Nvidia, a longtime industry leader, has had a rough few months. The bottom line, Bloomberg reports in ugly detail, is that the company’s stock is down 15% from its peak in May, even as projected earnings continue to rise. Relative to estimated earnings, the company’s stock is currently cheaper than the S&P average. Investors pay less per dollar of NVIDIA’s expected earnings than they would for a typical large American company.
Money is still flowing into AI infrastructure stocks, but most of it is going into memory companies. Over the same period, the value of Micron, one of the world’s largest makers of DRAM, a standard type of memory chip found in computers and servers, has nearly tripled, making memory the new bottleneck for data centers and hot new AI deals. The basic reason is simple. That’s because the GPU shortage that seemed so alarming last year has eased a bit. At the same time, data centers need all the memory money can buy.
For those who appreciate Nvidia’s technological achievements, this may be a bit disappointing. There are a number of truly impressive technologies behind Nvidia’s rise, including the development of CUDA, a widely adopted programming platform that has made NVIDIA GPUs the default engine for AI research, and pushed the pace of GPU development to speeds few thought possible. Nvidia’s success is something you could write a book about, and the GPU itself is one of the most complex devices ever produced, right at the cutting edge of human capabilities.
For memory companies like Micron, the story is much simpler. They build high-bandwidth memory chips (specialized components designed to move data in and out of processors as fast as possible), which have been gradually improving in performance over two decades. Unless the chips and companies change much, the services they provide suddenly become very valuable. And because demand is growing faster than supply, they’ve been able to increase prices tenfold over the past year.
Here’s how DRAM spot prices (the price buyers pay for chips on the open market, as opposed to long-term contract rates) will look like in 2023 and beyond, via Datatrack.
You might think that the summer of 2025 would be an amazing technological breakthrough, but instead, the entire industry vastly underestimated the amount of memory needed to scale data centers.
By comparison, this (via computing marketplace Ornn) shows how the hourly spot price for the Nvidia H100 GPU has changed over the last year.
Like Nvidia stock, it peaks in May (about $3.20 an hour) and then declines steadily. For better or worse, Nvidia’s value as a company is tied to the price of computing, and those prices are falling. Micron and its peers are pegged to DRAM prices, which continue to rise.
When I spoke to Wayne Nelms, co-founder and CTO of Orn, about the factors driving that disparity, he framed it as a simple issue of supply and demand. Google, Amazon, Microsoft, and even OpenAI have launched their own custom processors to reduce their dependence on Nvidia. Even if those chips aren’t as good as Nvidia’s latest models, they’re still enough to drive down the price of computing.
“More GPU and accelerator players are entering the market. Everyone wants to make their own silicon, but no one is making their own DRAM,” Nelms told me. “Until there’s a major technological breakthrough in HBM (high-bandwidth memory), a change in supply and demand, or someone new enters the (memory market), I think it will continue to be more or less the way it is today.”
This is a frustrating situation for Nvidia, and is largely a product of its own success. Having proven how valuable computing can be, the company found itself at the center of a market everyone wanted to be a part of. Meanwhile, simpler technologies and less interesting companies are enriching themselves on the sidelines.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
