Nvidia faces a shifting market as memory becomes the new AI bottleneck

The GPU scarcity that defined the last year has moderated, causing the hourly rental price for an H100 chip to decline from its May peak of approximately $3.20. As Google, Amazon, Microsoft, and OpenAI accelerate the development of their own custom silicon, Nvidia’s once-unrivaled dominance in compute is facing downward price pressure. Even if these alternative processors fall short of Nvidia’s bleeding-edge performance, they are sufficient to dilute the company's pricing power.

Conversely, Micron and other DRAM producers have seen their market value nearly triple. Unlike the complex evolution of GPU architecture, the demand for high-bandwidth memory has surged simply because the industry underestimated the total capacity required for data centers. Wayne Nelms, CTO and co-founder of Ornn, notes that while every major tech firm is attempting to design its own AI accelerators, none are producing their own memory. This creates a structural scarcity that allows memory manufacturers to raise prices while compute providers see their margins compress. Nvidia is now a victim of its own success, having validated the massive potential of AI infrastructure only to watch the financial gains migrate toward the simpler, less volatile bottleneck of memory supply.