OCI-MSA Sets the Spec for Optical AI Interconnects

By Amara Osei

OCI-MSA Gen 1 specifies eight wavelengths per fiber, four in each direction, with a 50 Gbps effective rate per wavelength after error correction. The 24-page Optical Compute Interconnect Multi-Source Agreement gives AI scale-up a published, open route for linking accelerators and switches across vendors.

The consortium’s six named founding members—AMD, Broadcom, Meta, Microsoft, Nvidia and OpenAI—include two GPU giants, the largest maker of custom AI compute, two hyperscalers and one of the two frontier-model companies. The recently formed consortium has received implementation-plan presentations from AMD, Broadcom, Meta and Microsoft at recent industry conferences.

At the September AI Infra Summit, Meta optical infrastructure director Drew Alduino appeared with Near Margalit, Broadcom VP and GM for Optical Systems, and AMD senior director for photonics Matthew Streshinsky. Alduino said: “We want AMD GPUs talking to Broadcom switches,” and “We want all of our compute and switches to be interoperable.”

Near Margalit said interoperability ultimately depends on what appears on the fiber.

When does optical reach inside AI racks?

Participants disagree on the precise copper-to-optical transition, but they broadly expect the move to all-optical interconnects over the next five years. Optical already dominates scale-out links. The next step is scale-up: connect racks in pods of hundreds to thousands of XPUs, then replace copper within each rack.

Today’s pluggable transceivers mostly use a single DR wavelength over two or four fibers. More performance would require power-hungry, faster physical-layer interfaces or additional fibers. Because connectors cost more than fiber, both routes raise system cost.

The design center is shifting toward bidirectional, multi-wavelength links, moving from near-packaged optics (NPO) to co-packaged optics (CPO). Other architectures may emerge, but OCI-MSA has emerged as the consensus candidate for scale-up.

Takeaway: The next market step is optical links inside AI racks, not only between them.

What must OCI-MSA standardize?

The consortium lists three objectives:

• Create an open, interoperable AI scale-up specification that removes proprietary constraints and supports cross-vendor compatibility in hyperscale clusters.
• Provide architectural guidance for low-power optical SerDes, delivering the latency, bandwidth and energy efficiency required for next-generation AI.
• Build a scalable, multi-vendor supply chain linking hyperscale end users and silicon innovators, reducing integration risk, accelerating development and limiting vendor lock-in during long-term deployments.

Meta’s stated priorities are:

• Standardize as little as possible while preserving interoperability and room for rapid innovation.
• Use multiple wavelengths so optical bandwidth does not depend on ever-faster SerDes.
• Make OCI operate like Ethernet.
• Move from Gen 1’s four 50 Gbps wavelengths in each direction to a Gen 2 design in definition that doubles bandwidth, followed by another doubling in Gen 3.
• Report that 200G × 16 wavelengths had already been shown as viable, a 16-fold increase over Gen 1 bandwidth.

Takeaway: OCI-MSA targets compatibility at both the link and supply-chain levels.

What does Gen 1 specify?

The Gen 1 document defines eight wavelengths per fiber, split four in each direction. Each wavelength runs at 53.125 Gbps using NRZ, and wavelengths within each directional group sit slightly more than 2 nm apart. Its reference link uses 500 m of SMF-28 fiber with 2.5 dB total insertion loss, driven mainly by connectors between transceivers. After error correction and miscellaneous signaling, the effective rate falls to 50 Gbps per wavelength.

The specification supports pluggables, NPO, CPO inside the package and CPO on the interposer.

At OCP in October, Broadcom hyperscale strategy and products director Rajiv Pancholy compared OCI-MSA with alternatives at equal bandwidth. For 200 Gbps aggregate bandwidth in one direction, Broadcom’s comparison placed CPO OCI at the best combination of low link power per bit and long reach. It matched pluggable reach at one-third of pluggable transceiver power and had a scalable bandwidth roadmap.

Takeaway: Gen 1 sets a 200 Gbps-per-direction baseline with a 500 m reference link.

How does OCI scale beyond Gen 1?

At PECC 2026 in October, AMD identified four ways to raise OCI bandwidth:

• Increase wavelengths from four to eight, 16, 32 and beyond, subject to the O band’s width limit.
• Raise per-wavelength rates from 50 Gbps to 100, 200, 400 Gbps and higher.
• Move from one polarization to two.
• Replace NRZ with PAM/QAM-n, with advanced signaling likely required at the highest rates.

Together, these options provide more than 100× scaling. If OCI doubles every one to two years, they create decades of per-fiber runway; adding fibers expands capacity further.

At the same event, Microsoft discussed OCI and the need for a common standard. Its first scale-up optical wave places CPO on the switch and NPO with the accelerator. The second puts CPO at both ends.

Takeaway: More wavelengths, higher lane rates, extra polarization and advanced modulation can expand the same fiber well beyond Gen 1.

When does interoperability become practical?

OCI-MSA is currently the only published optical scale-up standard receiving regular discussion at AI Infra, OCP and PECC from multiple backers, including AMD, Broadcom, Meta and Microsoft.

A likely 2027 milestone is the first fully functional Gen 1 implementations, possibly accompanied by interoperability demonstrations between suppliers. If several vendors adopt the specification, Alduino’s Ethernet analogy could become its practical role: a common layer above optical links. Other standards or proprietary interconnects remain possible, but OCI-MSA leads the field today.

Takeaway: OCI-MSA is the front runner, but interoperable Gen 1 hardware will test its claim to become a common standard.