Memory chips now AI's biggest constraint: Nvidia CEO

Nvidia, founded in 1993, became a dominant force in the artificial intelligence boom by supplying the graphics processing units (GPUs) that provide the compute power for training and scaling AI systems. At the outset of the current AI wave, the primary goal for major tech companies was straightforward: acquire as much compute power as possible to meet demand for AI infrastructure. That priority has now shifted. CEO Jensen Huang has highlighted that specialized memory chips—not compute—have become AI's biggest bottleneck.

The reason is rooted in how modern AI systems operate. AI training and deployment depend on processing trillions of data points, which must be stored and retrieved with speed and precision. As AI is applied to more complex tasks—such as running autonomous agents or processing intricate reasoning chains instead of simple recommendations—the demand for high-bandwidth memory has become critical. Nvidia recognized this shift and began constructing integrated hardware systems with multiple components, including specialized memory chips packed directly into its products. However, this strategy ties Nvidia to external suppliers: Micron, SK Hynix, and Samsung are among the companies providing the bulk of these memory chips. Manufacturing such specialized chips is technically challenging, and only a handful of producers make them at scale—a stark contrast to Nvidia's own position in GPU manufacturing.

This dependency creates both a risk and an opportunity for Nvidia. The downside is that Nvidia's supply chain is now constrained by the production capacity of its memory suppliers; if they cannot manufacture chips quickly enough, Nvidia will face delays that could affect its own business. The upside is Nvidia's financial firepower and scale. In the quarter ended April 26, Nvidia generated $48.6 billion（約7.8兆円） in free cash flow and held $13.2 billion（約2.1兆円） in cash and cash equivalents. This gives the company the ability to pay premiums for memory chips, buy them in bulk, and secure priority treatment from suppliers—effectively shutting out smaller competitors and further entrenching Nvidia's industry position. Rather than a crisis, Huang's acknowledgment of the memory bottleneck signals the next phase of AI infrastructure evolution. Nvidia's stock has been flat this year, up only 0.60% year to date as of July 29, reflecting broader sentiment around large technology companies rather than Nvidia's operational performance.