
CPO hasn't shipped yet, but NPO brings maintainability back first
Last week I wrote a field report on CPO, and the conclusion was that the power consumption looks good, but it's a pain to repair when it breaks, and it's still far from being usable by ordinary people. This week I went through the specs of that NPO near-packaged optical engine from Corephotonics, and my thinking has shifted a bit.
NPO can be seen as a compromise version of CPO. The optical engine isn't packaged into the chip substrate; it sits on the side of the chip, and the electrical signal only travels a few centimeters before being converted to light. Some solutions compress that electrical interconnect from about one meter down to five centimeters, and latency and power consumption drop accordingly, but the optical engine itself is still modular and can be swapped and pulled. That point matters more than anything. Last time I stood in the server room for forty minutes, and the worst part was that feeling of one channel failing and the whole board going down with it; the fact that the fiber leads out from the package edge and looks ugly was secondary. CPO gives up maintainability; NPO doesn't.
The current specs can also deliver. The 3.2T NPO optical engine integrates 32 100G channels in a single unit, uses linear direct drive, and has no DSP. The official rate capacity per unit area is 11.4 times that of a traditional 800G module. I believe half of that number, because the area is calculated based on the package, not the whole unit. This year Alibaba Cloud exhibited a 3.2T NPO module based on the OIF standard, and Accelink showed off the first 3.2T silicon photonic single-mode NPO module at OFC 2026, and it's said to have completed full-system verification at a leading domestic CSP.
Going from the lab to engineering deployment—this step has really been taken.
But between verification running through and mass production delivery, there's still yield and on-site maintenance in the way. What's most valuable about NPO right now is that it lets the people on the night shift still get a full night's sleep. As for whether this route can survive, in the end it still depends on who fixes it when it breaks.
Physix Frontier