Co-packaged optics was a topic dominating discussions at last week’s OFC, the optical conference and exhibition held in Los Angeles.
But for Broadcom’s Near Margalit, President and General Manager of the Optical Systems Division at Broadcom, the most important announcement wasn’t co-packaged optics but the newly announced Open Compute Interconnect Multi-Source Agreement (OCI MSA).
“For me it is probably the most important thing I’ve done at Broadcom, actually in my whole career,” says Margalit.
The six OCI MSA founding members are AMD, Broadcom, Nvidia, Meta, Microsoft, and OpenAI. Their aim is to make optics rather than copper the preferred interconnect approach for AI scale-up networks by building an industry consensus and moving away from proprietary designs.
“It is the first MSA in the industry that is intended to remove the high-speed SerDes [serialiser-deserialiser] that exists in most [switch or AI accelerator] products today,” says Margalit. “By eliminating the SerDes, you can greatly reduce the power consumption.”
The OCI MSA is a specification moving optical interconnects to a multi-wavelength-per-fibre framework: four optical wavelengths in one direction and four different wavelengths in the other.
Meta data
Margalit is also buoyed by Meta’s latest reliability analysis of Broadcom’s 51.2 terabit co-packaged optics-enabled ‘Bailly’ switches.
Meta tested the switch platforms for the runtime equivalent of 90 million optical transceiver hours.
The first phase, involving 70 systems, showed some minor non-fatal issues, for example pluggable lasers having power issues, yet they still outperformed optical module-based switches. “It is still 10 times better than optical transceivers,” says Margalit.
A second test phase involved 300 systems and 50 million runtime hours. In this phase, no failures occurred. “So maybe 100x better than the mean-time between failures of optical transceivers,” says Margalit, who is surprised that these early Broadcom co-packaged optics switch systems perform so well.
He attributes co-packaged optics’ superior results to a more uniform supply chain, system-level testing rather than component-level validation, and a manufacturing flow closer to semiconductor processes with fewer manual steps.
Those results also explain why co-packaged optics is now approaching commercial relevance. “On the cusp of hitting volume,” says Margalit, who is careful not to overstate the timeline.
His expectation is that co-packaged optics will account for perhaps 5 to 10 per cent of switch shipments in 2027. That suggests a gradual ramp rather than a sudden inflection, even though he believes companies such as Nvidia may push adoption more aggressively.
Open Compute Interconnect MSA
The more interesting question for Margalit is not when co-packaged optics reaches scale, but what comes after it. The core issue, he says, is that today’s optical interconnects are still constrained by their electrical origins.
High-speed optical links typically need the SerDes to drive data across copper before converting it to light. That architecture carries a power penalty.
By removing the SerDes electrical overhead, Broadcom believes optical links can cross a new power threshold. “Our switches and XPUs are going to be lower power with optics than with copper,” he says.
If that claim holds, it would represent a fundamental shift in interconnects. Optical links would be justified by power efficiency compared to electrical alternatives, not just by reach or bandwidth.
This is why Margalit views the MSA as so significant. “This is the future for co-packaged optics; it’s more than co-packaged optics,” he says.
“You could think of what we’ve done with co-packaged optics as an enabling technology. The OCI MSA is kind of the North Star.”
Margalit describes last week’s OFC as a contest among many competing approaches, ranging from traditional high-speed serial links to wide, parallel optics.
“There is a lot of positioning, a lot of people trying to establish their dominance, whether people pushing LEDs, VCSELs, silicon photonics, or the new XPO MSA,” he says.
XPO refers to a new 12.8-terabit high-capacity pluggable form factor for optical networking. XPO is a useful but incremental development, he says, particularly for scale-out architectures.
At the same time, he noted renewed interest in near-packaged optics (NPO). He sees this as evidence of lingering concerns about ecosystem control.
“The one thing that really bothers people with co-packaged optics is the lack of multi-vendor optics,” he says. Near-packaged optics offer certain benefits of integration while preserving supplier diversity.
Optics on the rise
Broadcom is determined to stay focused on its own roadmap. “We have a really clean roadmap for deep integration, we want to minimise any distraction.” That means optics to keep moving closer to compute chips over time. “The long-term vision is that fibre goes directly to the GPUs,” he says.
Margalit is cautious on the topic of optical circuit switches, despite growing interest and start-up activity. While acknowledging deployments at hyperscalers, Broadcom has yet to see a compelling role for the technology. “It doesn’t seem really impactful,” he says. “I wouldn’t invest in that right now.”
The key question it seems is no longer how much optics will be deployed but how it will be integrated; not a bolt-on to copper but an inherent part of the system itself.