NVIDIA AI Cities Expand Shared Infrastructure Globally - The Cryptonomist

NVIDIA's shift from selling GPUs to selling "AI cities" is fundamentally a pricing-power grab on the entire compute stack. Regionally pooled architectures push HBM bandwidth demand to a new order of magnitude—SK Hynix's HBM3E capacity will likely be fully allocated by end-2025, compressing Samsung and Micron's catch-up runway. NVLink and InfiniBand form an ecosystem moat: once compute is pooled at the city level, migration costs scale exponentially, elevating lock-in from chip to architecture. On compliance, "global expansion" is really a segmentation play. Export controls have split the market in two, forcing NVIDIA to build regulatory enclaves across the EU, Middle East, and Southeast Asia. TSMC in Taiwan, China remains a single-point supply risk. Energy cost is now overtaking silicon as the primary siting variable, redrawing the global compute map. Competitively, AMD's MI300X has narrowed the per-die gap, but the ROCm-versus-CUDA software chasm persists. The existential threat is hyperscaler in-house silicon—AWS Trainium and Google TPU are steadily eroding NVIDIA's inference share. Eighteen-month call: compute commoditization accelerates, inference costs fall 15-20% quarterly, and AI apps move from demos to scaled revenue. But power and cooling become the binding constraint—"chips without electricity" will define 2026.