Chip-industry commentator Jukan relayed his own reading of a Meta Engineering blog post, arguing that the metric for choosing AI data center storage is shifting from dollars-per-terabyte to dollars-per-GPU-hour. Meta published the storage design philosophy on its engineering blog, and a few hours later Jukan noted that Micron had also put out a related white paper with Meta. AI compute chips like Nvidia's GPUs roughly triple in processing speed every two years, but the storage and networking gear that feeds them data hasn't kept pace. When a GPU can't get its data fast enough, an expensive chip sits idle. Meta describes this as one of the biggest single causes of wasted GPU time in AI infrastructure.
Storage has traditionally been judged on how cheaply it can hold a terabyte of data, the dollars-per-terabyte metric. But in AI training, if slow storage leaves GPUs idle longer, that lost compute time can cost more than whatever was saved on storage. So Meta's engineering team argues storage should instead be valued by how well it keeps GPUs busy, a dollars-per-GPU-hour metric. A factory comparison makes this concrete. Skimping on a cheap part that ends up stalling the entire assembly line costs far more than the part itself was worth. An hour of GPU time costs far more than whatever gets saved by buying cheaper storage, which is why the math flips. If this yardstick becomes an industry standard, makers of fast, high-performance NAND flash, the kind of chip that keeps data even when the power is off, gain an easier case for charging more for storage that keeps GPUs fed. The competitive center of gravity shifts from raw capacity to raw speed. The fact that Micron co-published this framing with Meta itself signals that storage vendors are starting to use it as a marketing argument too.
Why it matters · If this yardstick sticks, storage competition shifts from capacity to speed, and the winners could be makers of the fastest high-performance NAND.
Worth asking · Will the dollars-per-GPU-hour yardstick reshuffle who wins in storage?