全球内存：内存会成为AI的负担吗？

Global Memory: Memory becoming a burden of AI too? We update our memory & industry models (models forHBM,DRAM,NAND,Samsung,SKhynix,Micron&KIOXIA) & maintain Outperform ratings on Samsung, SK hynix & Micron &Underperform on KIOXIA. Mark Li+852 2123 2645mark.li@bernsteinsg.com HBM price needs to go higher to narrow the profitability gap vs. conventional DRAM.From 3QCY25 to 2QCY26, conventional DRAM price has risen c. 4.5x, but HBM price islocked by annual contracts and hasn’t moved. As the result, we estimate in CY26 deployingcapacity to conventional DRAM will generate over 2x revenue & nearly 3x gross profit dollarper wafer capacity vs. HBM. This is why memory suppliers and GPU/XPU companies arenegotiating CY27 HBM price now to narrow the gap. Edward Hou, CFA+852 2123 2623edward.hou@bernsteinsg.com Yipin Cai, CFA+852 2123 2669yipin.cai@bernsteinsg.com The possible “markup” of GPU/XPU suppliers will amplify the cost burden onhyperscalers.Unlike conventional DRAM & NAND that hyperscalers can source directlyfrom memory suppliers, HBM is packaged in GPUs/XPUs and is part of the COGS of, forexample, NVIDIA. NVIDIA needs to mark up the HBM price hike 4x if NVIDIA has 75% grossmargin & wants to keep 75% unchanged, despite the HBM price hike. This will make theHBM price hike even a heavier burden for hyperscalers. With memory making hyperscalers’ capex 30% higher, a “re-calibration” isinevitable and may squeeze weaker suppliers.Using Vera Rubin (VR, NVL72) rack asan example, hyperscalers will find their capex higher by 30% if they deploy a data centerwith this rack and with the same memory capacity installed in the rack, because of higherconventional DRAM & NAND cost, higher HBM cost & the markup charged by NVIDIA. Withfunding availability and competitive pressure, we think hyperscalers will still invest in AI, butmay try to “re-calibrate” their costs across different component suppliers (& possibly eventoken prices across different customers too?). To reflect higher HBM price, a notable earnings revision is coming to support thestock prices of Samsung, SK hynix Micron but not KIOXIA.We raise HBM price 2-2.5x,not as much as the size of conventional DRAM price hike, as we assume memory suppliersrecognize the strategic value of HBM and choose to price less aggressively. We update ourforecast & find our FY27 EPS 25-40% above consensus as the result. As HBM negotiationgradually concludes in the next few months, we expect consensus to move upward and toprovide an upward support to the stock of Samsung, SK hynix and Micron, but not KIOXIAas KIOXIA has only NAND & does not have HBM. Please note this also means higher HBMexposure will result lower profitability for memory suppliers. MediaTek may benefit if hyperscalers prefer to source HBM directly to avoid thepossible markup. The stock went up c. 130% in the past 2 months but we reiterateOutperform. Reiterate Outperform on Samsung, SK hynix and Micron too.We raise target priceto KRW440,000 for Samsung (6.2x 1-year forward P/E), KRW3,300,00 for SK hynix( 6.2x 1-year forward P/E), and US$1,300 for Micron (7.7x 1-year forward P/E). MaintainUnderperform on KIOXIA (see KIOXIA: All parties come to an end - Underperform). BERNSTEIN TICKER TABLE PRICE TARGET CHANGE / ESTIMATE CHANGE IN BOLD INVESTMENT IMPLICATIONS Samsung, SK hynix, Micron:We maintainOutperformon Samsung, SK hynix and Micron and raise price targets toKRW440,000 (vs 225,000), KRW3,300,000 (vs 1,150,000)andUS$1,300 (vs 510)respectively, driven by increasedforecasts of conventional memory pricing as well as HBM pricing going forward. We now value them using 1-year forward P/Eand target6.2xfor Samsung,6.2xfor SK hynix and7.7xfor Micron. DETAILS Please feel free to download our industry (DRAM,NAND,CoWoW/HBM) and company models (Samsung,SK hynix,Micron&KIOXIA). We think HBM price will be higher. And added with the "markup" on HBM, the increase will make AI feel theburden of memory price hike too. A “re-calibration” to share the cost burden is probably inevitable. •As a result of the DRAM supply shortage, and skyrocketing conventional DRAM prices, we think HBM price is also boundto go up next year, by 2-2.5x vs this year (Exhibit 1-Exhibit 2), and will exert a heavy pressure on server rack cost and alsohyperscaler capex. The cost pressure possibly won’t be limited to just the higher purchase cost of HBM stacks paid tomemory suppliers. Take NVIDIA as an example. HBM is packaged inside NVIDIA’s GPUs, and the HBM cost paid memorysuppliers is part of NVIDIA’s COGS. Simply passing through this HBM cost increase to NVIDIA’s customers will diluteNVIDIA’s gross margin. Assume NVIDIA earns 75% gross margin on the VR200 racks, NVIDIA will need to increase its rackprice by 4x the cost increase of HBM in order to maintain margin at 75%. •We estimate before the HBM price increase HBM was 5% of VR200 price that NVIDIA charges. With the HBM price increasealone, the selling price of VR200 needs to be higher by 6% pric