Source: The Pomp Letter
As AI chip capacity expands, spending is consolidating around inference and deployment rather than model training—a shift that disrupts the traditional model where chip makers counted on sustained demand from data centers running expensive GPUs for R&D. Inference-optimized chip designs, edge deployment providers, and software layers handling model serving now compete for dollars that previously flowed to a handful of training-heavy hyperscalers. Price pressure on chip makers intensifies when customer spending fragments across dozens of inference use cases with different hardware requirements, not simply when capacity expands.