Bernstein Research Report Analysis: HBM Prices Set to More Than Double Next Year; Memory Becomes a Burden for AI
TechFlow Selected TechFlow Selected
Bernstein Research Report Analysis: HBM Prices Set to More Than Double Next Year; Memory Becomes a Burden for AI
Long-term HBM contracts with fixed pricing have resulted in significantly lower profitability compared to standard DRAM. Bernstein forecasts HBM prices will rise 2–2.5x by 2027 and has substantially raised its target prices for Samsung, SK Hynix, and Micron—triggering a structural revaluation across the memory industry.
Navigating Web3 tides with focused insightsBy: Xiao Bing
HBM prices remain locked under annual contracts, while standard DRAM prices have surged 4.5x. At the same wafer fab, producing standard memory generates twice the revenue—and nearly three times the gross margin—of producing HBM. This implies that HBM prices must rise 2–2.5x next year, or memory manufacturers will have no incentive to allocate capacity to it. The problem, however, is that HBM is soldered directly onto NVIDIA GPUs and sold as a bundled component; if HBM prices rise, NVIDIA—seeking to maintain its 75% gross margin—will amplify that price increase fourfold before passing it on to cloud providers.
In its global memory research report released on June 22, Bernstein’s Asia Tech team, led by Mark Li, maintained its Outperform rating for Samsung, SK hynix, and Micron—and significantly raised target prices: Samsung from ₩225,000 to ₩440,000; SK hynix from ₩1,150,000 to ₩3,300,000; and Micron from $510 to $1,300. Kioxia was kept at Underperform with its target price unchanged at ¥40,000. MediaTek retained its Outperform rating, with its target price raised to NT$4,380.
The underlying thesis of this report is that the memory industry is undergoing an unprecedented structural split.
On the same wafer, “what you make determines what you earn”—and that rule is being rewritten
From Q3 2025 through Q2 2026, standard DRAM prices rose roughly 4.5x. In contrast, HBM prices remained virtually flat due to binding annual contracts. As a result, Bernstein estimates that allocating capacity to standard DRAM instead of HBM will generate over double the revenue per wafer—and nearly triple the gross margin—in 2026.
Samsung and Micron explicitly stated in their Q1 2026 earnings calls that non-HBM DRAM profitability has already surpassed that of HBM—and that the gap is widening further as standard DRAM prices continue rising. Bernstein forecasts standard DRAM prices may still climb ~25% in 2027 before peaking.
This creates a stark set of numbers for HBM procurement negotiations: to match standard DRAM’s revenue per wafer, HBM prices would need to rise 3x. Yet memory makers are also aware that HBM is a critical AI infrastructure component; pricing too aggressively risks undermining the broader AI ecosystem’s healthy development. SK hynix stated on its earnings call that it will “prioritize optimal allocation between HBM and standard DRAM,” rather than maximizing revenue.
Factoring in all these considerations, Bernstein projects HBM average prices will rise 2–2.5x across 2027. Even so, HBM profitability will still lag behind standard DRAM—though the gap will narrow substantially compared to 2026.
The real impact of HBM price hikes lies buried within NVIDIA’s markup
Cloud providers can purchase standard DRAM and NAND directly from memory manufacturers. But HBM is different—it is packaged inside NVIDIA GPUs and forms part of NVIDIA’s cost of goods sold (COGS).
Assume NVIDIA maintains its 75% gross margin on its VR200 (Vera Rubin NVL72) rack. To absorb HBM price increases, NVIDIA must inflate its selling price fourfold. Bernstein’s calculation logic: HBM originally accounts for ~5% of the VR200’s selling price; after the HBM price hike, that share rises to ~6%. Yet to preserve its 75% gross margin, the rack’s overall selling price must increase by 24%.
For an AI data center deploying VR200 racks, the pass-through of HBM cost alone could raise total capital expenditure (including costs beyond the rack) by 4%–15%, depending on whether NVIDIA applies the full markup. Combined with standard DRAM and NAND price hikes (~14%), total AI capex for cloud providers could rise ~30% versus prior expectations.
The report terms this process “re-calibration,” arguing that cloud providers won’t slow AI investment—but will inevitably redistribute cost pressure across supply chain tiers, potentially even manifesting in differentiated token pricing for different customers.
A wave of earnings revisions is approaching—who benefits, who loses?
Bernstein raised its 2027 HBM average price assumption by 2–2.5x, resulting in significantly higher earnings forecasts than market consensus: Samsung’s 2027 EPS is projected 26% above consensus; SK hynix’s, 32% above; and Micron’s, 38% above. Analysts expect HBM annual negotiations to conclude progressively over the coming months, driving upward revisions in sell-side consensus—and further upside for stock prices.
However, HBM price hikes aren’t purely positive for memory makers. Bernstein specifically notes that greater HBM exposure translates into lower overall profitability, given standard DRAM’s exceptionally high margins. Samsung leads in HBM4 technology, and Korea’s memory export data—monitored in the report—shows Samsung’s unit export value surged notably in May, suggesting HBM4 shipments have begun. Yet Samsung itself has signaled pursuit of higher margins, potentially diverting more capacity toward standard DRAM instead of HBM.
Kioxia stands as the sole loser, as it operates only in NAND and lacks any HBM business—thus missing out entirely on this round of HBM-driven earnings revisions.
MediaTek may emerge as another beneficiary. The report argues that if cloud providers seek to bypass NVIDIA’s markup by purchasing HBM directly, ASIC (application-specific integrated circuit) service providers’ business model is well positioned to meet that demand. MediaTek’s execution on TPU projects has been steady, and supply-chain checks indicate upside risk to its 2028 outlook. The stock has risen ~130% over the past two months—but Bernstein maintains its Outperform rating.
Valuation shifts to P/E; target prices still hold 15%–26% upside
Given that memory makers’ return on equity (ROE) will reach unprecedented levels during this cycle—Samsung at 55%, SK hynix at 108%, and Micron at 85%—and cash accumulation is accelerating rapidly (cash holdings expected to reach 70%–80% of book value by 2027), traditional price-to-book (P/B) valuation has lost relevance. Bernstein instead adopts one-year forward price-to-earnings (P/E) valuation, setting target multiples near historical cycle troughs: 6.2x for Samsung and SK hynix, and 7.7x for Micron.
Corresponding target prices are: Samsung at ₩440,000 (26% upside), SK hynix at ₩3,300,000 (20% upside), and Micron at $1,300 (15% upside).
For 2028, the report expects storage prices to soften as additional cleanrooms come online, leading to YoY revenue declines across all three companies. Nevertheless, even amid this cyclical downturn, DRAM industry gross margins will remain at 70%—higher than peak levels in all prior upcycles except 2018.
This article is TechFlow’s summary and interpretation of a third-party brokerage research report. Ratings, target prices, earnings forecasts, and related judgments cited herein reflect the views of the brokerage’s analysts only, represent positions held by their respective institutions, and do not reflect TechFlow’s views nor constitute any investment advice.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@chaoxiangooo
