Nvidia has earmarked $279 billion for memory procurement through fiscal 2032, a figure that reframes the competition between Micron Technology and Sandisk. The commitment, disclosed during the chipmaker’s fiscal second-quarter earnings call, makes explicit what the market has sensed for quarters: memory bits are now the binding constraint on AI infrastructure build-outs, and the two suppliers sit on different sides of that bottleneck.
The number and the timeline
Chief Financial Officer Colette Kress described “extreme pricing conditions in memory,” noting that cost increases have already outrun internal forecasts and will climb further into next year. The margin impact is quantified: gross margin is guided to trough around 71 percent in the fourth quarter of fiscal 2027, down into the low-70 percentage range. Of the $279 billion total, roughly $92 billion falls due in the remainder of fiscal 2027, with $87 billion and $88 billion scheduled for the two subsequent years. Chief Executive Jensen Huang added that unconstrained demand would imply growth above 70 percent, but supply, memory included, is the governing factor.
Where each company sits in the stack
Micron supplies the high-bandwidth memory that sits atop GPU packages and the server DRAM that populates the rest of the system memory layer. Sandisk provides enterprise solid-state drives built on NAND flash. Both are selling into the same hyperscale data centers, and both are generating record revenue. The distinction is which layer is tighter. HBM conversion is slow and capital-intensive, leaving wafer capacity as the hard ceiling. NAND faces a different squeeze: token output from inference workloads is flooding storage tiers faster than new fab space comes online.
What the hyperscaler order book actually needs
Hyperscalers are not buying a single commodity called memory. They are procuring three distinct functions. The scarcest is the bandwidth layer that determines training throughput. The workhorse layer feeds CPUs and inference hosts. The storage layer absorbs the persistent output of those models. Nvidia’s multi-year reservation signals that the first layer will remain undersupplied through fiscal 2029, which favors the supplier with HBM volume. The storage layer is also tight, but the constraint there is driven by data retention economics rather than package-level integration complexity.
What to watch next
The $279 billion figure is a claim on future wafer starts, not a purchase order for today. Micron’s roadmap for HBM3E and HBM4 volumes will determine how much of that budget it captures. Sandisk’s exposure hinges on whether enterprise SSD pricing can sustain its current trajectory once the token flood normalizes. Nvidia’s margin trough in late fiscal 2027 is the next hard checkpoint, if memory costs inflect earlier, the supercycle narrative gets a revision.