Samsung, Micron, and SK Hynix are all accelerating development of HBM4 memory — the next generation of high-bandwidth memory designed for AI accelerators — with production targets pointing toward 2027. Early specifications indicate HBM4 will deliver over 2.5 TB/s bandwidth per stack, compared to approximately 1.2 TB/s for current HBM3e, while consuming less power per gigabyte transferred.
The Current Supply Crisis
The urgency behind HBM4 development is driven by a current market under severe strain. HBM3e prices have surged significantly over the past 18 months as AI accelerator production has consumed the majority of available output from all three manufacturers. The result is a constrained market with long lead times and elevated pricing that is flowing through to consumer products.
For NVIDIA's H100 and H200 data centre GPUs, HBM3e supply has been a production bottleneck for over a year. The Blackwell GB200 uses HBM3e at even higher capacity — 192 GB per GPU — which has not eased the constraint. Data centre operators are reporting months-long delivery timelines.
The Consumer Cascade
The consumer impact is indirect but measurable. Advanced packaging capacity — the specialised manufacturing process used to stack HBM dies — is shared with GDDR7 and LPDDR5X production at key process nodes. Memory manufacturers allocate this capacity based on margin and strategic priority, and AI memory currently commands both.
Sony's executives cited memory supply as a contributing factor to PS5 production constraints in their most recent earnings call. GPU manufacturers building GDDR7-based cards are experiencing similar lead time pressures.
What HBM4 Changes
HBM4's arrival in 2027 will partially relieve pressure as it takes over from HBM3e for the highest-performance AI applications, freeing some advanced packaging capacity for consumer memory production. However, the overall memory market is unlikely to normalise significantly until AI cluster investment cycles slow — which shows no sign of happening in the near term.