Broadcom's Optical Communications Playbook: CPO Isn't Just a Module, It's a System Platform Built for Volume

As AI data centers move to 800G, 1.6T and even 3.2T, Broadcom is no longer just a switch ASIC vendor but an industry driver building a complete "CPO platform from light source to packaging". From lasers, silicon photonics and VCSELs to automated manufacturing, Broadcom is steadily turning CPO from concept into reality — and it sums up its stance in one line:

"We're not building the flashiest module; we're building modules that ship in volume, are reliable, and can be serviced."

🔧 Spanning Lasers, Packaging and Modules: Broadcom's Core Capabilities

• Develops and mass-produces 50G, 100G and 200G VCSELs / EMLs.

• Focuses on platform integration of WDM lasers, silicon photonics (SiPh) and Co-Packaged Optics (CPO).

• Flagship platform: the Bailey CPO module, supporting 800G and 512 lanes at just 5.5W, already in commercial deployment. Next up is the Davisson CPO module.

🧠 Network Architecture Breakdown: Front-End, Scale-Out, Scale-Up

Broadcom approaches the problem from the network architecture angle, mapping different modules to three application layers:

Architecture	Applications and Module Focus
Front-End	Traditional Ethernet architecture, CPU → NIC → optical module; bandwidth still low.
Scale-Out	Dense data flows between XPUs; LPO, DSP and CPO all compete, and every picojoule becomes a unit of account.
Scale-Up	GPU-to-GPU links within the same rack, short reach and high speed; still mostly copper today, but gradually shifting to CPO/CPC.

⚡ The Optics-vs-Copper Tipping Point: 200G Is the Key Watershed

• 100G per lane: Multiple options work (DAC, DSP, LPO, CPO).

• 200G per lane: Copper hits a bottleneck:

  • Insertion loss, reflections and crosstalk surge;

  • Reach shrinks to <2m;

  • Retimers add power and weight.

• This is the "sweet spot" where optics replaces electrical, especially in high-density GPU architectures.

🚀 CPO's Value Isn't Just Optics — It's Whole-System Integration

• Shortening the electrical channel and eliminating PCB trace loss is CPO's real advantage.

• Dropping the DSP cuts power and compensation needs, improving system stability and thermal efficiency.

• Broadcom's Bailey module has been validated in volume production with BER and thermal stability at ASIC grade.

🔍 Scaling Into the Future: XPU + CPO + BiDi + Distributed AI Clusters

• Optical engines won't just live in switches; they will move onto GPU/XPU substrates:

  • A single die outputs 6.4T / 12.8T, forming the core of a 51.2T cluster.

• BiDi (bidirectional over a single fiber)* will halve fiber cost and double cabling efficiency.

• Distributed optical AI architectures emerge, with 9.6T per GPU → a single switch tier can directly connect hundreds of GPUs.

🏗 Manufacturing and Reliability: Not a Prototype, but a Product That Ships

• CPO packaging requires HBM-grade precision and thermal control.

• Broadcom has built automated volume production lines and introduced a link reliability data-reporting mechanism.

• Reliability targets are aligned with ASIC production quality requirements, making optics a system-level building block rather than a replaceable consumable.

💬 Key Takeaways from the Q&A

• 200G is copper's limit — and the starting point where optical modules take over.

• The efficiency gap between LPO and CPO comes from packaging and channel design, not from the optical components themselves.

• CPO won't be confined to switches; it will become the core interconnect choice for GPU and XPU architectures.

📌 Conclusion: The Endgame for Optical Modules Isn't Faster — It's More "System"

Broadcom's strategy centers not on the speed of light but on system stability and manufacturing feasibility. CPO is a platform, not a single module. What will truly solve AI clusters' power, thermal, bandwidth and deployment-efficiency problems isn't showmanship but "solutions you can actually build," like Bailey.