Bernstein Research Report Interpretation: Next year, HBM prices will definitely double or more, and storage will become a burden for AI

By: Xiao Bing

HBM is still locked in annual contract pricing, while ordinary DRAM has already risen by 4.5 times. In the same wafer fabrication plant, the profit earned from producing ordinary memory is 2 times that of producing HBM; revenue is 2 times, and gross profit is nearly 3 times. This means that next year’s HBM prices must increase by 2–2.5 times; otherwise, no memory manufacturer will be willing to allocate capacity to it. The issue is that HBM is bundled and sold together inside NVIDIA GPUs. Once HBM prices rise, NVIDIA will amplify the price increase by 4 times and pass it on to cloud providers in order to maintain its 75% gross margin.

On June 22, Mark Li from Bernstein’s Asia Technology team published a global memory research report. It maintained Outperform ratings for Samsung, SK Hynix, and Micron, and significantly raised target prices: Samsung raised its target from 225,000 KRW to 440,000 KRW; SK Hynix from 1.15 million KRW to 3.3 million KRW; and Micron from $510 to $1,300. KIOXIA (KIOXIA) was maintained at Underperform, with its target price of 40,000 JPY unchanged. For MediaTek, the rating was maintained as Outperform, with a target price of 4,380 TWD.

The underlying logic of this report is that the memory industry is undergoing an unprecedented structural split.

Rewriting what earns profit on the same wafer

From Q3 2025 to Q2 2026, the price of ordinary DRAM increased by about 4.5 times. Meanwhile, HBM, bound by annual long-term contracts, saw almost no price movement. As a result, Bernstein estimates that in 2026, allocating capacity to ordinary DRAM will generate more than 2 times the revenue per wafer compared with HBM, with gross profit approaching nearly 3 times.

In their Q1 2026 earnings call, Samsung and Micron have already stated clearly that the profit margin of DRAM that is not HBM has surpassed that of HBM, and that this gap will continue to widen as ordinary DRAM prices keep rising. Bernstein expects ordinary DRAM prices could still rise by about 25% in 2027 before reaching their peak.

This creates a set of sharp numbers for HBM procurement negotiations: in order for HBM revenue per wafer to catch up with ordinary DRAM, HBM prices would need to rise by 3 times. But memory makers also know that HBM is a key component of AI infrastructure, and overly aggressive pricing would damage the healthy development of the entire AI ecosystem. SK Hynix stated on the call that it would “prioritize achieving the optimal configuration between HBM and ordinary DRAM,” and would not pursue maximizing revenue.

After aggregating these factors, Bernstein’s judgment is that in 2027, the average HBM price would rise by 2–2.5 times for the full year (Exhibit 1–2). Even so, HBM profitability will still be lower than that of ordinary DRAM, but the gap will narrow significantly compared with 2026.

The real impact of HBM price hikes is hidden in NVIDIA’s markup

Cloud providers can purchase ordinary DRAM and NAND directly from memory manufacturers. But HBM is different: it is packaged inside NVIDIA 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) chassis, the portion of the HBM price increase would require NVIDIA to amplify pricing by 4 times. Bernstein’s estimation logic is as follows: HBM originally accounts for about 5% of the VR200 selling price; after HBM prices rise, this proportion becomes 6%. But if NVIDIA wants to keep the 75% gross margin unchanged, the increase in the chassis selling price would need to reach 24%.

For an AI data center deploying VR200 chassis, passing through only the HBM cost would increase total capital expenditure (including costs outside the chassis) by 4%–15%, depending on whether NVIDIA applies the price increase. When combined with the roughly 14% increase in ordinary DRAM and NAND prices, all of these together mean that cloud providers’ AI capex needs to be about 30% higher than before (Exhibit 3).

The report refers to this process as “re-calibration.” It believes cloud providers will not slow down AI investment as a result, but they will inevitably share the cost pressure across each link in the supply chain, and it may even show up in charging different token prices to different customers.

A profit correction wave is approaching—who benefits and who is harmed

Bernstein raised its 2027 average HBM price assumption by 2–2.5 times, which corresponds to profit forecasts that are far above market consensus: Samsung’s 2027 earnings per share (EPS) is 26% higher than consensus, SK Hynix’s is 32% higher, and Micron’s is 38% higher (Exhibit 11–13). Analysts believe HBM annual negotiations will wrap up over the next few months, and sell-side consensus will be revised upward, which in turn will push the stock prices higher.

HBM price increases are not purely positive for memory manufacturers. Bernstein specifically points out that a larger HBM exposure implies lower overall profitability because the profit margins of ordinary DRAM are simply too high. Samsung is leading in HBM4 technology, and the South Korean memory export data monitored in the report also shows that Samsung’s export unit value in May rose significantly, implying that HBM4 has started shipping (Exhibit 8). However, Samsung also expressed a desire to pursue higher profit margins, and may allocate more capacity to ordinary DRAM rather than HBM.

KIOXIA (KIOXIA) is the only loser because it has only the NAND business and no HBM, so it cannot benefit from the profit revision brought by this round of HBM price increases.

MediaTek could be another beneficiary. The report believes that if cloud providers, to avoid NVIDIA’s markup, require direct procurement of HBM, then the business model of ASIC (application-specific integrated circuit) service providers would be well positioned to take advantage of this demand. MediaTek’s execution on the TPU project is solid, and supply-chain research indicates there is upside risk in the 2028 outlook. The stock has risen by about 130% over the past two months, but Bernstein still maintains an Outperform rating.

Valuation method switches to P/E, with target prices still having 15%–26% upside

Because in this cycle the return on equity (ROE) of memory manufacturers’ net assets will reach an unprecedented level—Samsung at 55%, SK Hynix at 108%, and Micron at 85% (Exhibit 18)—and cash accumulation is astonishing (in 2027, cash could account for 70%–80% of book value), the traditional price-to-book (P/B) valuation method has lost reference value. Bernstein instead uses a 1-year forward price-to-earnings (P/E) ratio and sets target multiples near historical cycle troughs: Samsung and SK Hynix at 6.2x, and Micron at 7.7x.

The corresponding target prices are: Samsung at 440,000 KRW (up 26%), SK Hynix at 3.3 million KRW (up 20%), and Micron at $1,300 (up 15%).

For 2028, the report believes that as more cleanrooms come online and begin production, memory prices will soften and the revenues of the three companies will decline year over year. However, even in 2028 during a downturn in the cycle, the DRAM industry’s gross margin will still be around 70%, higher than the peak levels of all prior upcycles except 2018 (Exhibit 17).

This article is a compilation and interpretation of research reports from third-party brokerage firms by ChaoXiang Research. The ratings, target prices, earnings forecasts, and related judgments cited in the article are solely the views of the analysts at that brokerage, and only represent the stance of their respective institution. They do not represent the views of DeepTech TechFlow, nor do they constitute any investment advice.