Ben Bajarin
CEO & Principal Analyst of Creative Strategies, Semiconductors-to-Software Research
Optical networking's broader rollout will be slow because scale-up and scale-out require orders of magnitude more connections than today's limited optical distance links. Supply-chain discussions echo that slow pace. NPO/CPO starts in switches and reaches XPUs and GPUs much later, with new challenges. Optical backplanes will be necessary as racks move to optical, though copper will remain in many implementations. x.com x.com x.com x.comÂ
Low initial attach will limit optical networking's near-packaged optics ramp even when it scales in 2028. OCP confirms that timing, with co-packaged optics later still. In response to a claim of 2027 CPO scale-out shipments, those would be very small and for switches, not compute; CPO scale is a 2028-29 or later story. x.com x.comÂ
Higher optical-fiber use per GPU could expand demand for Applied Optoelectronics. Corning said fully optical scale-up would raise fiber count from about 16 to 160 per GPU; that 10x case is hypothetical for now, with a mix of technologies likely near term, but it shows the direction of travel. x.com x.comÂ
Pluggable optics have a new scaling path alongside CPO and NPO, supporting demand for light from Lumentum, Coherent and Applied Optoelectronics. Nvidia and Cisco's planned USFP packs 16 full-duplex 400G/lane connections into an OSFP-sized footprint, targeting four times the bandwidth density of 200G/lane modules while retaining replaceability. Volume-production timing and adoption remain unknown. x.com x.comÂ
Co-packaged optics standards could make finished transceiver modules smaller and more interchangeable as chip designers build optical engines inside their packages. That is another reason to avoid Lumentum, Coherent and Applied Optoelectronics and prefer Soitec. Lumentum and Coherent now earn a lot from finished modules, leaving them exposed if the architecture changes. x.comÂ
Scale-up moving to optics adds demand for optical suppliers even if new AI architectures need fewer switches. Bigger data centers and more GPUs add to that need; Jensen Huang expects future optics requirements to be several orders of magnitude above what NVIDIA ships today. This challenges the optics bear case for Lumentum, Coherent, Tower Semiconductor and Credo. x.comÂ
Networking is the next AI hardware trade after memory, whose share of rack value rose from 15% to 35%. Optics is moving inside the server, as Nvidia's Feynman architecture builds in co-packaged optics and silicon photonics. Scale-across networking could grow 30% a year just linking existing data centers, as inference moves closer to users and training spreads across sites. Supply is tight: Lumentum's CEO said it can meet only 30% of expected 2027 demand and sees optics tighter than memory. x.comÂ
Applied Optoelectronics has a better chance than Coherent of delivering a 300-400 mW laser with tolerable RIN and linewidth. Its transceiver capacity ramp is also larger and is not gated by Sumitomo's 6-inch InP wafers. Long AAOI against short COHR is an unnecessarily dangerous pair trade idea. x.comÂ
More lasers per module and new 400G DSPs could keep the pluggable-optics runway open for years, supporting Lumentum, Coherent, Marvell, Applied Optoelectronics and Sivers even as co-packaged optics approaches. NVIDIA and Cisco co-designed a module with four times the bandwidth in the same slot; fewer modules per GPU look bearish, but the added optical content points the other way. x.comÂ
Coherent's 6-inch substrate access, yields and margin concerns look mostly like timing issues in the optical buildout. Coherent is now one of the preferred ways to gain exposure and should participate meaningfully, despite those constraints. x.comÂ
Lumentum and Coherent face a threat from microLED optics, which could replace laser optics with 20x less power and no temperature sensitivity by late 2027. The use case is short-reach, high-volume links under 30 meters. The copper versus optics debate is a distraction, since every AI cluster needs both and Credo now does both while Applied Optoelectronics handles long-haul; the bigger question is what microLED does to laser makers. x.com x.comÂ
Copper scale-up is dead, making optics hyper bullish, especially Lumentum. Nvidia's Rubin NVLink Nyquist cut from 56 to 47 GHz signals failure: reflections matter, and bidirectional links make them worse. Feynman will have to use 16-wavelength DWDM ring-modulator CPO at roughly 50-64G NRZ. x.com x.comÂ
