Storage becomes an AI burden—what does this mean for investors?

The overall rise in storage chip prices is spreading cost pressures from the consumer side to AI infrastructure, forcing hyperscale cloud computing vendors to reassess their capital expenditure return models.

According to Chasewind Trading Desk and Bernstein Research’s latest report, HBM (High Bandwidth Memory) prices are expected to rise 2 to 2.5 times next year. Combined with substantial increases in traditional DRAM and NAND prices, the capital expenditures for hyperscale cloud vendors deploying AI data centers may rise by about 30%.

The transmission path of this cost shock is not simple. Take Nvidia as an example: HBM is packaged within its GPU products, and if Nvidia wants to maintain its 75% gross margin, it will need to amplify the HBM cost increase about fourfold and pass it on to customers, further increasing the capital pressure for cloud vendors. Bernstein believes that this pressure will not slow hyperscale cloud vendors' AI investment, but a "cost rebalance" across supply chain stages is inevitable, and weaker suppliers may be squeezed in this process.

For investors, the direct impact of this pattern is: the per share earnings forecasts for Samsung, SK Hynix, and Micron for fiscal year 2027 will see significant upgrades, with Bernstein’s forecasts currently 25% to 40% above consensus; pure NAND supplier Kioxia will miss out on this round of upside.

HBM Pricing Inversion, Price Increase Pressures Mount

The core logic behind the HBM price rise lies in the widening profit gap between HBM and traditional DRAM.

According to Bernstein’s estimates, from Q3 2025 to Q2 2026, traditional DRAM prices will have risen about 4.5 times cumulatively, whereas HBM prices, constrained by annual contract lock-ins, have essentially stayed flat. This means that in 2026, the per wafer revenue from traditional DRAM production will be about twice that of HBM, and gross profit nearly three times higher.

Samsung clearly stated during its Q1 2026 earnings call that non-HBM DRAM profit margins have surpassed HBM, and with traditional DRAM prices still rising, the gap keeps widening. Samsung also hinted that it expects the margin gap to “narrow significantly” in 2027, implying HBM price increases are inevitable. SK Hynix also emphasized that it will optimize capacity allocation between HBM and regular DRAM, rather than simply pursue maximum revenue.

Based on these considerations, Bernstein forecasts HBM prices will rise 2 to 2.5 times next year—less than the threefold increase needed for HBM revenue to catch up with traditional DRAM—reflecting vendors’ recognition of HBM’s strategic value and concerns that excessive pricing could harm the overall AI ecosystem. Even so, HBM profitability will still lag traditional DRAM next year, but the gap will narrow significantly.

Cost Amplification: The "Markup Multiplier" of GPU Vendors

The impact of HBM price hikes on hyperscale cloud vendors goes far beyond the increase in direct procurement costs.

Bernstein uses Nvidia’s Vera Rubin (VR200) rack as an example: before the HBM price hike, HBM accounted for about 5% of the VR200 price. If HBM prices rise 2 to 2.5 times, this alone would increase VR200’s price by around 6%. But if Nvidia, to maintain a 75% gross margin, amplifies the HBM price increase fourfold and passes it on, the rack price could rise by about 24%.

In sum, if only direct HBM cost transmission is considered, hyperscale cloud vendors' total AI data center capital expenditures will increase by about 4%; factoring in Nvidia's full markup, the increase expands to about 15%. Additionally, the sharp rises in traditional DRAM and NAND prices will contribute another 14% increase. Adding these three, hyperscale cloud vendors need to raise AI capital expenditures by about 30% to cover the higher storage costs.

Bernstein notes that GPU and accelerator vendors may argue they do not mark up HBM, treating it as “pass-through” revenue and emphasizing the integrated value of HBM with logic chips, packaging, and software. Regardless, without marking up prices, HBM’s rising costs will directly compress GPU vendors’ gross margins, creating strong pricing motivation.

Hyperscale Cloud Vendors Face a "Cost Rebalance"

Facing about 30% higher capital expenditures, hyperscale cloud vendors must recalculate their AI investment return models.

Bernstein believes competition and access to capital will support maintaining AI investment, but “cost rebalance” is inevitable. This process may include multiple dimensions: pressuring supply chain stages to share costs, and potentially adjusting the price of computing services billed to different customers. Weaker suppliers are at risk of being squeezed, while those within Bernstein’s coverage are well positioned, with some likely to benefit.

Among them, MediaTek is seen as a potential beneficiary. Bernstein points out that if hyperscale cloud vendors bypass GPU vendors’ HBM markup by purchasing HBM directly, Asian ASIC service providers’ business models are well-suited for this demand, and MediaTek stands to benefit. The stock has surged approximately 130% over the past two months, and Bernstein maintains an “outperform” rating with a target price of NT$4,380.

Earnings Forecasts Shift Up Sharply, Target Prices Raised for Samsung, SK Hynix, Micron

The expected rise in HBM prices hasn’t been fully reflected in sell-side consensus forecasts, meaning a wave of significant earnings upgrades is imminent.

Bernstein has raised its HBM price forecast by 2–2.5 times, along with similar upgrades for traditional DRAM and NAND prices, resulting in fiscal 2027 per share earnings forecasts 25%–40% higher than consensus: Samsung up about 26%, SK Hynix up about 32%, Micron up about 38%. As HBM pricing negotiations are resolved in the coming months, Bernstein expects consensus forecasts to be revised upward, supporting the stock prices of all three companies.

For valuation, Bernstein has switched the basis for these three companies from price-to-book to price-to-earnings, as current profit margins and return on equity will reach unprecedented levels, making historical price-to-book less relevant. Using near-historic-bottom PE multiples for cycle peak earnings, Bernstein raises Samsung’s target price to 440,000 KRW (6.2 times 1-year forward PE, about 26% upside), SK Hynix’s to 3,300,000 KRW (also 6.2 times, about 20% upside), Micron to $1,300 (7.7 times, about 15% upside), with all three rated “outperform”.

Notably, the earnings upgrade does not benefit pure NAND suppliers. Kioxia, which does not engage in HBM business, cannot benefit from the HBM price increase. Bernstein maintains an “underperform” rating, with a target price of 40,000 JPY.

~~~~~~~~~~~~~~~~~~~~~~~~

Contents above courtesy of Chasewind Trading Desk.

For more detailed analysis, including real-time interpretation and frontline research, please join [Chasewind Trading Desk Annual Membership]

Risk Warning and DisclaimerThe market involves risks; investments require caution. This article does not constitute personalized investment advice and does not address any individual user's specific investment objectives, financial situation, or needs. Users should consider whether any opinion, viewpoint, or conclusion in this article applies to their particular circumstances. Investing based on this is at your own risk.