Optical Scale-Up Architecture Is Settled, Manufacturing Is the Bottleneck

The Optical Compute Interconnect Multi-Source Agreement (OCI MSA), formed by AMD, Broadcom, Meta, Microsoft, Nvidia and OpenAI, has settled the architectural debate over co-packaged optics (CPO) scale-up networks for AI infrastructure. The consortium converged on a slow-and-wide NRZ modulation combined with wavelength-division multiplexing (WDM); the GEN1 spec delivers 200 Gb/s per fiber direction using four wavelengths at 50 Gb/s each, with a roadmap scaling to 1.6 Tb/s. This choice is driven by energy-per-bit economics: NRZ requires far less optical power than PAM-4 and keeps forward error correction light.

While the architecture is settled, how bandwidth continues scaling remains open. The answer lies in multiplying wavelengths rather than pushing symbol rates higher, letting system bandwidth multiply while per-channel electronics stay unchanged. But the real bottleneck to this strategy isn't architectural anymore — it's manufacturing precision laser arrays at industrial volume.

The piece traces three eras of photonic integration: discrete optical assembly, silicon photonics (which excluded lasers), and now heterogeneous integration — combining III-V gain material with silicon on a single wafer process. This mirrors the transformation CMOS underwent in electronics. Both competing manufacturing paths — shared-laser and dedicated-laser — hit reliability and cost walls as wavelength counts rise, making the choice of laser-source architecture the decisive factor going forward.

For engineers, the stakes are concrete: today's wavelength-scaling decisions determine what cluster sizes — and therefore what model scales — will be achievable by 2028 and beyond. Whether a supplier's laser architecture can extend to 8, 16 wavelengths and beyond without redesign is becoming a critical criterion for system architects.