High-bandwidth memory (HBM) contract prices will more than double next year, according to a report published by Businesskorea. The figure marks one of the sharpest projected price escalations on record for the memory segment and lands amid sustained demand pressure from AI accelerator manufacturers.

The report's central claim is straightforward: HBM pricing, which suppliers negotiate with major customers on long-cycle contracts, will rise by a factor exceeding two over the coming year. For buyers, that means budgeting for input costs on next-generation AI hardware at more than twice the levels that prevailed in the current cycle.

Such an increase would be unusual for the memory industry, where DRAM and NAND pricing has historically moved through pronounced boom-and-bust cycles. HBM, however, has decoupled from that pattern over the past two years. The stacked-memory technology, which packages DRAM dies vertically and sits close to the compute silicon in AI accelerators, has become the binding constraint on how many AI GPUs and custom accelerators the industry can ship.

Three structural factors explain why HBM pricing carries leverage that commodity DRAM never did.

First, capacity is scarce by design. HBM production consumes far more wafer starts per bit than conventional DRAM because of lower yields on stacked dies and the larger die sizes involved. Every HBM stack shipped displaces conventional DRAM output from the same fabrication lines.

Second, the customer base is narrow and deep. A handful of AI chip designers account for the overwhelming majority of HBM procurement, and they compete with one another for allocation. When buyers with urgent, price-insensitive demand bid against each other, contract prices detach from cost curves.

Third, supply remains concentrated. Only three memory makers — Samsung Electronics, SK hynix and Micron — produce HBM at scale, and qualification cycles with major accelerator vendors take months. A supplier that wins qualification on a leading platform locks in allocation advantages that persist across product generations.

The Businesskorea report does not break down the projected increase by HBM generation, customer, or supplier, and it does not specify whether the doubling applies to average contract prices across the market or to specific negotiated deals. What it does signal is direction and magnitude: memory vendors will enter the next pricing round with materially stronger hands than in any recent cycle.

For system builders and cloud operators, the implications are direct. Memory cost as a share of total accelerator bill of materials will rise. Systems architects may respond by pushing harder on memory capacity per node to amortize the price increase, or by extending refresh cycles on existing HBM-equipped hardware.

For the memory makers themselves, the pricing surge translates into margin expansion at a scale the DRAM industry has rarely sustained. The open question is timing: pricing that doubles in one year invites capacity expansion, and HBM capacity already committed by the major suppliers will come online progressively. Whether elevated prices hold through that supply response will depend on whether AI accelerator demand keeps absorbing every stack the industry can produce.

For now, the report's headline number stands as the planning assumption buyers will carry into negotiations: budget for HBM at more than double current pricing next year.