AI
Nvidia's dominance dims as memory chips become AI's new bottleneck
The GPU maker's stock has tumbled 15% as competitors muscle in on custom chips, leaving memory suppliers to capture the real margin surge.
TechCrunch — AI
Nvidia's reign as AI infrastructure's undisputed champion has hit a rough patch. The chipmaker's stock has lost 15% of its value since its May peak, and on a price-to-earnings basis, it now trades below the broader S&P 500 average — a striking reversal for a company that has anchored the AI boom.
The culprit isn't a lack of demand for artificial intelligence itself. Rather, Nvidia finds itself squeezed by two opposing forces: rivals entering the GPU and accelerator market with competent alternatives, and a critical shortage of memory capacity that's driving those components toward astronomical prices.
The GPU Marketplace Splinters#
The GPU shortage that defined 2025 has eased considerably. Major cloud providers—Google, Amazon, Microsoft, and OpenAI—have all rolled out custom processors. While these chips don't match Nvidia's latest offerings, they're sufficiently capable to dent pricing power.
This fragmentation is a direct consequence of Nvidia's success. By establishing CUDA as the AI industry standard and proving the enormous returns on compute, Nvidia invited exactly the kind of competition now weighing on its margins. Meanwhile, the spot price for compute time on an Nvidia H100 GPU has declined steadily from a May peak near $3.20 per hour.
Memory Becomes the Gating Factor#
What Nvidia loses, memory companies gain. Micron and other DRAM suppliers have tripled in value over the same period, riding demand that suppliers vastly underestimated. The data center buildout requires far more memory than the industry anticipated in 2024, and unlike GPUs, memory production remains bottlenecked.
High-bandwidth memory (HBM) chips, though incremental refinements of decades-old technology, command prices ten times higher than a year ago. Supply cannot keep pace with demand, allowing memory makers to raise prices with little pressure to innovate.
Why This Persists#
According to Wayne Nelms, co-founder of the compute marketplace Ornn, the asymmetry reflects straightforward market mechanics: "More GPU and accelerator players are entering the market. Everyone wants to make their own silicon, but no one is making their own DRAM." Without a memory breakthrough, new supply, or an entrant bold enough to tackle chip manufacturing, memory margins will likely stay elevated while Nvidia's compute pricing continues to compress.
