Everyone is talking about AI. Everyone is talking about GPUs. But I believe the next major bottleneck won’t be the processors themselves. It will be the infrastructure that connects them. As AI clusters continue to scale from thousands to tens of thousands of GPUs, moving data efficiently has become just as important as generating compute. This is why Co-Packaged Optics (CPO) has rapidly moved from a research topic to one of the semiconductor industry’s most strategic technologies.
However, the real question is no longer whether CPO works. The real question is whether the industry can manufacture it economically at hyperscale. History has shown that many promising semiconductor technologies do not succeed or fail because of performance alone. They succeed—or fail—because of manufacturing yield, supply-chain maturity, packaging capability, reliability, and field serviceability. In my opinion, CPO is entering exactly this phase.
The Supply Chain Is Becoming the Competitive Advantage
Much of the recent attention has focused on companies such as NVIDIA and Broadcom, but they represent only one part of a much larger ecosystem. The value chain behind CPO spans multiple highly specialized technologies:
| Value Chain | Key Players |
|---|---|
| AI Accelerators & Switch ASICs | NVIDIA, Broadcom, Marvell |
| Silicon Photonics | Intel, Ayar Labs, Lightmatter, Celestial AI |
| Lasers & Optical Components | Coherent, Lumentum |
| Advanced Packaging | TSMC, ASE, Amkor |
| Fiber & Connectivity | Corning, Furukawa Electric |
| System Integration | Foxconn, Quanta, Wistron |
| End Customers | Microsoft, Google, Meta, Amazon |
No single company can deliver CPO alone. Success depends on the maturity and coordination of the entire ecosystem.
Where I See the Biggest Challenges
From my perspective, the greatest risks are not optical bandwidth or photonic performance. They are manufacturing and operational challenges.
Advanced packaging remains one of the industry’s most constrained resources. CPO requires heterogeneous integration of switch ASICs, silicon photonics, lasers, and fiber interfaces with micron-level precision.
Thermal management is another critical hurdle. High-performance AI switch ASICs can consume well over one kilowatt of power, while optical components require stable operating temperatures to maintain performance and reliability.
Manufacturing yield becomes increasingly important because a defect in a single optical engine can affect an entire high-value assembly.
Finally, serviceability cannot be overlooked. Traditional pluggable optics allow failed modules to be replaced individually. With CPO, operators must rethink maintenance strategies when optical components become integrated with the switching platform.
These are engineering challenges—but they are also supply-chain challenges.
Looking Beyond the Headlines
AI chips and GPU demand dominate today’s headlines. Tomorrow’s competitive advantage may belong to the companies enabling the infrastructure behind them. The winners may not only be GPU designers. They may also include companies solving advanced packaging, silicon photonics, laser integration, optical testing, and large-scale manufacturing. CPO has already demonstrated what is technically possible. Now comes the harder challenge: industrialization.
Which companies will transform technical innovation into scalable manufacturing, resilient supply chains, and sustainable business models?
As AI infrastructure enters its next phase, I believe this question may matter more than who designs the next generation of chips.
