The AI boom’s least glamorous winner just filed for a $5bn listing, and it wires Google and Amazon’s data centres

The market is still pricing AI as a GPU story, but the more immediate earnings sensitivity may be in the networking stack: transceivers, optical modules, and the vendors that relieve rack-level bandwidth bottlenecks. That shifts the near-term winners toward names like COHR, LITE, CIEN, and potentially component-heavy ODM/supply-chain exposures, while NVDA remains a beneficiary only indirectly through sustained hyperscaler capex. The second-order effect is that interconnect content per deployed GPU cluster keeps rising, so revenue growth can outpace the headline compute cycle even if GPU shipments merely normalize.

The key risk is that capital raises and capacity expansions can quickly destroy pricing power. If Chinese and global optical suppliers all expand into the same bandwidth bands, margins can compress faster than revenue grows, especially over a 6-18 month horizon as new fabs and packaging lines come online. For NVDA, this is less about demand destruction and more about mix: stronger networking spend validates AI infrastructure budgets, but it also signals that customers are trying to lower the total cost of ownership, which is a mild long-term check on GPU pricing power.

The contrarian view is that the consensus may be underestimating how much of AI capex migrates from “compute” to “connect.” That argues for a relative-value long in the optical ecosystem versus pure-play AI compute, not a blanket bullish bet on the whole semiconductor complex. The thesis would be falsified if hyperscaler capex guides roll over, or if transceiver pricing weakens before volume growth shows up in reported margins; that would indicate the cycle is more saturated than the market is assuming.