Memory price inflation is becoming a new pressure point for AI capex

Overview

This report updates the models for HBM, DRAM, NAND, Samsung, SK hynix, Micron and KIOXIA. Its core view is that memory price inflation is no longer affecting only consumer electronics, but is also being transmitted into AI data centers. Bernstein believes tight supply and rapid price increases in conventional DRAM will force HBM to be repriced in the next year, while HBM embedded in GPUs/XPUs may also see markups passed through to hyperscale cloud providers, further amplifying AI infrastructure cost pressure.

Core views

First, HBM prices need to rise to close the profitability gap with conventional DRAM; due to advantages in price, bit density and yield, conventional DRAM may deliver materially higher revenue and gross margin per wafer. Second, if GPU/XPU suppliers mark up HBM costs to preserve gross margins, capex pressure on hyperscale cloud providers' AI servers and data centers will intensify. Third, despite rising cost pressure, the report believes funding availability and competitive pressure will still support continued AI investment, but the supply chain, customer pricing and weaker suppliers may face a recalibration. Fourth, Samsung, SK hynix and Micron will benefit from higher HBM and DRAM price assumptions, while KIOXIA, with only NAND exposure, is unlikely to benefit.

Analysis framework

The report cross-analyzes memory industry supply and demand, annual HBM contract pricing, unit wafer economics for conventional DRAM versus HBM, AI rack cost breakdowns, GPU/XPU supplier gross-margin pass-through, company earnings models and valuation multiples. On valuation, it shifts from the traditional P/B framework to 1-year forward P/E, because the memory leaders' ROE, cash accumulation and earnings peaks may reach historically rare levels, reducing the usefulness of book-value multiples.

Methodology notes

• Cost pass-through analysisHBM markups and AI rack capex sensitivity

  Model the impact of HBM price increases, supplier gross-margin protection and conventional DRAM/NAND price increases on AI data center capex.

  Using NVIDIA's Vera Rubin/VR200 rack as an example, the report assumes HBM is part of GPU/XPU cost; if suppliers amplify the pass-through of HBM cost increases in order to maintain a 75% gross margin, rack prices and total data center capex would rise significantly.

• Earnings forecastMemory prices and CY2027 EPS upgrades

  Update company models with HBM and conventional DRAM price assumptions.

  The report raises HBM prices by 2-2.5x and incorporates stronger-than-expected conventional DRAM prices, concluding that Samsung, SK hynix and Micron FY27 EPS will be significantly above consensus.

• Valuation method1-year forward P/E valuation

  Use a low P/E multiple to evaluate peak-cycle earnings.

  The report believes current ROE and cash accumulation may reach historically rare levels, making P/B less useful as a historical reference; therefore, it uses 6.2x 1-year forward P/E for Samsung and SK hynix, and 7.7x 1-year forward P/E for Micron.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Samsung Electronics Co Ltd

  Memory leader benefiting from higher HBM and conventional DRAM prices

  Strengths

  The report believes Samsung leads in HBM4 technology and may increase HBM share; meanwhile, highly profitable conventional DRAM can support overall earnings.

  Weaknesses

  Greater HBM exposure does not necessarily mean the highest margins, because conventional DRAM may be more profitable on a per-wafer basis.

  Comparison

  Samsung is one of the main beneficiaries alongside SK hynix and Micron, but may stand out more in HBM4 technology progress.

  Risks

  An early end to the favorable memory price environment, weakening demand, rising supply, changing investor valuation sentiment and progress in China's memory industry.

• SK hynix Inc

  Beneficiary of the HBM and DRAM upcycle

  Strengths

  The report raises earnings forecasts and the target price, believing SK hynix benefits from higher HBM and DRAM prices.

  Weaknesses

  After the cycle peaks, if CY28 prices normalize, revenue may decline year over year.

  Comparison

  Like Samsung, it is valued at 6.2x 1-year forward P/E, and the report implies about 20% upside.

  Risks

  DRAM price pullback, supply expansion, and market discounts on peak earnings.

• Micron Technology Inc

  U.S. memory leader benefiting from higher DRAM and HBM price assumptions

  Strengths

  The report expects Micron FY27 EPS to be about 38% above consensus and raises the target price to US$1,300.

  Weaknesses

  The valuation depends on sustained high profitability, and CY28 price normalization could cause revenue to decline.

  Comparison

  Valued at 7.7x 1-year forward P/E, above Samsung's and SK hynix's 6.2x.

  Risks

  An earlier-than-expected memory cycle reversal, demand below expectations, faster-than-expected supply recovery and valuation compression.

• KIOXIA Holdings Corp

  Pure NAND supplier, unlikely to benefit from HBM upgrades

  Strengths

  If NAND demand proves stronger than expected, NAND prices improve, or Japanese policy support emerges, upside risk could materialize.

  Weaknesses

  It lacks HBM exposure, so it cannot enjoy the earnings revision from this HBM price increase cycle.

  Comparison

  Relative to Samsung, SK hynix and Micron, KIOXIA is in a less favorable position under the report's core HBM thesis.

  Risks

  Weak NAND demand, cost declines falling short of expectations, and policy support below expectations.

• MediaTek Inc

  May benefit from hyperscale cloud providers directly purchasing HBM and from the Asian ASIC services model

  Strengths

  The report believes its first and second TPU projects are executing well, and supply-chain checks indicate upside risk to 2028 forecasts.

  Weaknesses

  The stock has already risen about 130% in the past two months, so near-term expectations may be high.

  Comparison

  Unlike GPU/XPU suppliers, MediaTek's business model may allow hyperscale cloud providers to buy HBM directly to avoid markups.

  Risks

  Stronger Qualcomm competition in the 5G market, slowing 5G penetration, weaker global smartphone demand, slower diversification beyond smartphones, and a broad semiconductor downcycle.

• Hyperscalers and AI data center operators

  The main parties under pressure from rising AI memory costs

  Strengths

  Funding availability and continued competitive pressure in AI still support investment.

  Weaknesses

  HBM, conventional DRAM and NAND price increases may sharply raise AI project capex budgets, requiring ROI recalculations.

  Comparison

  Relative to upstream memory leaders, hyperscale cloud providers are more likely to bear the burden of cost redistribution.

  Risks

  Capex overruns, lower project ROI, supplier markups, and eventual pressure to adjust token pricing or customer pricing.

Key data

• HBM Price Forecast2-2.5x YoY increaseUsed to narrow the profitability gap between HBM and conventional DRAM.
• Conventional DRAM Price Changeroughly 4.5x from 3QCY25 to 2QCY26, with another ~25% possible in CY27The report believes DRAM prices may peak in CY27.
• AI Data Center Capex Impact~+30%Includes HBM price increases, possible HBM markups, and conventional DRAM and NAND price increases.
• FY27 EPS vs. Consensus25%-38% aboveCovers the earnings upgrade view for Samsung, SK hynix and Micron.
• Samsung Target PriceKRW440,000Based on 6.2x 1-year forward P/E, implying about 26% upside according to the report.
• SK hynix Target PriceKRW3,300,000Based on 6.2x 1-year forward P/E, implying about 20% upside according to the report.
• Micron Target PriceUS$1,300Based on 7.7x 1-year forward P/E, implying about 15% upside according to the report.
• KIOXIA Target Price¥40,000Underperform is maintained because it is mainly exposed to NAND and does not benefit from HBM.
• MediaTek Target PriceNT$4,380The report believes that if hyperscale cloud providers want to source HBM directly to avoid markups, Asian ASIC service models may benefit.

Impact & implications

For investors, the key takeaway is that the memory upcycle may continue to drive earnings and share prices for the leaders, but cost allocation across the AI ecosystem is becoming more complex. Higher HBM and conventional DRAM prices benefit Samsung, SK hynix and Micron, which have high-end memory exposure, while also potentially forcing cloud providers to recalculate the ROI of AI projects and redistribute costs among GPU/XPU suppliers, memory suppliers, ASIC service providers and end customers. If cloud providers shift to directly purchasing HBM to avoid markups, Asian ASIC service providers such as MediaTek that can support this model may benefit.

Risks

• The favorable memory price environment may end early if demand weakens or supply increases.
• It is uncertain whether AI accelerator suppliers will markup HBM costs; if costs are simply passed through, capex pressure would be lower than the report's high-case estimate.
• Hyperscale cloud providers may use a mix of accelerators, CPU servers or different memory configurations, meaning a single VR200 rack assumption may not fully represent actual deployments.
• As more cleanrooms and capacity come online, DRAM and NAND prices may normalize in CY28, causing revenue to pull back.
• China's progress in memory, especially NAND, may create downside risk to industry pricing and competition dynamics.
• MediaTek faces risks from Qualcomm competition, slowing 5G penetration, weaker smartphone demand and a semiconductor downcycle.

What to watch

• The outcome of CY27 annual HBM pricing negotiations over the next few months.
• Whether Bloomberg and other market consensus estimates for Samsung, SK hynix and Micron CY2027/FY2027 EPS are revised upward.
• Whether conventional DRAM prices continue to rise and peak around CY27.
• Whether GPU/XPU suppliers markup HBM costs or allow cloud providers to purchase HBM directly.
• Whether cloud providers adjust AI data center ROI models, procurement methods or token pricing.
• Samsung HBM4 shipments, Korea export value/volume indicators and changes in HBM share.
• The speed at which DRAM and NAND prices normalize after new CY28 capacity comes online.