OCI-MSA Gen 1 Specifies 200G per Fiber for AI Scale-Up

By Sophie Lindqvist

OCI-MSA’s 24-page Gen 1 specification defines 200 Gbps aggregate in one direction over each optical fiber, using four effective 50 Gbps wavelengths.

The recently formed Optical Compute Interconnect Multi-Source Agreement wants an open, interoperable optical link specification for AI scale-up.

Founding members include AMD, Broadcom, Meta, Microsoft, Nvidia and OpenAI. The mix spans two GPU giants, the largest maker of custom AI compute, two hyperscalers and one of two frontier-model companies.

AMD, Broadcom, Meta and Microsoft have already presented OCI-MSA plans at recent industry conferences. The source identifies no corresponding implementation plans from Nvidia or OpenAI.

What changes as AI clusters scale?

Optical interconnects already dominate scale-out. Their next target is scale-up: connecting racks in pods of hundreds to thousands of XPUs, meaning AI compute units. Optical links would then move inside racks and replace copper.

Timing remains disputed. The source nevertheless reports almost universal agreement that all-optical interconnects will arrive within five years. That remains a directional forecast, not a dated product commitment.

Today, optical systems rely mainly on pluggable transceivers using one wavelength, or DR, with two or four fibers. More bandwidth would require power-hungry faster physical-layer interfaces, additional fibers or both.

Connectors cost more than fibers, so adding fibers also raises cost. The preferred path uses multiple wavelengths and bidirectional links, moving through near-packaged optics, or NPO, to co-packaged optics, or CPO.

Alternatives may develop, but the source identifies OCI-MSA as the emerging consensus.

What is OCI-MSA trying to standardize?

The consortium frames interoperability as a supply-chain problem as much as an optics problem, aiming to let hyperscale users combine silicon from competing vendors without proprietary links. It lists three objectives:

• Create an open, interoperable AI scale-up specification that breaks proprietary constraints and enables cross-vendor hyperscale clusters.
• Provide architectural guidance for low-power optical SerDes with the latency, bandwidth and energy efficiency required for next-generation AI.
• Build a scalable, multi-vendor supply chain linking hyperscale users and silicon innovators, reducing integration risk, accelerating deployment and limiting vendor lock-in.

At a September AI Infra Summit panel, Meta optical infrastructure director Drew Alduino appeared with Broadcom Optical Systems Division VP and GM Near Margalit and AMD senior photonics director Matthew Streshinsky.

Alduino said: “We want AMD GPUs talking to Broadcom switches. We want all of our compute and switches to be interoperable.”

Margalit said interoperability ultimately depends on what appears on the fiber. Alduino outlined Meta’s position:

• Standardize as little as possible while preserving interoperability and room for rapid innovation.
• Use multiple wavelengths so optical bandwidth can rise without ever requiring faster SerDes.
• Make OCI look like Ethernet today.
• Gen 1 uses four wavelengths at 50 Gbps in each direction. Gen 2, now in definition, doubles bandwidth; Gen 3 doubles it again.

Alduino also said a 200G-per-wavelength, 16-wavelength configuration had already proved “viable,” delivering 16 times Gen 1’s bandwidth.

The source provides no test date, supplier or link length for that demonstration, so it does not prove production readiness.

What does Gen 1 specify?

The public Gen 1 specification runs 24 pages and supports four physical formats:

• Pluggable transceivers, NPO, CPO in the package and CPO on the interposer.
• Eight wavelengths, or channels, per fiber, divided into four in each direction.
• Each wavelength carries 53.125 Gbps NRZ. The wavelength groups sit slightly more than 2 nm apart.
• An effective 50 Gbps per wavelength remains after error correction and miscellaneous signaling.
• A reference 500-meter SMF-28 link has 2.5 dB total insertion loss, mainly from connector losses.

At OCP in October, Broadcom’s Rajiv Pancholy, director of hyperscale strategy and products, presented OCI-MSA’s claimed power and reach profile against alternatives at equal bandwidth.

Can optical links cut power?

Broadcom’s presentation claimed that OCI in CPO format offered the best combination of low link power per bit and long reach at 200 Gbps aggregate in one direction.

It reached the same distance as pluggables at one-third of a pluggable transceiver’s power, according to Broadcom.

That is Broadcom’s own comparison. The source presents no independent deployment test, production pricing or volume data.

How far can the roadmap scale?

At PECC 2026 in October, AMD presented several paths beyond Gen 1:

• Raise the wavelength count from four to eight, 16 and 32, with an upper limit set by the O-band width.
• Raise the rate from 50 Gbps to 100, 200 and 400 Gbps.
• Move from one polarization to two.
• Move from NRZ to PAM or QAM-n signaling, likely for the highest rates.

Together, these paths permit more than 100 times Gen 1 bandwidth. If OCI doubled every one to two years, the source says, that would provide decades of runway. Fiber count could also increase.

Gen 2 has no implementation date in the source. Neither does Gen 3. The roadmap describes capability targets, not committed shipments.

What has Microsoft committed?

At PECC 2026, Microsoft discussed its OCI plans and the need for a common standard.

Its first-wave scale-up optical plan places CPO on the switch and NPO with the accelerator. A second wave places CPO at both ends.

Microsoft names no products, suppliers or deployment volumes for either wave.

What happens next?

At present, the source identifies OCI-MSA as the only optical scale-up standard both published and repeatedly discussed at AI Infra, OCP and PECC by AMD, Broadcom, Meta and Microsoft.

The source forecasts likely fully functional Gen 1 implementations, and perhaps supplier interoperability demos, in 2027. Alduino’s vision is for a broad supplier base to make OCI the Ethernet of optical interconnects.

Other standards or proprietary optical links could still prevail. The commercial test is whether vendors produce fully functional Gen 1 systems in 2027 and then demonstrate interoperability; until then, the 200 Gbps target and roadmap toward more than 100 times Gen 1 bandwidth remain specifications and ambitions.