Nvidia has been seen as the biggest beneficiary of the artificial intelligence investment boom, but it has had a tough time in recent months.

On July 9, in local time, TechCrunch reported that Nvidia shares were down 15 percent from a May peak, based on Bloomberg reporting. Expected revenue keeps rising, but its valuation is now below the S&P 500 average. That means investors are paying less per $1 of Nvidia’s expected profit than they do for typical large U.S. companies.

Money is not leaving AI infrastructure overall. It is going elsewhere. Over the same period, Micron, a major DRAM maker, has seen its market value jump by nearly three times. Memory has emerged as a new bottleneck in data-centre expansion and a new focal point for AI investment.

Some who highly value Nvidia’s technical achievements may find that news somewhat bitter. Nvidia’s growth has been underpinned by impressive technology. It developed CUDA, the programming platform that became a core engine for AI research, and accelerated the pace of graphics processing unit development at a speed no one expected. Nvidia’s success is symbolic enough to fill a whole book, and GPUs themselves are considered among the most complex devices created by humankind.

By contrast, the story of memory companies is much simpler. The high bandwidth memory chips they make are designed to exchange data with processors as fast as possible, and have only gradually improved performance over the past 20 years. Chips and the companies themselves have not changed much, but the value of the service they provide has suddenly surged. As supply growth has failed to keep up with rising demand, memory makers have been able to raise prices by as much as 10 times over the past year.

According to a Datatrack tally, DRAM prices have risen sharply since 2023 based on spot prices in open markets paid by individual buyers. It may be easy to think there was a technological leap in the summer of 2025, but in reality the industry as a whole severely underestimated memory demand needed for data-centre expansion.

By contrast, a tally by compute marketplace Ornn shows the hourly spot price for Nvidia’s H100 peaked at about $3.20 in May over the past year and then steadily fell. As with Nvidia’s share-price trend, weakness has continued since May. In the end, Nvidia’s enterprise value is linked, for better or worse, to compute prices, and those prices are now falling. By contrast, the value of memory makers including Micron is linked to DRAM prices, and those prices keep rising.

Wayne Nelms (웨인 넬름스), co-founder and chief technology officer of Ornn, said the force creating the gap is a simple supply-and-demand issue. Google, Amazon, Microsoft and OpenAI are rolling out their own customised processors to reduce dependence on Nvidia. Even if these chips are not as strong as Nvidia’s latest products, they are enough to bring compute prices down.

Nelms said, "More GPU and accelerator makers are entering the market. Everyone wants to make their own silicon, but no one makes their own DRAM." He added, "Unless there is a major technological breakthrough in HBM, the supply-demand structure changes, or a new entrant appears in the memory market, the current situation will generally continue."

For Nvidia, it is a somewhat frustrating situation, and much of it is also the result of its own success. It is standing in the middle of a market that everyone wants to enter, as the price of proving the value of compute. In the meantime, companies with relatively simple technology and little flash are quietly building wealth on the market’s periphery.

Ultimately, the current AI infrastructure market is placing higher value on the scarcity of memory supply that supports computing power than on computing performance itself. There are 2 points to watch. One is whether compute prices will keep falling as GPU supply expands and in-house chips spread. The other is whether the current price advantage will weaken as memory supply, including HBM, catches up with demand. Where AI infrastructure investment concentrates on the scarcest resource is likely to keep dividing related companies’ share-price trends.