Nvidia, long the dominant force in AI chips, has seen its share price fall around 15% since a peak in May 2026, even as its projected revenue keeps rising. The stock is now cheaper relative to expected earnings than the average S&P company. The shift matters because it signals that the AI infrastructure bottleneck — and the money that follows it — has moved away from graphics processors and towards memory chips, upending assumptions about where value in the AI buildout accrues.

The driver is straightforward supply and demand. Last year's alarming GPU shortage has eased, partly because Google, Amazon, Microsoft and OpenAI have built their own custom processors that, while not matching Nvidia's best, are good enough to push compute prices down; the spot price for an hour on an Nvidia H100 has fallen since May. Meanwhile data centres cannot get enough DRAM, and memory makers such as Micron — whose value has nearly tripled — have raised prices roughly tenfold in a year. As Ornn's Wayne Nelms notes, everyone wants to make their own silicon but no one makes their own DRAM, so the imbalance is likely to persist until a breakthrough or new supplier emerges.