According to the latest research, the significant price increase for conventional DRAM since the second half of 2025 reflects a market situation of supply falling short of demand. However, the annual pricing negotiations for HBM among the three major manufacturers have prevented HBM contract prices from promptly reflecting the quarterly upward trend in the market.
As the timeline progresses into the second quarter of 2026, buyers and sellers are currently negotiating the supply of HBM4, the mainstream product for 2027. The analysis suggests that, based on the tight DRAM supply-demand dynamics, the high manufacturing complexity, and elevated costs associated with both new and existing HBM generations, the three major manufacturers are likely to substantially increase their HBM quotations in 2027.
Tracking the per-wafer revenue for HBM and conventional DRAM, which is estimated based on die size, yield, and price per gigabit, the per-wafer revenue for HBM was surpassed by that of DDR5 64GB RDIMM in the first quarter of 2026. Consequently, HBM's profit margin has also fallen below that of DDR5 64GB RDIMM since that quarter.
Therefore, manufacturers will adjust capacity allocation between HBM and conventional DRAM based on the price levels achieved in HBM negotiations. This strategy aims to ensure HBM can continue to serve as a core component for AI training and inference infrastructure, driving the development of the broader AI ecosystem. This, in turn, is expected to stimulate comprehensive demand for conventional DRAM, including RDIMM, Server LPDDR, and memory for edge devices.
Primary Demand Drivers for 2026 and 2027
In terms of demand drivers, HBM demand is expected to remain robust from 2026 to 2027, fueled by the accelerated construction of AI infrastructure, though the specific drivers differ slightly between the two years.
The primary demand momentum for HBM in 2026 is anticipated to come from capacity upgrades in AI ASICs, which will substantially increase the HBM capacity per AI chip from 96/192GB to 216/288GB. While the HBM capacity per GPU in the NVIDIA Rubin platform remains flat compared to the previous generation, overall demand will still be pushed higher by growth in shipment volumes.
In 2027, NVIDIA's Rubin Ultra platform is projected to further increase HBM capacity per GPU to 384GB. Meanwhile, AI ASICs like Google's TPU will amplify their bit demand for HBM due to an increase in the number of chips deployed.
It is estimated that HBM wafer starts by the three major manufacturers will account for approximately 18%, 22%, and 30% of total DRAM wafer starts in 2025, 2026, and 2027, respectively, based on year-end figures. In terms of bit supply, HBM is projected to constitute about 8%, 9%, and 13% of the total DRAM bit supply in those respective years.
In summary, as HBM evolves into its next generation by 2027, die sizes are expected to expand again alongside rising demand. This will intensify the crowding-out effect on conventional DRAM production capacity, providing manufacturers with ample justification to raise HBM prices. Consequently, they are likely to secure clear pricing dominance in the 2027 HBM negotiations.