Bernstein: DRAM 2027 will continue to surge, earning twice as much per wafer for HBM, the era of memory superprofits has just begun

Bernstein Reports on June 22: Conventional DRAM Prices to Surge 4.5x from 2025Q3 to 2026Q2, Revenue per Wafer Already Double that of HBM, Forecast to Continue Rising in 2027.
(Background: The nine major US industries jointly petition the Trump administration: Memory price hikes and shortages are pushing us to the brink.)
(Additional context: SK Hynix and Micron's market caps surpass Bitcoin! BTC has fallen to the 19th largest asset globally.)

Table of Contents

Toggle

• Conventional DRAM per-bit selling price surpasses HBM
• HBM still needs three times the price to catch up with conventional DRAM
• Taiwan supply chain benefits, AI cost pressures rise in tandem

In a recent report released on June 22, analyst Bernstein pointed out that after a roughly 4.5-fold surge in conventional DRAM prices from Q3 2025 to Q2 2026, prices may continue to rise in 2027. This indicates that the "superprofit cycle" in the memory industry has not yet peaked, and that conventional products are actually more profitable than HBM.

Conventional DRAM per-bit selling price surpasses HBM

Bernstein notes that currently, the average unit price per bit for conventional DRAM has already matched or even exceeded that of high-bandwidth memory (HBM). This means that in the AI chip market, the most talked-about HBM product has, in terms of unit price, actually been overtaken by what appears to be "traditional" conventional DRAM.

According to the report's estimates, considering higher bit density and yield, the revenue per wafer for conventional DRAM in 2026 could be twice that of HBM, with significantly higher profit margins. For example, SK Hynix's HBM accounts for about 23% of wafer usage, but the remaining 77% is contributed by conventional DRAM, which actually yields higher profits per wafer.

HBM still needs three times the price to catch up with conventional DRAM

Bernstein provides an intuitive figure: HBM prices need to increase by about three times to match the revenue per wafer of conventional DRAM. This also explains why the narrative in the AI chip market focuses on HBM, but in reality, conventional memory remains the main profit driver for memory giants.

The firm also warns that storage chip manufacturers' pricing strategies are unlikely to be so aggressive. They are aware that excessively high HBM costs could hinder the overall AI ecosystem development, ultimately suppressing storage demand. Bernstein writes in the report that if the three major manufacturers simultaneously raise prices for both HBM and conventional DRAM, the cost of AI servers will increase cloud service rates, indirectly boosting capital expenditures for tech giants.

Taiwan supply chain benefits, AI cost pressures rise in tandem

The rising prices of conventional DRAM directly impact Taiwan's memory supply chain—foundries at Nanya and packaging/testing plants are operating near full capacity. Recently, the nine major US industries jointly petitioned the Trump administration, pointing out that memory price hikes and shortages are affecting downstream manufacturing costs. The progress of US-based fabs by SK Hynix and Samsung has also accelerated.

The 4.5x surge from Q3 2025 to Q2 2026 has already pushed the memory industry into a similar "superprofit cycle." The three major manufacturers are simultaneously raising prices for both HBM and conventional DRAM, creating a scenario of volume and price increases. If prices continue to rise in 2027, hardware costs for AI training and inference could further escalate, impacting the pricing strategies of large models.