How hyperscaler AI capex splits between merchant GPUs, custom silicon, and interconnect One 2026 hyperscaler capex pool feeds three buy-paths - merchant GPU, custom XPU, optical interconnect - each with different terminal economics. US cloud/AI capex pool 2026 ≈ $660-700B guided · ~2x 2025 Merchant GPU NVDA direct path CUDA stack GPU die Blackwell · Rubin Merchant GPUs richest margin per capex $ risk rides NVDA roadmap demand booked years out Custom XPU Broadcom · Marvell custom dies Custom XPU hyperscaler cost control design-win-cyclical Optical interconnect Marvell path PAM4 DSP 60-65% DSP share pluggable module Interconnect capacity-constrained scales with cluster size not tied to one chip winner rich margin cost control capacity-bound