BERLIN – Nvidia is anticipating optical interconnect demand to power switch sales “several orders of magnitude more” than it currently ships today, according to CEO Jensen Huang.

During a press Q&A, SDxCentral asked the Nvidia founder about co-packaged optics (CPO) demand and whether increasing its optics capacity was in line with network architecture shifts towards fewer switches connecting clusters, like through the OpenAI-led Multipath Reliable Connection (MRC) protocol.

Having announced plans to scale optics manufacturing capacity fifteen-fold during its event in Germany, Huang responded by saying that adding scale-up (interconnectivity inside the rack) to the optical domain, the total volume “is going to be, well, 15x.”

“These data centers are getting larger and larger, and the radius are getting larger, and the number of GPUs is increasing tremendously, and so I'm fairly certain the number of switches we need is going to increase, and I'm fairly certain the amount of optics that we need in the future will be several orders of magnitude more than we currently ship today,” he added.

Nvidia’s CPO-infused switch lines have been initially tested inside by neoclouds like CoreWeave and Lambda, with volume production well underway. The Spectrum-6 switches come in a far larger form factor than their traditional cousins, but are capable of supporting a staggering 409.6 Tb/s of bandwidth across 512 ports at 800 Gb/s.

Talking further to the press on the emerging optical portion of its portfolio, Huang struck a vindicated view, saying: “People thought that CPO would take much longer than we ultimately were able to achieve, and so our bit error rate, our data rate will be much higher. Our error rate will be lower, and our quality will be higher. That's the benefit. And the energy consumed will also be a lot lower.”