AI demand is keeping HBM tight while the rest of the chip cycle waits

Samsung’s semiconductor campus outside Hwaseong does not sit still. In August, the upstream line processing high-bandwidth memory ran at near capacity, with chip trays bound for Nvidia’s latest AI accelerator, while the parallel line for more ordinary DRAM lagged behind, underbooked and quiet, according to Shattered on August 28, 2026.

At stake is not just one segment running hot while the rest cools, but an industry whose margins and margins-of-safety have diverged. AI’s hunger for HBM pulled every wafer the majors could spare into a single, volatile channel. This is not another story of overshoot and flush in semiconductors. For the first time, the profitability and equipment spend of the biggest memory companies now turns on a segment that is both capacity-constrained and—at these prices—profitable enough to make the rest wait their turn.

Micron Technology told investors in late August that its full-year HBM output was already sold out, with next-generation HBM4 commanding a 55% to 70% price premium over the older HBM3 series, according to 24/7 Wall St. on August 25, 2026. Fortune reported on February 15, 2026, that SK Hynix, the largest supplier by volume, said memory sales for AI had doubled in each of the last two years, and the firm’s June earnings call confirmed that more than two-fifths of its 2026 DRAM revenue is now from HBM, up from less than a tenth three years prior.

HBM manufacturing is fundamentally less flexible than legacy DRAM. The technology stacks multiple memory dies with through-silicon vias, each addition bumping the odds of a ruined stack. Samsung said reaching 80% yield with HBM4 took three separate learning cycles and forced a partial pullback in conventional DRAM production, Shattered reported on August 28, 2026.

PC makers, phone OEMs, and auto suppliers are competing for leftovers.

The incentive is plain. SK Hynix and Samsung have no reason to shift toward idle commodity contracts when every HBM wafer is contracted at a premium. The effect is a chip cycle that runs on two clocks: one fast, thin, and high-value. The other is slow, thick, and languishing, with spillover only if AI demand broke, not before.

HBM now accounts for 41% of global DRAM revenue, a jump from just 8% in 2023, Great Money reported in August 2026. Even so, HBM physically constitutes only a small fraction of all DRAM bits shipped, a sign that ordinary device customers are neither outbidding the AI market nor finding substitute supply.

AI server builders are absorbing price increases without pushback.

Nvidia’s Rubin launch, relying on 20-stack HBM4 modules, passed the cost directly into system prices, with secondary buyers reporting waitlists into early 2027 for certain configurations, according to Shattered on August 28, 2026. Auto makers and traditional server OEMs are not so buffered. They either take non-HBM DRAM, now a sideline for the majors, or defer launches until the market loosens.

The supply risk moved from inventory to the order-books—nobody pays for it until the shortage arrives.

What looks like a positive for the memory makers—a part of the business that is both tight on supply and flush with margin—also means that manufacturer earnings are newly vulnerable to a single channel’s cycle. A surprise drop in AI buildouts would leave them with an installed base tailored for a customer who can simply wait, as the rest of the chip industry already is. The risk has shifted from inventories to order-books, and the gap is not visible in conventional market statistics.

AI server buyers are measured in weeks, able to absorb immediate cost and supply risk.

Consumer and device memory customers are on quarters, forced to defer and hope for relief in the next round of equipment buildouts. If HBM expansion lands too late, it brings margin compression without share. Samsung and SK Hynix have seen this before, but almost never in a context where the leading edge is already sold forward a year out.

One risk has never been priced. Contract revenue is locked in, but capacity for non-AI memory is not. Should AI demand break, memory makers would face both margin collapse in HBM and a market share race that their own expansion plans have yet to resolve.

The industry is betting on continued AI demand not just for growth, but for the maintenance of a two-tier memory market whose base remains unsteady.