The infrastructure category Nvidia built is now the one squeezing its margins — a classic incumbent problem in a market it invented.
Nvidia established the modern AI compute market. Now it's watching that market's second-order effects erode the pricing power that fueled its rise. New entrants, alternative accelerators, and hyperscaler in-house silicon are all squeezing what used to be an uncontested moat.
Nvidia's GPUs remain the technical default for training frontier models. But around Nvidia, an entire supporting infrastructure has emerged — cloud brokers, spot markets, per-token API layers — that commodifies compute delivery. That means Nvidia sells the same GPU, but customers extract more value per dollar and reserve less premium capacity.
Meanwhile, AMD's MI-series, Groq's LPUs, and now the wave of hyperscaler custom chips give buyers substitutes at inference. Nvidia still wins training, but training is a smaller fraction of long-term revenue than inference.
Nvidia's move up the stack — CUDA-X, AI Enterprise, and NIMs — is designed to counter exactly this squeeze. Whether those higher-margin products can replace the pure-hardware margins is the question the next four quarters will answer.
Source: TechCrunch