NVIDIA just dropped optical-integrated switches: Spectrum-X and Quantum-X.

Here's the upgrade path they're betting on:

Pluggable optics → LPO → Silicon photonics → CPO (co-packaged optics).

Right now everyone's still on pluggable modules. Signal travels from ASIC → PCB → SerDes → DSP → optical module. Works fine, very standardized. But at 800G/1.6T? Power, latency, signal loss, heat… all becoming a nightmare.

LPO is the band-aid. Still pluggable, but strips out DSP to cut power and cost. Good for short distances in hyperscale data centers. Downside: needs tight coordination between switch, SerDes, and module.

Silicon photonics is the real device-level upgrade. Integrates optical functions onto silicon chips, boosts density and bandwidth. Doesn't kill lasers — actually makes high-end lasers, coupling, and advanced packaging way more critical.

CPO is the architecture shift. Puts optical engine right next to the ASIC, shortens electrical paths dramatically. Lower power, lower latency, higher density, fewer failure points. Trade-off: packaging, testing, cooling, and repairs get way harder.

What NVIDIA is really saying: AI cluster bottlenecks aren't about "buying more optics" anymore. It's about rethinking the whole stack — switches + photonics + packaging + system integration.

Value is shifting from pluggable modules to:

- Switch ASICs

- Silicon photonics

- Optical engines

- External lasers

- Advanced packaging

- System integrators

Who benefits:

$II $LITE (Coherent, Lumentum): Lasers, photonics, CPO engines

$FN (Fabrinet): High-end optical manufacturing

$GLW (Corning), Sumitomo, Senko: Fiber, connectors, high-density cabling

$TSM $ASX (TSMC, ASE): Silicon photonics fab, advanced packaging, heterogeneous integration

$AVGO $MRVL (Broadcom, Marvell): High-speed switch ASICs, SerDes, DSP, AI networking chips

Basically: the money's moving upstream. 💡