Most of the AI boom's costs land somewhere abstract: data-center balance sheets, utility bills, the national grid. This week a Reuters report tracked one of them to a much more ordinary place, the memory chips inside cheap phones and laptops. As hyperscalers buy up the world's supply of DRAM to feed AI servers, smaller device makers are being squeezed hard enough to redesign their products, order components months early, and, in a telling detail, test incoming chips to make sure they are not counterfeit. The shortage that started in the server hall has reached the checkout aisle.
The clearest sign of strain is where the money now goes inside a budget device. Raymond van Eck, chief executive of the repairable-phone maker Fairphone, says memory can account for almost 60 percent of the bill of materials on a phone costing around $400. That is a startling figure for a component buyers rarely think about, and it hits the low end hardest, where there is no premium margin to absorb it. It arrives, too, at a fragile moment for the industry: Counterpoint Research expects global smartphone shipments to fall 13.9 percent this year, to 1.08 billion units, the steepest annual decline on record.
For the smaller manufacturers, simply paying more is not always an option, because the deeper problem is availability rather than price. Their responses have turned into a small catalogue of improvisation. Finnish maker Jolla has designed two versions of its motherboard so it can switch between memory packages depending on what it can actually get, and it now tests samples from every batch to catch refurbished chips being passed off as new. Framework, the modular-laptop company, leans on its design so customers can reuse memory from old machines, and places non-cancellable orders well in advance without knowing the final price or the delivery volume.
None of this looks like it will ease soon. Back in July, SK Hynix chief executive Kwak Noh-jung called 2027 the "worst year" for the memory crunch and said he expects it to run until 2030. The forces behind it are structural, not a passing spike: every gigabyte of high-bandwidth memory a manufacturer commits to an AI accelerator is a gigabyte not going into consumer DRAM, and the accelerators pay far more for it. Framework's own experience captures the whiplash, nearly doubling prices on some laptop configurations after a supplier update, then cutting them again, retroactively, once it sourced cheaper inventory.
The through-line is quietly significant. We tend to measure the AI build-out in the things it consumes directly, chips, power, water, capital. The memory squeeze is a reminder that it also reshapes markets it never touches on purpose. A student choosing a first laptop or a family buying an entry-level phone is now, in a small way, competing with a data center for the same silicon, and losing. That is not a scandal so much as a spillover, the ordinary economics of a boom rippling outward. But it is worth watching, because spillovers like this are how an abstract technology story turns into a concrete one, on a shelf, with a higher price tag.