Bandwidth is per-lane speed times lane count. Push speed and you hit the SerDes energy floor; push lane count and you hit fiber and connector cost. The new fact this year: the second wall arrived first.

Google reviews five TPU generations: an architecture set in 2017 survived the Transformer takeover and scaled 3600x. For optical readers, the key line: OCS, one of only two designs still unique to TPU, is what keeps a near-10K-chip machine alive, schedulable and deployable in stages. Power has now replaced cost as the new ceiling.

OCS has left the lab, but today it is less a market than a customer list with one name on it. Cignal AI's ECOC 2026 numbers: ~95% of OCS ports go into Google's TPU network, and whether GPU scale-up becomes the next killer app depends on CPO going multi-wavelength — an 8x port-count gap that will split the supply chain.

iPronics turns silicon photonics OCS architecture selection into a computable graph-theory benchmark. Verdict: Double Dilated Benes is today's practical 128×128 design, while Dilated Banyan performs best but hits the reticle limit. The question is no longer whether SiPh OCS can scale, but which architecture to pick.