Ken Kuang, President and Founder, Torrey Hills Technologies, LLC

Better Chips for AI: Why We Need a Mix of Materials in 2026

The rapid use of AI is forcing engineers to totally rethink how we build chips. The modern design is no longer just about the silicon chip; it's about the complex package where electrical signals and light-based optical signals meet. New technologies like Co-Packaged Optics (CPO) and high-speed pluggable modules (400G/800G) create a critical challenge: extreme heat.

The AI chips and the sensitive lasers in the optical parts all get hot, causing different parts of the assembly to expand at different rates. This difference is called the Coefficient of Thermal Expansion (CTE). When the CTEs don't match up, the materials push against each other, creating huge stress that is the biggest reason why systems fail.

To solve this, our focus for 2026 is a Heterogeneous Material Strategy, moving past simple materials like copper and aluminum, which simply can't match the low CTE of silicon. We must use advanced refractory metal composites, specifically copper tungsten (CuW) and copper molybdenum (CuMo). We will deploy CuW as the rigid, high-strength baseplate or heat spreader in CPO assemblies . Its ability to precisely match the silicon’s low CTE ensures that expensive optical and electrical components stay perfectly aligned and stable over many heating and cooling cycles.

Complementing this structural base is CuMo, which is ideal for the heat sink lids and smaller thermal components within the optical modules (like QSFP-DD and OSFP). CuMo offers a superior balance of thermal conductivity (good heat escape) and lower density, making it easier to machine into the intricate shapes needed for compact pluggable modules.

By using the specialized strengths of CuW for structural integrity and CuMo for balanced heat dissipation, we can effectively manage the competing demands of the AI interconnect ecosystem. This tailored, dual-material approach is the critical foundation for ensuring the long-term reliability and scalability of the world's AI infrastructure.