Nvidia loses ground as memory shortage reshapes the AI compute market

On July 9, 2026, Nvidia found itself in an unusual position: despite rising revenue forecasts, its shares had fallen 15% from their May peak, while memory maker Micron had nearly tripled its market value. The reason is an easing GPU shortage and a new memory shortage in data centers, which is changing where the money goes in AI infrastructure.

Why Nvidia is getting cheaper

The market has begun paying less for compute on Nvidia GPUs, even though demand for AI infrastructure persists. According to data cited by TechCrunch, the company's shares have lost 15% since their May high; at the same time, the market values its expected earnings at less than those of the average large US company in the S&P index. This does not mean revenue is falling: forecasts for it keep rising. But investors see that the key resource of the AI boom, access to GPUs, is no longer as scarce as it was a year earlier.

• On July 9, 2026, TechCrunch reported that Nvidia shares had fallen 15% from their May 2026 peak.
• The hourly spot price of the Nvidia H100 reached about $3.20 in May 2026 and then declined steadily.
• Micron's market value nearly tripled over the same period.
• Over the past year, memory makers have been able to raise DRAM prices roughly tenfold.

Nvidia built the technological foundation of today's market: its CUDA software platform made its GPUs the standard tool for AI research and model training. The company also refreshed its GPU lineups quickly, and those chips rank among the most complex mass-produced electronic devices. But the success of that strategy attracted competitors: cloud platforms and labs began looking for ways to reduce their dependence on a single supplier.

What became the new bottleneck

The new constraint for data centers is memory, not the accelerators themselves. Micron makes DRAM and high-bandwidth memory (HBM), components that move data quickly between the processor and memory. These technologies have been developing for a long time, but the industry underestimated how much memory the mass build-out of AI data centers would require. Demand outpaced the expansion of production, so spot-market prices jumped sharply.

This is an important shift: the cost of renting Nvidia H100 compute is falling while the price of DRAM is rising. Data center owners can get more accelerators, or rent them more cheaply, but they still run into limited memory availability. For Micron and comparable manufacturers, a familiar technology and limited supply are enough to profit from the new wave of AI infrastructure capital spending.

The spot price reflects deals with immediate delivery, not long-term contracts. On the Datatrack chart included in the article, the jump in DRAM prices begins in August 2025. The author stresses that it is not tied to a sudden breakthrough in memory technology. The market simply recognized the scale of data center demand too late.

Why compute is getting cheaper

The supply of accelerators is expanding in two directions at once: more GPU and accelerator makers are entering the market, and the largest buyers are developing their own processors. Google, Amazon, Microsoft and OpenAI have already released custom chips to depend less on Nvidia. Such chips may trail Nvidia's latest products, but they are good enough for some workloads and for putting pressure on compute prices.

Wayne Nelms, co-founder and CTO of the Ornn marketplace, describes the situation as ordinary supply and demand. In his words, companies want to design their own silicon, but almost nobody makes their own DRAM. Until a notable breakthrough in HBM, a shift in the supply-demand balance or a major new memory player appears, the imbalance may persist.

"Everyone wants to make their own silicon, but nobody makes their own DRAM," said

Wayne Nelms, co-founder and CTO of Ornn.

This does not erase Nvidia's technological lead. Rather, the market is turning the GPU from a scarce resource into a more competitive commodity. As alternatives multiply, excess profit moves to the next scarce element of the system: memory. It also shows that the valuation of AI companies and infrastructure suppliers depends increasingly not only on chip quality but on the specific constraint in the supply chain.

What this means

The AI boom has not ended, but its financial center is shifting: Nvidia remains the key supplier of accelerators, while memory makers increasingly capture the short-term gains. For developers and data center operators, this means the cost and availability of HBM and DRAM are becoming just as important as the choice of GPU.