亚洲半导体与全球存储器：2026年仍将围绕AI展开

Asia Semiconductors & Global Memory: 2026 is still all about AI 2026 is still about AI.Sufficient visibility shows that 2026 demand for AI accelerators up. But unfortunately 2026 marks the 4thyear of AI investments & investors are asking if AIis a bubble. We see an upside risk to earnings forecasts, and hope that OpenAI & Anthropicwill release positive news to drum up their IPOs which reportedly are scheduled for 2H26.Unable to find other theses meaningful, we recommend to stay in AI but focus on quality. TPU is gaining steam but NVIDIA’s GPU remains the majority, in terms of CoWoSsupply. Amkor & ASE (both not covered) will offer some, but limitedly, CoWoS capacityhence should be still a key bottleneck in 2026. Broadcom is set to supply most of TPUs shipment in 4Q26 should allow NIVIDIA to mount a counter-attack. Leading-logic capacity will be tight, but TSMC should still stay capex disciplined.2026 capacity has been fully booked, & N2 capacity will be tight in 2027 too. We model Mature nodes may see mixed demand, & high memory price squeezing end demandis an additional headwind. debate. We model CoWoS/HBM shipment to rise 73%/98% YoY & the capacity build-upshould present an investment opportunity. Geopolitically, leading logic has become 2 distinct investment markets & investorsshould be mindful of memory IPOs from China.SMIC/Huawei just rolled out “N+3”node but all indicators suggest it’s not as good as N5 & China still lags TSMC by 5 years, orlonger if scale & yield are considered. Threat from China, however, will be still a controversyfor mature nodes. YMTC & CXMT reportedly will go public this year & YMTC’s expansion without EUV is why we’re structurally negative on pure-NAND suppliers.As for companies - Margin headwind is temporary & we like MediaTek on its TPU projects (report).Memory price should begin to normalize in late 2026, and so will margin pressure. We model TPU to be 7%/36% of MediaTek’s total EPS in 2026/2027 & recent checks suggestan upside risk. (more on next page) For memory, we sense stronger demand and price and raise valuation multiples and targetprices for Samsung/SK hynix/Micron to KRW140K/KRW750K/US$330. The recordprice upcycle is the biggest driver. The pace of the price hike may be ahead of our model& accelerate from 4Q25 to 1Q26. Overall price will stay strong nearly all 2026, but we stillbelieve the momentum will taper off in late 2026, as companies hikes capex by 30-40%and more bit supply comes online. And for 2027 we expect supply/demand to normalizeand price/margin to fall, but remain at or above healthy levels as AI demand continues toconsume capacity. HBM bit shipment is expected to double in 2026 with largely stableprice, and we expect Samsung to gain share. Chinese suppliers will likely grow bit share and worth monitoring but unlikely to alter the shortage situation in 2026.UMC/Vanguard/Novatek:Squeeze from memory & broad-based inflation is a headwind BERNSTEIN TICKER TABLE PRICE TARGET CHANGE / ESTIMATE CHANGE IN BOLD O - Outperform, M - Market-Perform, U - Underperform, NR - Not Rated, CS - Coverage SuspendedMU estimate is Adjusted EPS; 2330.TT, TSM, 2303.TT, UMC, 5347.TT valuation is P/B (x); MU valuation is Adjusted P/E (x); MU base year is 2025;Source: Bloomberg, Bernstein estimates and analysis. INVESTMENT IMPLICATIONS Samsung Electronics: We rate Samsung Electronics Outperform with target price raised from KRW 130,000 to KRW140,000,driven by higher valuation multiple from 1.5x 2-year forward P/B to 1.6x. SK hynix: We rate SK hynix Outperform with target price raised from KRW 650,000 to KRW 750,000, driven by highervaluation multiple from 1.7x 2-year forward P/B to 1.9x. Micron: We rate Micron Outperform with target price raised from US$270 to US$ 330.00, driven by higher valuation multiple from 2.4x 2-year forward P/B to 2.9x.KIOXIA:We rate KIOXIA Underperform with target price of JPY 7,000.00. MediaTek: We rate MediaTek Outperform with PT=NT$ 1,640.00. Novatek: We rate Novatek Market-Perform with PT lowered to NT$ 390.00 based on the same 14x target P/E and reviseddown Q5-Q8 EPS due to headwinds from cost inflation and end demand destruction. UMC: We rate UMC Underperform with PT=NT$ 32.00. DETAILS WE FIRST DISCUSS THEMATIC THESES, ON AI, TECHNOLOGIES, GEOPOLITICAL CONSIDERATIONS. After 3 years, unfortunately AI is still the main investment thesis for 2026. •Other theses hardly matter, as the intensity of AI investments, and the scope of possible implications is far bigger andbroader. Admittedly the worry on AI bubble or not will affect valuation, but on the other hand OpenAI and Anthropic very likely to release positive news in advance to drum up their possible IPOs (linkandlink) should support the sentiment. Wehave sufficient visibility that demand is so strong that supply struggles to keep up, and it is also possible sovereign projects should sustain demand beyond those from private sector. Earnings growth hence will be very stro