AI Infrastructure Bottleneck Migrates From Compute to Optics and Memory

A cluster of research and conference commentary points to a decisive shift in where scarcity — and therefore economic rent — sits within the AI supply chain. Optical component shortages are described as widening rather than easing, positioning Lumentum and contract manufacturer Fabrinet as structural beneficiaries of hyperscaler cluster scale-out, where interconnect bandwidth increasingly gates cluster performance. In memory, SK hynix retains leadership in high-bandwidth memory, an oligopolistic market where supply is contracted well in advance and pricing has decoupled from the commodity DRAM cycle. Thermal management is a third emerging chokepoint, with Asia Vital Components reporting into strengthening liquid-cooling demand. On the demand side, neocloud operators such as Nebius are compounding AI cloud revenue at rates that validate the capex cycle. The counterweight is valuation discipline: independent analysis argues the math of the AI trade is changing, as depreciation schedules, financing costs and the rising sovereign discount rate erode the return-on-invested-capital case for buildouts funded at yields materially above post-crisis norms. The strategic implication for allocators is a rotation toward suppliers with contracted backlogs and pricing power, and away from balance-sheet-intensive capacity owners exposed to the cost of capital.