Nvidia stock retreats as DRAM price surge reshapes AI infrastructure economics
Nvidia stock has fallen about 15% from its May high even as company revenue forecasts remain strong, a shift that highlights changing market dynamics between GPUs and memory. The move has left Nvidia’s valuation trading at levels closer to the broader market despite continued demand for compute in AI workloads.
Nvidia stock slides after May peak
Nvidia’s share price decline since May has outpaced declines in the broader technology sector, narrowing its valuation premium versus the S&P 500. Investors are now paying less per dollar of projected profit for Nvidia than they did at the height of the AI-driven rally. The change reflects both evolving supply conditions for GPUs and surging prices elsewhere in the data-center stack.
DRAM prices soar, memory firms surge
Memory manufacturers have become the market’s new beneficiaries as demand for high-bandwidth memory outstrips supply. Companies such as Micron have seen dramatic share-price gains this year, driven by a roughly tenfold rise in some DRAM spot prices since late 2025. That spike has turned memory into a bottleneck for data-center operators building capacity for large-scale AI models.
Server builders and cloud operators are now prioritizing capacity for DRAM and HBM, viewing memory as the component that most constrains performance for modern AI workloads. The result is a sharp reallocation of capital: money that flowed into compute-focused names earlier in the year has increasingly gone to producers of fast memory modules.
Compute-hour costs fall as GPU supply expands
At the same time, the market price for renting GPU hours has softened after a May peak, according to marketplace snapshots of H100 availability. Greater supply of accelerator options and improved procurement by cloud providers have pushed per-hour compute costs down. Lower spot rates for GPUs have a direct effect on expectations of future revenue growth for companies whose valuations rely heavily on high-priced compute.
This change in the price of compute is one reason investors are recalibrating Nvidia’s premium. Even though demand for GPUs remains elevated, the arrival of more capacity and substitute accelerators has muted earlier fears of a persistent GPU shortage.
Hyperscalers build custom accelerators
Major cloud providers have pursued a strategy of developing their own processors, reducing reliance on off-the-shelf GPUs. Firms including Google, Amazon, Microsoft and others have deployed internal accelerators or custom silicon for specific training and inference workloads. Those chips are not always superior to the latest Nvidia models, but they are sufficiently capable to drive down market prices for compute services.
Industry executives point out a contrast: hyperscalers can and do build their own compute silicon, but no large cloud provider manufactures DRAM at scale. That asymmetry helps explain why memory prices remain elevated while compute pricing softens.
Technological leadership versus market realities
Nvidia’s technical contributions remain foundational to modern AI, from its GPU architectures to the widespread adoption of programming tools that made GPUs the default for researchers and enterprises. The company’s engineering achievements underpin a large portion of contemporary model training and inference workloads. Those strengths have not evaporated, even as market valuations adjust.
Yet success has also made Nvidia the focal point of competition. As more firms chase parts of the AI stack, the company faces pressure from alternatives and from shifts in component economics. The evolving mix of in-house accelerators and cheaper compute offerings has complicated the investment case that once simply equated Nvidia with the entire AI infrastructure story.
Implications for data-center economics and investors
For data-center operators, the immediate implication is a tighter focus on memory procurement and inventory planning. Capital allocation now emphasizes DRAM and HBM expansion to avoid becoming memory-constrained during large model deployments. Procurement teams are scrambling to lock in supplies and negotiate longer-term contracts to mitigate volatile spot prices.
For investors, the market is signaling a more nuanced distribution of profits across the AI supply chain. Hardware winners this year may not be the most glamorous chip designers but the firms that supply essential, hard-to-scale components. Unless there is a major technological breakthrough in high-bandwidth memory or new entrants materially expand global DRAM capacity, the current premium enjoyed by memory suppliers may persist.
Nvidia remains central to AI compute, but the market’s attention has broadened: valuation swings now reflect the interplay of GPU pricing, memory scarcity and the rise of custom accelerators rather than a single-company narrative. In this environment, tracking DRAM supply dynamics and compute-hour pricing will be as important as watching Nvidia stock itself.
