Co-Packaged Optics Goes to Volume: The Supply Chain That Has to Be Built by 2028

For three years, co-packaged optics lived in an uncomfortable middle ground: technically proven, architecturally necessary for next-generation AI switches, but not yet in the supply chain at a scale anyone could build a business around. That phase ended in 2026. Broadcom has been shipping its 102.4-terabit Tomahawk 6 Davisson CPO switch since last October, Nvidia put Spectrum-X Ethernet Photonics into mass production in August with CoreWeave, Lambda and Oracle as launch customers, and TSMC confirmed on its July earnings call that its COUPE photonic platform has begun volume production. The technology question is settled. The question that remains is whether the component supply chain underneath it can be built fast enough to meet the 2028 scale-up deployments that the biggest buyers are now planning around.

The scale of the ramp matters. Nvidia's flagship Spectrum-X Photonics configuration packs 409.6 terabits per second into a single chassis, with 512 ports of 800 gigabit Ethernet, and the company says it uses roughly 75 percent fewer lasers and 80 percent less power than an equivalent pluggable-based system. Broadcom's Davisson uses sixteen 6.4-terabit optical engines on TSMC's COUPE process at 200 gigabits per lane, with a 400-gigabit successor in development. Every one of those products draws on the same three constrained inputs: high-power lasers, photonic wafer starts, and advanced packaging capacity. None of the three is where it needs to be.