Author: Xiao Bing
HBM is still locked into annual contract prices, while regular DRAM has surged 4.5 times. From the same wafer fab, the profit from making regular memory is twice that of HBM, with revenue twice as high and gross margin nearly three times higher. This means next year's HBM prices must double to 2.5 times, otherwise no memory manufacturer will allocate production capacity to it. The problem is, HBM is soldered into NVIDIA GPUs and sold as a package. Once HBM prices increase, to maintain its 75% gross margin, NVIDIA will pass the price hike onto cloud vendors by amplifying it by a factor of 4.
In the global memory research report released on June 22 by Bernstein's Asia tech team led by Mark Li, they maintained an 'Outperform' rating on Samsung, SK Hynix, and Micron, significantly raising target prices: Samsung from 225,000 won to 440,000 won, SK Hynix from 1.15 million won to 3.3 million won, Micron from $510 to $1,300. They maintained an 'Underperform' rating on Kioxia, with the target price unchanged at 40,000 yen. They maintained an 'Outperform' rating on MediaTek with a target price of NT$4,380.
The underlying logic of this report is that the memory industry is undergoing an unprecedented structural split.
The Value of a Wafer Is Being Redefined
From Q3 2025 to Q2 2026, regular DRAM prices rose by about 4.5 times. HBM, tied to annual long-term contracts, saw almost no price movement. As a result, Bernstein calculates that in 2026, allocating capacity to regular DRAM generates over twice the revenue per wafer compared to HBM, with gross margin nearly three times higher.
Samsung and Micron have clearly stated on their Q1 2026 earnings calls: the profit margin for non-HBM DRAM has already surpassed that of HBM, and as regular DRAM prices continue to rise, this gap is widening. Bernstein expects regular DRAM prices may have about 25% more upside in 2027 before peaking.
This presents a stark set of numbers for HBM procurement negotiations: To match the revenue per wafer of regular DRAM, HBM prices would need to triple. However, memory makers are also aware that HBM is a critical component of AI infrastructure, and overly aggressive pricing could harm the health of the entire AI ecosystem. SK Hynix stated on its call that it would "prioritize achieving optimal allocation between HBM and regular DRAM," not pursue revenue maximization.
After considering these factors, Bernstein's judgment is: HBM prices will increase by an average of 2 to 2.5 times for the full year 2027 (Exhibit 1-2). Even so, HBM's profitability will remain lower than that of regular DRAM, though the gap will narrow significantly compared to 2026.
The real impact of HBM price increases is hidden in NVIDIA's markup.
Cloud vendors can buy regular DRAM and NAND directly from memory makers. HBM is different; it is packaged within NVIDIA's GPUs and is part of the latter's Cost of Goods Sold (COGS).
Assuming NVIDIA maintains a 75% gross margin on the VR200 (Vera Rubin NVL72) rack, the portion of the HBM price increase would need to be amplified by a factor of 4 in NVIDIA's pricing. Bernstein's estimation logic: HBM originally accounted for about 5% of the VR200's selling price; after the HBM price increase, this rises to 6%. But if NVIDIA is to keep its 75% gross margin unchanged, the selling price of the rack would need to increase by 24%.
For an AI data center deploying VR200 racks, the pass-through cost of HBM alone would increase total capital expenditure (including costs outside the rack) by 4%-15%, depending on whether NVIDIA applies a markup. Combined with price increases for regular DRAM and NAND (about 14%), all together, cloud vendors' AI capital expenditure would need to be about 30% higher than originally planned (Exhibit 3).
The report terms this process "re-calibration," believing cloud vendors will not slow down AI investment because of it, but will inevitably spread the cost pressure across supply chain segments, possibly even reflected in charging different token prices to different customers.
Profit Revision Wave Is Approaching, Who Wins and Who Loses
Bernstein raised its 2027 average HBM price assumption by 2 to 2.5 times, leading to profit forecasts significantly above market consensus: Samsung's 2027 Earnings Per Share (EPS) is 26% higher than consensus, SK Hynix 32% higher, Micron 38% higher (Exhibits 11-13). The analyst believes the annual HBM negotiations will conclude over the coming months, leading to upward revisions in sell-side consensus, further driving stock prices higher.
The HBM price hike is not pure positive for memory makers. Bernstein specifically points out that greater HBM exposure means lower overall profitability because regular DRAM margins are exceptionally high. Samsung is leading in HBM4 technology; Korea memory export data monitored by the report also shows Samsung's export unit value rose significantly in May, hinting at HBM4 shipments beginning (Exhibit 8). But Samsung has also expressed a pursuit of higher profit margins, potentially allocating more capacity to regular DRAM over HBM.
Kioxia is the sole loser, as it only has NAND business, no HBM, and thus cannot benefit from the profit revisions brought by this round of HBM price increases.
MediaTek could become another type of beneficiary. The report believes if cloud vendors, to avoid NVIDIA's markup, demand to directly purchase HBM, the business model of ASIC (Application-Specific Integrated Circuit) service providers perfectly fits this demand. MediaTek's execution on TPU projects has been solid, and supply chain checks suggest upside risk for 2028 outlook. The stock has risen about 130% in the past two months, but Bernstein still maintains its 'Outperform' rating.
Valuation Switch to P/E, 15%-26% Upside Remains for Target Prices
Since the Return on Equity (ROE) of memory makers will reach unprecedented levels in this cycle, Samsung 55%, SK Hynix 108%, Micron 85% (Exhibit 18), and cash accumulation is astonishing (cash could account for 70%-80% of book value by 2027), the traditional Price-to-Book (P/B) valuation method has lost relevance. Bernstein has switched to using a 1-year forward Price-to-Earnings (P/E) ratio, setting the target multiple near historical cycle lows: Samsung and SK Hynix 6.2x, Micron 7.7x.
Corresponding target prices are: Samsung 440,000 won (26% upside), SK Hynix 3.3 million won (20% upside), Micron $1,300 (15% upside).
For 2028, the report believes that as more cleanroom capacity comes online, memory prices will soften, and revenue for all three vendors will decline year-on-year. However, even in the downcycle of 2028, the DRAM industry's gross margin is still projected at 70%, higher than the peak levels of all historical upcycles except 2018 (Exhibit 17).
This article is a collation and interpretation by Tide Research of a third-party brokerage research report. The ratings, target prices, profit forecasts, and related judgments cited herein are the views of that brokerage's analyst, representing only the stance of their affiliated institution. They do not represent the views of DeepChain TechFlow and do not constitute any investment advice.
