelerator vendors and the leading hyperscalers suggests strong industry momentum behind the effort.