TSMC’s COUPE: Bringing Optical Connections Closer to AI Chips

As AI systems grow, moving data between chips at higher speeds requires increasingly power-hungry electrical connections. Optical communication uses light to carry data, but it still needs electronics to control transmission and process incoming signals. TSMC’s COUPE platform brings these electronic and optical components together, reducing the distance electrical signals must travel and the power needed to connect them.

COUPE stands for Compact Universal Photonic Engine. It integrates an electronic integrated circuit, or EIC, with a photonic integrated circuit, or PIC. The electronics drive and receive signals, while the photonic components manipulate and detect the light carrying data. Three-dimensional integration shortens the connection between them, creating a compact optical engine.

This addresses one part of a broader networking challenge. In conventional equipment, removable optical modules sit at the edge of a circuit board. Signals travel electrically between the main networking chip and those modules before continuing through optical fiber. As speeds increase, maintaining signal quality along these electrical paths consumes more power.

Co-packaged optics, or CPO, brings optical engines onto the same package as the networking chip, shortening those paths. COUPE supplies the integrated electronic and photonic building block for this arrangement. The distinction matters: COUPE is TSMC’s optical-engine platform, while CPO describes how optical engines and the main chip are packaged together.

TSMC has announced that its COUPE-on-substrate CPO solution will begin production in 2026. It has reported twice the power efficiency and one-tenth the latency of a board-mounted pluggable implementation. Those comparisons apply to the connections being measured, rather than the electricity consumption of an entire data center.

Nvidia and Broadcom have announced products using COUPE. Nvidia’s Quantum-X Photonics and Spectrum-X Ethernet Photonics switches use optical engines to connect AI systems. Broadcom’s Tomahawk 6–Davisson combines 16 COUPE-based optical engines and claims a 70% reduction in optical-interconnect power consumption compared with traditional pluggable solutions. Their adoption gives TSMC a role in the networking equipment supporting AI expansion.

Turning these designs into reliable products requires more than integrating two chips. Manufacturers must supply lasers, attach fibers precisely, control heat, and test the completed assemblies. At Semicon Taiwan, TSMC advanced packaging executive K.C. Hsu emphasized the coordination needed across this supply chain as companies prepare CPO for mass production. The immediate task is making these connections practical for widespread data center deployment.

Samsung and Intel are developing alternatives. Samsung is building a silicon photonics foundry platform, while Intel has demonstrated an optical communication chiplet packaged with a processor. Competition will depend on manufacturing reliability and cost as well as technical performance.

COUPE’s longer-term potential extends beyond networking switches. Bringing optical connections closer to processors could enable larger groups of AI accelerators to exchange data while limiting the power required to connect them. For customers, the benefit would be the ability to expand computing capacity without an equally steep increase in communication power consumption. For TSMC, COUPE extends its role beyond making processors and packaging them with memory. It provides a manufacturing foundation for the optical connections that increasingly powerful AI systems will need.