Lightmatter has sampled its Passage co-packaged optics (CPO) chiplet with Qualcomm Technologies’ 112G PAM4 optical SerDes chiplet, and the companies say the combination hit 1.6 Tbps of throughput per fiber for AI optical interconnect.

The reported 1.6 Tbps/fiber result comes from a 16-wavelength DWDM (dense wavelength division multiplexing) design running at 112G per SerDes lane. Lightmatter says the architecture delivers up to 8X more bandwidth per fiber than existing near-packaged optics (NPO) and CPO solutions, targeting the bandwidth-density limits that show up fast when AI clusters scale and link counts explode.

For data center engineers, the practical hook is straightforward: if throughput per fiber goes up, you can move the same bisection bandwidth with fewer fibers. That can ease physical routing pressure, reduce cable bulk at the rack and row level, and potentially trim optics-related capex. But the big system question is power at scale—CPO is often sold on pJ/bit improvements, and those gains only matter if the packaging, thermals, and serviceability don’t create new operational pain.

Qualcomm Technologies Executive Vice President and General Manager Tony Pialis called the validation “a critical step forward for the industry,” adding that the work with Lightmatter provides “a clear path to 100 Tbps of total I/O per package.” Lightmatter Founder and CEO Nick Harris said the milestone “proves that our photonic engines are delivering the massive throughput and power efficiency that hyperscale data centers demand today.” Analyst Vlad Kozlov, Founder and CEO of LightCounting, pointed to fiber cabling and physical space constraints, and said that “moving toward a 100 Tbps per-package I/O threshold is a key milestone for the ecosystem.”

Lightmatter said the Passage L-Series photonic interconnects are available through Passage Evaluation Kits for lead customer testing. More details are available at lightmatter.co.

Source: Lightmatter