AI infrastructure capital is rotating from GPU makers to memory suppliers, and Micron's stock has nearly tripled, because renting AI compute by the hour has made memory the constraint.
The AI buildout's bottleneck has moved from compute to memory, and capital is following. Nvidia, the dominant supplier of the processors that train and run AI models, is down roughly 15 percent from its May peak. Micron, the largest US-listed memory chipmaker, has nearly tripled over the same window. The split tracks a quieter shift inside the same buildout: the constraint buyers pay up for isn't the GPU anymore, it's the DRAM and high-bandwidth memory that sits next to it.
The mechanism behind that rotation is one of Nvidia's own making. CUDA became the lingua franca of AI compute, and hyperscalers, neoclouds, and brokered compute marketplaces now publish availability and pricing close to real time, which lets a model lab arbitrage the supply it cares about. When GPUs were scarce in 2024 and 2025, that meant paying up for Nvidia silicon. When GPU supply caught up this year, the same machinery let buyers redirect spending upstream to the next constraint.
Micron reported this spring that Q3 FY2026 revenue roughly quadrupled year-over-year, driven by data-center pricing for both conventional DRAM and the stacked HBM parts that pair with AI accelerators (CNBC). Spot pricing on standard DDR4 parts moved the same way; data tracked by SiliconAnalysts shows 8Gb DDR4 contract and spot quotes climbing sharply through the spring into early summer. The HBM market, where Micron competes with SK hynix and Samsung, has tightened further as Nvidia's own Blackwell-class ramps pull in allocations. A Motley Fool piece from early July called the move a forecast rather than backward-looking commentary (fool.com).
This isn't a story about Nvidia losing customers to Micron. The relationship is indirect: the same hyperscaler or model lab that signs a multi-billion-dollar GPU contract with Nvidia is also signing a memory contract with Micron or SK hynix to fill out the rack. Nvidia's revenue is still projected to grow; the move in its stock is multiple compression, not a demand collapse. A SeekingAlpha note this month pointed out that Nvidia now trades at a forward P/E below the S&P 500 average, the first such instance in over a decade (SeekingAlpha). The Tech Marketer put the closing valuation gap at levels last seen in the early 2010s (TheTechMarketer). The market is paying Micron for solving a problem Nvidia's own transparency made legible.
A small but telling lane is forming around that visibility. Ornn, a startup that publishes itself as "Financial Products for Compute," treats compute capacity as something with a yield, a duration, and a price (Ornn). When compute can be rented like a financial asset, the next constraint on it, today memory, can be financed and arbitraged in the same language.
TechCrunch called Nvidia "a victim of the compute marketplace it created," the cleanest articulation of the rotation thesis (TechCrunch); independent confirmation that the rotation is structural, rather than a one-quarter reset, is light. HBM supply is concentrated in three vendors and any of them opening supply (Samsung catching up, SK hynix expanding HBM3e, a memory glut returning) would blunt the trade. GPU supply can tighten again if a frontier-model release pulls demand forward.
Both Micron's Q4 FY2026 print and Nvidia's next datacenter earnings call land before mid-August, and either update will test whether the rotation is structural.