Written by: Xiaobing

HBM is still priced under annual contracts, while regular DRAM has increased by 4.5 times. At the same wafer fab, the money made from regular memory is twice that from HBM, with revenue double and gross profit nearly three times higher. This means that HBM prices must double or increase by 2.5 times next year; otherwise, no memory manufacturers would be willing to allocate capacity to it. The issue is, HBM is sold packaged with NVIDIA GPUs; if HBM prices rise, NVIDIA will further amplify the price increase by 4 times to maintain its 75% gross margin, passing it on to cloud vendors.

Bernstein's Asia Technology team, led by Mark Li, maintained an Outperform rating for Samsung, SK Hynix, and Micron in a global storage research report released on June 22, and significantly raised the target prices: Samsung's target was raised from 225,000 KRW to 440,000 KRW, SK Hynix from 1,150,000 KRW to 3,300,000 KRW, and Micron from 510 USD to 1,300 USD. Kioxia was maintained at Underperform with a target price unchanged at 40,000 JPY. MediaTek was also maintained at Outperform with a target price of 4,380 TWD.

The underlying logic of this report is that the storage industry is undergoing an unprecedented structural rupture.

The rules about what to produce and profit from in the same wafer are being rewritten

From the third quarter of 2025 to the second quarter of 2026, regular DRAM prices rose by about 4.5 times. Meanwhile, because HBM is bound by long-term contracts, its price has remained almost unchanged. The result is that Bernstein estimates that allocating production capacity to regular DRAM in 2026 would generate revenue more than twice that of HBM per wafer, with gross profit almost three times higher.

Samsung and Micron made it clear in their Q1 2026 earnings call that the profit margins for non-HBM DRAM had already surpassed HBM, and as regular DRAM prices continue to rise, this gap is widening. Bernstein anticipates that regular DRAM prices may have about 25% more room for increase before reaching a peak in 2027.

This presents a sharp figure for HBM procurement negotiations: for each wafer's revenue of HBM to catch up to that of regular DRAM, HBM prices would need to increase threefold. However, memory manufacturers are also aware that HBM is a key component for AI infrastructure, and excessively aggressive pricing could harm the healthy development of the entire AI ecosystem. SK Hynix stated during the earnings call that it would "prioritize the optimal allocation between HBM and regular DRAM" and would not pursue maximum revenue.

Considering these factors, Bernstein's assessment is that HBM prices will average increase by 2-2.5 times in 2027 (Exhibit 1-2). Even so, the profit level for HBM will still be below that of regular DRAM, though the gap will significantly narrow compared to 2026.

The true impact of HBM price increases is hidden within NVIDIA's price hikes.

Regular DRAM and NAND can be directly purchased by cloud vendors from memory manufacturers. However, HBM is different; it is packaged within NVIDIA's GPUs and is part of the latter's cost of goods sold (COGS).

If NVIDIA maintains a 75% gross margin on the VR200 (Vera Rubin NVL72) cabinet, then the portion of the HBM price increase would need to be inflated by 4 times for pricing. Bernstein's estimation logic is that HBM originally accounted for about 5% of the VR200's selling price, and after the HBM price increase, this proportion would increase to 6%. However, if NVIDIA wants to keep its 75% gross margin unchanged, the selling price of the cabinet would need to rise by 24%.

For an AI data center deploying VR200 cabinets, just the HBM cost transfer alone could raise total capital expenditure (including cabinets and external costs) by 4%-15%, depending on whether NVIDIA raises prices. Combined with the rise in regular DRAM and NAND (about 14%), cloud vendors' AI capital expenditures would need to be approximately 30% higher than originally planned (Exhibit 3).

The report refers to this process as "re-calibration" and suggests that cloud vendors will not slow down their AI investments due to this, but will inevitably distribute cost pressures among various supply chain segments, and this could possibly manifest in charging different token prices to different customers.

A wave of profit corrections is approaching; who benefits and who suffers

Bernstein has raised the assumption for average HBM prices in 2027 by 2-2.5 times, leading to significantly higher profit forecasts than market consensus: Samsung's EPS for 2027 is expected to be 26% higher than consensus, SK Hynix 32% higher, and Micron 38% higher (Exhibit 11-13). Analysts predict that the annual HBM negotiation will wrap up in the coming months, and seller consensus will be revised upward, driving stock prices higher.

The increase in HBM prices is not purely beneficial for memory manufacturers. Bernstein specifically pointed out that a larger HBM exposure means lower overall profitability because the profit margin of regular DRAM is simply too high. Samsung leads in HBM4 technology, and monitored data on South Korean memory exports show that in May, Samsung's unit export value significantly increased, indicating that HBM4 has already begun shipping (Exhibit 8). However, Samsung has also expressed its pursuit of higher profitability, which may result in allocating more capacity to regular DRAM rather than HBM.

Kioxia is the only loser, as it has only NAND business and no HBM, making it unable to benefit from this round of HBM price increases and profit corrections.

MediaTek may become another type of beneficiary. The report believes that if cloud vendors demand direct procurement of HBM to avoid NVIDIA's price increases, ASIC (Application-Specific Integrated Circuit) service providers are well-positioned to meet this demand. MediaTek has executed solidly on the TPU project, and supply chain research shows that there is an upside risk for the outlook in 2028. The stock has risen approximately 130% in the past two months, but Bernstein still maintains an Outperform rating.

Valuation method switched to price-to-earnings ratio, target price still has 15%-26% upside potential

Due to the net return on equity (ROE) of memory manufacturers in this cycle reaching unprecedented levels—Samsung 55%, SK Hynix 108%, Micron 85% (Exhibit 18)—and the accumulating cash at an astonishing pace (by 2027, cash could account for 70%-80% of book value), the traditional price-to-book (P/B) valuation method has lost its relevance. Bernstein instead employs a 1-year forward price-to-earnings (P/E) ratio, with target multiples set near historic cycle lows: 6.2 times for Samsung and SK Hynix, and 7.7 times for Micron.

The corresponding target prices are: 440,000 KRW for Samsung (upside 26%), 3,300,000 KRW for SK Hynix (upside 20%), and 1,300 USD for Micron (upside 15%).

For 2028, the report anticipates that with more clean rooms going into production, storage prices will soften, and all three manufacturers' revenues will decline year-on-year. However, even in the down cycle of 2028, the DRAM industry's gross margin will remain at 70%, higher than the peak levels of all historical upcycles except for 2018 (Exhibit 17).

This article is a compilation and interpretation of third-party brokerage research reports by Chao Xiang Research. The ratings, target prices, profit forecasts, and related judgments quoted in the article are the views of the brokerage's analysts and only represent the position of their respective institutions, not the views of Deep Tide TechFlow, nor do they constitute any investment advice.