Premium: Breaking AI's Compute Limits — Decoding NVIDIA's and Broadcom's Two CPO Paths and a Full Optical Component Volume Model
Executive Takeaway
The key takeaway of this in-depth report: the ultimate CPO battle — a sweeping inflection across technology, architecture, and supply chain
Why CPO Is Inevitable, and What Drives It
The bottleneck has shifted: In the AI compute race, the bottleneck has moved from “compute” to “transmission.” Traditional copper has hit its physical limit at 224G per lane, causing signal attenuation (36 dB) and a surge in power consumption (over 30% of switch power).
CPO's core advantage: Co-packaged optics (CPO) co-packages the optical engine with the ASIC, minimizing transmission distance and saving 30% to 50% in power (NVIDIA/Broadcom data show roughly 3.5x power savings) — the only way forward in the “post-1.6T era.”
The Two Giants' Architectural Showdown and Technology Trends
NVIDIA (Scale-Up Route):
Goal: Build a “single giant virtual GPU” (NVL72/NVL576), pursuing ultra-low latency (< 20 ns).
Technology choice: A firm backer of MRM (micro-ring modulators), whose tiny footprint meets the needs of ultra-high-density interconnect (despite extreme thermal sensitivity, which must be solved through advanced packaging).
Broadcom (Scale-Out Route):
Goal: Build a standardized Ethernet platform, optimizing for cost efficiency ($/Gbps).
Technology shift: Moving from the earlier, stable MZM to MRM, because MZM's large footprint can't deliver the edge density required by next-gen 102.4T switches. Physical size is the final arbiter.
Future challenger: Thin-film lithium niobate (TFLN), with ultra-high bandwidth and ultra-low drive voltage, could enable 200G/400G per lane in the future — and may even eliminate the power-hungry DSP.
Why External Laser Sources (ELSFP) Are Critical
Decoupling the heat source: Because switch ASIC power is extremely high (>1000W), laser chips integrated alongside would see their lifetime shortened drastically by high temperatures (>70°C). The industry consensus is therefore to externalize the laser (ELSFP).
Demanding specs: The ELSFP module becomes a pure “power source” and must deliver:
Extreme optical power (UHP/SHP 23-26 dBm): It must feed multiple optical engines simultaneously.
Polarization control (PER ≥16 dB): Requires expensive polarization-maintaining fiber (PMF).
Thermal management: High power classes (UHP/SHP) will make liquid cooling standard.
102.4T volume estimate: A single 102.4T switch requires roughly 16 optical engines, 512 modulators, 1,024 fibers, and 16 ELSFP modules (128 high-power laser chips in total).
A Redistribution of Power in the Supply Chain
Crisis and transformation for traditional module makers: CPO shifts the core of optical engine manufacturing to advanced packaging at foundries (TSMC, Intel) and OSATs (ASE). Traditional module makers must pivot to focus on ELSFP module supply and fiber cabling solutions.
Emerging winners and dividends: The most certain growth areas over the next 3 to 5 years are:
ELSFP external laser source suppliers.
Advanced packaging technology (die bonders, active alignment equipment and services).
Specialty fiber and connectors (MPO connectors, polarization-maintaining fiber PMF).
New scale-up players: Seven startups including Ayar Labs, Celestial AI, and Lightmatter are tackling chip-to-chip (scale-up) bandwidth and memory disaggregation bottlenecks through optical chiplets and optical interposers, with volume production expected in 2025-2026.
Comments