OCI MSA nails optical architecture, but manufacturing decides if AI scale keeps working

Earlier this spring, AMD, Broadcom, Meta, Microsoft, NVIDIA, and OpenAI formed the Optical Compute Interconnect Multi-Source Agreement (OCI MSA). Their goal: bring coherence to AI infrastructure, and establish a specification for co-packaged optics (CPO) scale-up networks. The consortium settled the architecture debate. They aligned on a slow and wide non-return-to-zero (NRZ) modulation paired with wavelength-division multiplexing (WDM).

In OCI GEN1, the details are concrete: four wavelengths at 50 Gbps per channel, for 200 Gbps per direction per fiber. The roadmap then scales to 1.6 Tbps per fiber per direction. That is the “how the link talks” answer. But the big unresolved question is the “how the link gets built” answer, and it matters because optics scaling is not just a design problem. It is a manufacturing yield and supply-chain physics problem, and it hits cost, reliability, and timelines.

Here is the key logic from the MSA’s architecture choice: slow-and-wide is an energy-per-bit argument, not a marketing vibe. Low symbol rates and simple encoding go together. NRZ carries one bit per symbol. PAM-4 can carry two, but it requires roughly three times the optical power to hit the same bit-error rate (BER). NRZ keeps BER low enough that forward error correction (FEC) stays light, latency stays tight and predictable, and the link stays within its energy budget. On the electrical side, SerDes power per bit at 50 GBaud is roughly one-third that at 100 GBaud. So the architecture is trying to preserve a specific power and latency regime as bandwidth rises.

OCI’s bet is that WDM lets bandwidth scale without “walking back” that regime. Symbol-rate escalation would do the opposite: it pushes the architecture out of the slow-and-wide comfort zone. Wavelength multiplication keeps the per-channel electronics in the same general operating point while increasing system throughput. The MSA roadmap assumes more wavelengths on the same fiber infrastructure, so the basic physical plan is: more lanes, more throughput. No new per-channel design burden in proportion to the total bandwidth.

This is also why first movers already pre-debated the “how many wavelengths” question. OCI’s founding members endorsed four wavelengths as the GEN1 starting point. Ayar Labs has been pushing an eight-to-16 wavelength path for years, and NVIDIA’s published roadmap, “A Roadmap Toward Sub-1 pJ/b Optical Interconnect,” models a 16-wavelength interconnect as the route to reach energy targets the MSA aims at. OCI is effectively standardizing the first step, not the endgame. It treats GEN1 as a practical production starting point that can demonstrate the architecture in silicon and coordinate the supply chain, while leaving the “after four wavelengths” work to the next phase of engineering and manufacturing.

So what does the specification actually do, and where does it stop? The GEN1 spec is four wavelengths at 50 Gbps NRZ, with MSA channel spacing of 400 GHz. It is a minimum multi-wavelength standard, not the maximum. The step-up happens with dense wavelength-division multiplexing (DWDM), where channel spacing tightens at wavelengths of eight or more. System-level bandwidth scales with wavelength count: more lanes, more capacity. In the cleanest version of the story, doubling wavelengths doubles bandwidth without doubling design cost.

The bigger payoff is what cluster size unlocks, not just raw throughput. Larger, flatter, low-latency scale-up domains increase working memory, extend context windows, and add transformer layers. That can translate into deeper reasoning, fewer network stalls, and higher GPU utilization. In other words, wavelength count sets a ceiling for which models the resulting cluster can run in 2028 and beyond. The reader who thinks “that is just networking” is missing the dependency chain: interconnect scaling constrains the model scaling agenda. OCI’s architecture choice is trying to buy runway for that.

Now comes the part the architecture cannot paper over: the wavelength staircase only works if the light source can be manufactured and supplied at scale. The MSA roadmap to 1.6 Tbps per direction per fiber is described as achievable on paper, but “the harder question is which manufacturing approach gets the industry there.” The source breaks the manufacturing challenge into two structural paths.

One path is a shared-laser approach. Multiple lasers are combined and then split to feed a multi-wavelength source. But splitting losses scale with channel count. Every additional output tapped off the network costs the laser optical power it has to make up at the input. That pushes the lasers harder. Drive currents climb, reliability margins erode at every wavelength step, and the economics that can work at four wavelengths do not extend cleanly to eight or 16. The second path is the dedicated-laser approach: one laser per wavelength. That can sound straightforward, but assembly complexity scales with channel count when using discrete optics. A single module supplying 16 wavelengths across eight fibers would imply roughly 128 lasers, 128 fiber alignments, and 128 monitoring photodiodes. That is the supply chain and manufacturing wall in a single sentence: more parts, more alignment steps, more opportunities for yield loss.

Put bluntly for decision-makers: OCI MSA settled the “architecture debate,” but it did not solve the “manufacturing decides the curve” problem. For boards and operators, the non-obvious risk is that timeline slippage may not come from system design or standards bodies, it may come from which light-source architecture survives precision laser array manufacturing in massive quantities. If the supplier’s light-source architecture cannot extend to eight, 16, and beyond without re-engineering, then redesign is already on the calendar, with a two-year delta baked in. In AI infrastructure, a two-year delta is not an inconvenience. It can be the difference between scaling with the next GPU generation or getting stuck paying a “networking tax” at the exact moment demand is highest.