The 2023-2025 AI compute cycle is the largest hardware capex buildout in technology history measured in absolute dollars. Hyperscaler capex — Microsoft, Alphabet, Meta, Amazon Web Services, and Oracle, plus the next tier of OpenAI / Anthropic-aligned compute and the sovereign AI programs — aggregated to approximately $200-260 billion in calendar 2024 (industry analyst estimates triangulated from public hyperscaler disclosures). A substantial fraction was allocated to Nvidia GPUs and the surrounding compute infrastructure.

The unit economics of the cycle are determined by a small number of products: the Nvidia H100, the H200, and the B100 / B200. Each generation of these accelerators sits at the intersection of the constraints covered in earlier lessons: HBM supply (sc1_l5), advanced packaging (sc1_l12), TSMC leading-edge capacity (sc1_l4), and the export control regime that governs which customers can buy them (sc1_l8). This lesson is the unit economics framework: per-GPU cost, training run economics, hyperscaler capex flow, and the central question — is the buildout demand-constrained or supply-constrained, and what does each scenario imply for the cycle's durability.