Custom silicon has crossed from design-wins to multi-gigawatt committed programmes -- Broadcom, Marvell, and Qualcomm XPU platforms are creating a $100B+ parallel compute economy alongside NVIDIA's $1 trillion Blackwell+Rubin trajectory

The hyperscaler custom silicon commitment has passed the point of no return. Meta is deploying 1+ GW of custom Broadcom silicon. Amazon's internal chips business is at a $20B+ run rate with $225B in Trainium revenue commitments. Google is considering selling TPU hardware to external customers. Broadcom's AI XPU platform with Apollo and Blackstone will deploy more than 20 GW of compute capacity, with the first trench valued at $35B. AMD secured a 6 GW Meta Instinct deal. This sits alongside, not against, NVIDIA's dominance. NVIDIA expects AI infrastructure spending to reach $3-4 trillion annually by decade-end, with cumulative Blackwell+Rubin revenue visibility approaching $1 trillion and quarterly guidance of $91B. The key insight is that total AI compute demand is so massive that no single supplier can satisfy it. Credo sizes neoclouds at roughly 20% of its revenue, and Broadcom has $6B in purchase orders from unnamed customers. The parallel compute economy creates structural opportunity across the enabling semiconductor stack. Every custom chip requires the same foundry capacity (TSM capex at the high end of $52-56B), advanced packaging (50%+ growth at AMAT, LRCX, KLAC), power delivery (MPWR, ADI), connectivity silicon (CRDO, ALAB, MRVL), and test (TER, KEYS). The PCIe Gen 6 transition is a direct beneficiary as fabric switches, retimers, and signal conditioners sit at the heart of every next-gen AI cluster, custom or merchant.