AI Stack — The Silicon Railroad

Pricing power concentrates at the un-substitutable toll booths — ASML EUV, TSMC leading-edge + CoWoS, Nvidia CUDA/NVLink, Broadcom/Marvell merchant fabric, the Cadence/Synopsys EDA duopoly, the SK Hynix/Micron/Samsung HBM oligopoly — and at the scarce physical inputs, firm nuclear baseload and the sold-out turbine/transformer/switchgear/CDU supply chain, plus the hyperscaler-distributors (Microsoft, Alphabet, Amazon) who bundle AI into seats customers already pay for. Value is competed away in the commoditizing middle: raw model calls and open-weight-compressed frontier pricing, merchant GPU rental, commodity DRAM/NAND/HDD, building shells, and any app without proprietary data, distribution, or workflow lock-in. Pre-revenue option plays (SMR, fusion, uranium developers, thin agents) carry theme exposure but no present pricing power.

Firm energized power + multi-year grid interconnection — the stack's true binding constraint; new baseload can't be plugged in faster than queues, NRC licensing, and enrichment allow, so incumbents with interconnected assets price at scarcity.

Nvidia (NVDA) — owns the accelerator layer via the CUDA + rack-scale NVLink moat, taxing every training and inference run regardless of which lab or cloud wins.