Co-Packaged Optics Moves the AI Bottleneck Again

The value moves from the faceplate into the package

A conventional high-speed optical connection separates the switch ASIC from a removable optical transceiver at the front of the system. Electrical signals travel across the board to the pluggable module, where DSPs, drivers, lasers, modulators, photodetectors and other components turn those signals into light.

CPO changes that architecture. The optical engine is placed beside the switch silicon, shortening the electrical path and moving optical conversion much closer to the ASIC. Broadcom's Tomahawk 6 Davisson, for example, combines its switch architecture with sixteen 6.4 Tbps optical engines while using field-replaceable external laser modules. [2]

That distinction matters economically. A transition toward CPO can reduce the amount of value captured in a self-contained pluggable transceiver while increasing the strategic importance of components that remain indispensable after the module disappears.

The first is the laser.

Moving the optical engine into the switch package does not mean embedding every laser inside the package. External laser sources allow designers to keep a heat-generating and failure-sensitive component replaceable while feeding light into the integrated photonics system. This is why suppliers with high-power continuous-wave indium phosphide laser technology become particularly important.

Lumentum is a clear example. In March, NVIDIA entered a multiyear agreement with the company that includes a multibillion-dollar purchase commitment, future capacity rights for advanced laser components, and a $2 billion investment intended partly to expand manufacturing capacity. [3] That is stronger evidence of where NVIDIA sees supply-chain scarcity than a simple supplier announcement. Capacity itself is becoming strategic.

Coherent sits in a similar position. It supplies lasers and other optical components while also having capabilities across indium phosphide, silicon photonics and optical assembly. The architectural shift therefore does not necessarily punish every incumbent optical supplier. Companies that can move upstream from complete transceivers into the components CPO still requires may preserve or expand their role.

Packaging becomes part of the networking supply chain

The second shift is toward semiconductor manufacturing and packaging.

Once optical engines sit beside high-value switching silicon, yield, thermal management, fiber attachment and package-level integration matter much more. The optical network is no longer assembled only as a collection of interchangeable modules connected to a finished switch. Part of the network is manufactured with the switch.

That pulls TSMC and advanced packaging capacity deeper into the networking bill of materials. NVIDIA's CPO ecosystem has included TSMC alongside Coherent, Corning, Fabrinet, Foxconn, Lumentum, Sumitomo Electric and specialist optical suppliers. [1] The significance is not simply that these companies participate in photonics. It is that CPO forces several previously separate manufacturing steps to work as one qualified system.

This changes the nature of supplier risk. A merchant transceiver can often be replaced independently of the switch. A photonic engine, fiber interface or packaging process qualified inside a CPO platform is much more tightly connected to that platform's manufacturing yield and reliability.

The same logic raises the importance of optical connectivity suppliers such as Corning and Sumitomo Electric. CPO still has to move light out of the package and into the data center. As electrical traces get shorter, precision fiber coupling, connectors, fiber arrays and optical routing become a larger part of the physical problem rather than disappearing with the transceiver.

Manufacturing specialists such as Fabrinet also deserve attention, although the direction is less automatic. CPO removes some conventional pluggable assembly content, but the underlying optics become more difficult to manufacture and test. The relevant question is whether a supplier can migrate its production expertise from standardized modules toward integrated optical engines, laser assemblies and package-level photonics. CPO rewards capability migration, not simply existing exposure to optical transceivers.

The relative losers are functions, not necessarily companies

The cleanest way to analyze CPO is by function rather than by ticker.

The function most directly pressured is the standalone pluggable transceiver surrounding the switch. DSP content inside those modules is also vulnerable because one purpose of placing optics beside the ASIC is to avoid the electrical losses that require substantial retiming and signal processing in conventional designs.

But that does not translate neatly into a list of companies to short. Broadcom, for example, sells optical DSPs while simultaneously leading CPO development. Coherent and Lumentum participate in conventional pluggables while supplying technologies that remain essential in integrated architectures. The transition can move revenue between product categories inside the same supplier.

Nor will pluggables disappear quickly. CPO offers its strongest economics in extremely high-bandwidth AI fabrics where power density and reliability justify tighter integration. Traditional optical modules remain attractive where serviceability, standardization and deployment flexibility matter more.

The investment question is therefore not whether CPO replaces every transceiver. It is what happens to incremental optical spending as the fastest AI networks reach bandwidths where the traditional architecture becomes increasingly costly.

The next bottleneck sits below the switch

The important CPO beneficiaries are not limited to companies selling switches. The architecture elevates suppliers whose products cannot be removed from the optical path even after the transceiver is removed.

That puts high-power laser capacity from companies such as Lumentum and Coherent near the center of the transition. It increases the importance of TSMC and other advanced packaging capabilities as optics and switching silicon converge. It preserves critical roles for precision fiber and connectivity suppliers such as Corning and Sumitomo Electric. And it creates a manufacturing challenge for companies such as Fabrinet and Foxconn as optical assembly moves closer to semiconductor packaging.

The broader conclusion is that CPO moves the AI networking bottleneck upstream. In the pluggable era, investors could treat the optical module as the relevant unit of supply. In the CPO era, the more useful map starts one layer deeper: laser wafers, photonic engines, package integration, fiber attachment, testing and qualified manufacturing capacity.

As AI clusters become larger, the companies controlling those layers may matter more than the name printed on the transceiver that CPO is designed to replace.

For Altsets research conventions and interpretation limits, see the research methodology.

Sources

1. NVIDIA, "NVIDIA Vera Rubin Ramps Into Full Production to Power Agentic AI Factories Worldwide," May 31, 2026, https://nvidianews.nvidia.com/news/vera-rubin-full-production-agentic-ai-factory

2. Broadcom, "Broadcom Announces Tomahawk 6 Davisson, the Industry's First 102.4-Tbps Ethernet Switch with Co-Packaged Optics," October 8, 2025, https://investors.broadcom.com/news-releases/news-release-details/broadcom-announces-tomahawkr-6-davisson-industrys-first-1024

3. Lumentum, "NVIDIA Announces Strategic Partnership With Lumentum to Develop State-of-the-Art Optics Technology," March 2, 2026, https://investor.lumentum.com/financial-news-releases/news-details/2026/NVIDIA-Announces-Strategic-Partnership-With-Lumentum-to-Develop-State-of-the-Art-Optics-Technology/default.aspx

How to Cite This

For research inquiries or data access: press@altsets.com

Go Deeper