AI servers consume far more high‑bandwidth memory than traditional kit, prompting the industry's biggest producers to reroute wafer capacity. Samsung, SK Hynix and Micron are shifting production toward high‑bandwidth memory (HBM) for AI servers, while IDC forecasts DRAM and NAND supply growth will lag historical norms in 2026. Hyperscalers' large data‑centre and AI commitments are encouraging longer contracts and tighter markets, which can raise costs and squeeze device makers.

Why AI is reshaping memory demand

Large language models and other generative AI systems need far more memory per server than traditional enterprise applications. High‑bandwidth memory (HBM) sits next to accelerators like GPUs and is designed to feed those chips at high speed, making each AI server a much bigger buyer of fast memory than a laptop or phone.

And the customers are large. OpenAI and major cloud and social platforms have outlined multi‑year data‑centre projects and sizeable AI infrastructure budgets. Those commitments dwarf the orders typical of handset and PC makers, so hyperscalers place large, steady orders that use wafer capacity intensively — and manufacturers are responding.

Memory production is concentrated: Samsung, SK Hynix and Micron control most of the DRAM and NAND market used in consumer devices and also make HBM. When those producers switch lines or allocations toward HBM and next‑generation server memory, commodity DRAM and NAND volumes available to device customers shrink. Analysts at IDC have warned supply growth for DRAM and NAND will be below historical norms in 2026, a change from earlier years when capacity additions usually brought prices down.

How makers and markets are reacting

Investors have already priced in the shift: memory stocks have risen as traders and portfolio managers anticipate sustained AI demand. Executives and suppliers report the market dynamics feel different from past cycles.

  • Contracting: Suppliers say customers increasingly favour long‑term contracts over the one‑year agreements that were common before. Long contracts let hyperscalers secure capacity and give makers predictable revenue, but they also lock a larger share of output to a few big buyers.
  • Allocation: A number of manufacturers have deliberately redirected wafer starts and module assembly toward HBM and advanced low‑power DDR variants used in AI servers and next‑generation racks.
  • Commercial drivers: HBM typically commands higher prices and margins than commodity DRAM, and suppliers see strategic value in long, profitable partnerships with hyperscalers that can keep plants busy for years.

Prices, shortages and the device market

The shift is visible in spot and contract pricing. Technical notes and market reporting cited recent jumps in contract rates for certain server memory parts in late 2025, an example of how quickly rates can move when capacity tightens and buyers bid up available supply.

That squeeze filters down to consumer hardware. IDC models suggest global smartphone volumes could face pressure in 2026 under tighter supply and higher component costs, and device makers may see higher bills or constrained availability for some parts.

Related Articles

IDC's forecast that DRAM and NAND supply growth will fall short of historical norms in 2026 provides a clear forward marker: if manufacturers continue prioritising HBM, higher prices and tighter supplies could become a lasting feature of the memory market.

This article was created with AI assistance.