Fortinet三季度业绩三项超预期并上调全年营收及Billings指引，盘后下跌22.5%

{"visible_text": "Fig. 39: Global NAND revenue market share | Fig. 40: Global DRAM revenue market share\nFig.39 Time periods: 3Q24, 4Q24, 1Q25, 2Q25, 3Q25, 4Q25, 1Q26, 2Q26\nFig.39 Companies: Samsung, SK Hynix, Micron, Kioxia, SanDisk, YMTC\nFig.40 Time periods: 3Q24, 4Q24, 1Q25, 2Q25, 3Q25, 4Q25, 1Q26, 2Q26\nFig.40 Companies: Samsung, SK Hynix, Micron, CXMT, Nanya, Others\nSource: Counterpoint, Nomura research", "image_type": "stacked bar charts (2 figures)", "companies": ["Samsung", "SK Hynix", "Micron", "Kioxia", "SanDisk", "YMTC", "CXMT", "Nanya"], "industries": ["NAND Flash Memory", "DRAM Memory", "Semiconductors"], "key_numbers": {"NAND Samsung 3Q24": "37%", "NAND Samsung 2Q26": "28%", "NAND Samsung decline": "-9pp from 3Q24 to 2Q26", "NAND YMTC 3Q24": "4%", "NAND YMTC 2Q26": "14%", "NAND YMTC growth": "+10pp from 3Q24 to 2Q26", "DRAM Samsung 3Q24": "40%", "DRAM Samsung 2Q26": "38%", "DRAM SK Hynix peak 2Q25": "39%", "DRAM SK Hynix 2Q26": "25%", "DRAM CXMT 3Q24": "3%", "DRAM CXMT 2Q26": "10%"}, "table_data": [["Fig.39 NAND Market Share", "3Q24", "4Q24", "1Q25", "2Q25", "3Q25", "4Q25", "1Q26", "2Q26"], ["Samsung", "37%", "34%", "31%", "31%", "31%", "27%", "29%", "28%"], ["SK Hynix", "22%", "21%", "16%", "20%", "18%", "22%", "18%", "19%"], ["Micron", "12%", "14%", "15%", "13%", "12%", "12%", "13%", "15%"], ["Kioxia", "14%", "16%", "17%", "14%", "16%", "15%", "14%", "14%"], ["SanDisk", "11%", "11%", "13%", "11%", "12%", "12%", "13%", "11%"], ["YMTC", "4%", "4%", "8%", "11%", "10%", "12%", "14%", "14%"], ["Fig.40 DRAM Market Share", "3Q24", "4Q24", "1Q25", "2Q25", "3Q25", "4Q25", "1Q26", "2Q26"], ["Samsung", "40%", "38%", "34%", "33%", "33%", "36%", "38%", "38%"], ["SK Hynix", "33%", "35%", "36%", "39%", "33%", "32%", "29%", "25%"], ["Micron", "21%", "22%", "25%", "22%", "26%", "22%", "22%", "24%"], ["CXMT", "3%", "4%", "4%", "5%", "5%", "8%", "8%", "10%"], ["Nanya + Others", "~3%", "~1%", "~1%", "~1%", "~3%", "~2%", "~3%", "~3%"]], "chart_insights": ["Samsung loses NAND share from 37% to 28% over 2 years; competitor YMTC nearly quadruples from 4% to 14%", "DRAM market shows SK Hynix peaking at 39% in 2Q25 then declining sharply to 25% by 2Q26", "CXMT rises steadily in DRAM from 3% to 10%, indicating Chinese DRAM capacity expansion", "Micron relatively stable in both NAND (12-15%) and DRAM (21-26%)", "Samsung DRAM share recovers from 33% trough to 38% by 2Q26", "Chinese players (YMTC in NAND, CXMT in DRAM) gaining meaningful share"], "key_information": ["Global NAND revenue market share forecast through 2Q26 showing Chinese player YMTC gaining ground at expense of Samsung", "Global DRAM revenue market share forecast showing SK Hynix losing share to Samsung and CXMT", "Forward-looking forecast data (2Q25-2Q26 are projections)", "Counterpoint and Nomura research source suggests institutional/investment research report"], "source_date_clues": ["Data spans 3Q24 through 2Q26 (8 quarters)", "2Q26 data likely represents forecast/projection", "Source: Counterpoint + Nomura research indicates investment bank research report", "Nomura branding suggests equity research publication"], "research_relevance": "High relevance for memory semiconductor industry analysis. Useful for tracking competitive dynamics among Samsung, SK Hynix, Micron (Big 3) versus Chinese players (YMTC, CXMT). Important for NAND/DRAM market sizing, competitive positioning, and China self-sufficiency theme. Investment implications for Samsung Electronics, SK Hynix, Micron, and Chinese memory plays (YMTC indirectly via CXMT/parent companies)."}

Forward-looking forecast that Chinese memory players (YMTC in NAND, CXMT in DRAM) are projected to capture meaningful share through 2Q26；SK Hynix projected to lose DRAM share materially after a 2Q25 peak

# Weekly theme: Samsung's 3Q results, global forecast updates, de-spec impact Industry Overview ## Samsung's earnings and outlook confirm super-cycle Samsung Electronics (SEC) reported preliminary 3Q OP of W107tn, broadly in line with our estimate and consensus. However, this still confirms the memory super-cycle. Ahead of the 29 Oct earnings call, our analysis highlights three factors: DRAM and NAND contributed $>100\%$ of total OP, offsetting large losses in smartphones and foundry; memory ASP still rose by $20\%+$ QoQ after prices more than doubled in 1H; and OPM remained high at $70\%+/-$ (DRAM: $75\%$ ; NAND: $65\%$ ). We also expect OP to increase further in 4Q vs 3Q and in 2027 vs 2026. SEC's shareholder returns (dividends and buybacks) remain high, potentially exceeding W150tn pa using $50\%+$ of 2027-28, equivalent to $8-10\%$ of market cap. Memory capex will continue to rise (our assumptions: W67tn/W85tn/W92tn in 2026/27/28 vs W43tn in 2025), but we estimate annual FCF of W300tn+ in 2027-28. In our view, this also supports the super-cycle thesis. ## Fine-tune global forecasts for higher NAND ASP We updated our global memory forecasts to reflect stronger-than-expected NAND ASP. Micron's Aug-end quarter NAND ASP rose by \~30% QoQ (vs consensus of 10-20%). We also see upbeat 3Q NAND ASP trends at Korean and Japanese chipmakers, with increases of more than 20% QoQ. Against this backdrop, we raise our NAND ASP assumptions by 4% for 2H26 and 2027-28. For 3Q, we lifted our global forecast for the QoQ ASP change by 5ppt to +20%, from +15%. Although NAND spot prices have fallen since August (1Tb: down more than 10%), we continue to see strength in contract prices, led by high-end eSSD, boot-drive, etc. Overall, our latest industry model indicates that the global memory market will reach US\$2.1tn in 2030, vs US\$0.9tn+/- in 2026. We leave our DRAM assumptions unchanged, with ASP still rising through 4Q and the quarters of 2027. Taiwan tech companies' September sales were broadly strong, led by Nanya Tech and Phison Electronics. Based on August results, our BofA memory indicator remained well above the 2017/22 peak levels (179 vs 120-130; mid-cycle). ## De-spec could be growth catalyst rather than concern We acknowledge that some chipmakers may supply 8-hi HBM4 (8 x 3GB DRAM die = 24GB) for Rubin GPU next year, vs the current mainstream 12-hi configuration (36GB). As Rubin typically requires 8 units of HBM, a shift to 8-hi would reduce memory content from 288GB (8 x 36GB) to 192GB (8 x 24GB). However, we do not view this as memory de-spec or a demand cut, for three reasons: 1) total GPU shipments could rise quickly in response to the HBM shortage; 2) the incremental GPUs require additional memory – SOCAMM using LPDDR5 and key value cache eSSD – to be used with the counterpart logic processor CPU (e.g., Vera CPU); and 3) Big Tech should pay a price premium for 8-hi as a legacy memory vs 12-hi. Overall, the shift to 8-hi from 12-hi could support higher GPU volumes and, consequently, additional orders for SOCAMM and eSSD. We therefore see 8-hi-based de-spec as another growth catalyst for memory. That said, we have not observed any meaningful wafer input for 8-hi HBM4. As a result, 8-hi HBM4 volume will ## 09 October 2026 Equity Global Technology Simon Woo, CFA >> Research Analyst Merrill Lynch (Seoul) +82 2 3707 0554 simon.woo@bofa.com Dai Shen >> Research Analyst Merrill Lynch (Hong Kong) dai.shen@bofa.com Vivek Arya Research Analyst BofAS vivek.arya@bofa.com Mikio Hirakawa >> Research Analyst BofAS Japan mikio.hirakawa@bofa.com Matt Shin >> Research Analyst Merrill Lynch (Seoul) matt.shin2@bofa.com ## Exhibit 1: 2030 DRAM+NAND sales likely US\$2tn BofA global memory forecasts YoY 26E 27E 28E 30E DRAM $bn 573 917 1,025 1,396 NAND $bn 367 498 513 654 DRAM sales 328% 60% 12% 21% NAND sales 353% 36% 3% 17% DRAM ASP 242% 34% -5% 3% NAND ASP 276% 14% -13% 1% DRAM bit 25% 19% 18% 17% NAND bit 20% 19% 18% 16% DRAM capex 70% 37% 5% 3% NAND capex 57% 25% 10% 2% DRAM capa 12% 11% 14% 5% NAND capa 5% 7% 7% 5% Source: BofA Global Research estimates BofA GLOBAL RESEARCH Exhibit 2: DDR5 up WoW vs NAND decline Spot-market price trend US$ Current WoW QoQ YoY 16Gb DDR5 58.6 1% 23% 603% 16Gb DDR4 83.6 0% 8% 474% 8Gb DDR4 46.4 0% 24% 605% 1Tb NAND 30.2 -1% 25% 418% Source: DRAMeXchange BofA GLOBAL RESEARCH See glossary at the end of this report likely remain small in 1H27 or even 3Q, given the manufacturing cycle of $6+$ months. ## Global memory forecasts Exhibit 2: 2030 TAM can reach US\$2.1tn (DRAM+NAND); the annual run-rate closely approaching US\$1.0tn as of 3Q26 (US\$282bn; four-quarter equivalent: near-US\$1.0tn); 2026 ASP or contract prices rose 3-4x YoY, vs volume growth of 20%+; record-high ASP is sustainable, with limited price cuts through 2030 amid globally tight supply and strong AI chip demand; annual bit growth should remain in the high teens in 2026-30 Global memory forecast summary – top-down approach 1Q26 2Q26 3Q26E 4Q26E 1Q27E 2Q27E 3Q27E 4Q27E 2023 2024 2025 2026E 2027E 2028E 2029E 2030E Revenue US$bn DRAM 83.4 130.6 171.0 188.0 202.7 218.0 240.9 255.5 47.3 87.9 133.8 572.9 917.1 1,024.8 1,158.5 1,396.1 NAND 48.4 85.7 111.5 121.5 121.8 121.7 126.0 128.3 42.4 78.1 81.0 367.2 497.8 513.3 558.5 654.4 DRAM+NAND total 131.8 216.3 282.5 309.5 324.5 339.7 366.9 383.8 89.7 166.0 214.8 940.1 1,414.9 1,538.1 1,717.0 2,050.5 % QoQ DRAM 77% 56% 31% 10% 8% 8% 11% 6% NAND 82% 77% 30% 9% 0% 0% 4% 2% DRAM/NAND total 79% 64% 31% 10% 5% 5% 8% 5% % YoY DRAM 256% 363% 388% 299% 143% 67% 41% 36% -39% 86% 52% 328% 60% 12% 13% 21% NAND 216% 370% 432% 359% 152% 42% 13% 6% -37% 84% 4% 353% 36% 3% 9% 17% DRAM/NAND total 241% 366% 404% 320% 146% 57% 30% 24% -38% 85% 29% 338% 50% 9% 12% 19% Blended ASP US$/unit DRAM 8Gb equiv 8.1 11.8 14.6 15.7 16.5 17.0 17.3 17.4 1.8 2.9 3.7 12.7 17.1 16.2 15.4 15.8 NAND 256Gb equiv 5.4 8.9 10.7 11.1 11.1 10.7 10.2 9.9 1.6 2.6 2.4 9.2 10.4 9.1 8.5 8.6 % QoQ DRAM ASP 73% 46% 23% 8% 5% 3% 2% 1% NAND ASP 87% 67% 20% 3% 0% -3% -5% -3% % YoY DRAM ASP 155% 267% 307% 236% 104% 44% 19% 11% -45% 62% 29% 242% 34% -5% -5% 3% NAND ASP 130% 310% 364% 285% 106% 20% -5% -11% -46% 65% -8% 276% 14% -13% -7% 1% Shipments bn units DRAM 8Gb equiv 10.3 11.0 11.7 12.0 12.3 12.8 13.9 14.6 26.7 30.6 36.0 45.1 53.7 63.3 75.3 88.1 NAND 256Gb equiv 9.0 9.6 10.4 11.0 11.0 11.4 12.4 13.0 26.6 29.5 33.2 40.0 47.7 56.3 65.9 76.4 % QoQ DRAM bit growth 2% 7% 6% 2% 3% 4% 8% 5% NAND bit growth -2% 6% 8% 6% 0% 3% 9% 5% % YoY DRAM bit growth 40% 26% 20% 19% 19% 16% 18% 22% 10% 15% 18% 25% 19% 18% 19% 17% NAND bit growth 38% 14% 15% 19% 22% 18% 19% 18% 18% 11% 13% 20% 19% 18% 17% 16% Capacity 300mm k wpm DRAM wafer capacity 2,028 2,071 2,141 2,200 2,222 2,284 2,374 2,499 1,615 1,747 1,887 2,110 2,345 2,684 2,871 3,015 NAND wafer capacity 1,685 1,712 1,726 1,755 1,787 1,824 1,851 1,894 1,835 1,775 1,638 1,720 1,839 1,963 2,061 2,164 DRAM+NAND total 3,713 3,783 3,867 3,955 4,009 4,108 4,225 4,393 3,450 3,522 3,525 3,830 4,184 4,646 4,932 5,179 % QoQ DRAM capacity total 3% 2% 3% 3% 1% 3% 4% 5% NAND capacity total 2% 2% 1% 2% 2% 2% 1% 2% DRAM/NAND total 2% 2% 2% 2% 1% 2% 3% 4% % YoY DRAM capacity 11% 12% 13% 11% 10% 10% 11% 14% 4% 8% 8% 12% 11% 14% 7% 5% NAND capacity 2% 6% 6% 6% 6% 7% 7% 8% 4% -3% -8% 5% 7% 7% 5% 5% DRAM+NAND total 7% 9% 10% 9% 8% 9% 9% 11% 4% 2% 0% 9% 9% 11% 6% 5% Capex spending DRAM 25.3 24.3 23.3 24.3 30.6 32.0 34.6 36.0 28.8 43.6 57.3 97.2 133.2 139.9 166.5 171.5 NAND 8.0 8.3 8.7 9.7 11.2 10.8 10.4 10.8 20.5 22.7 22.1 34.7 43.2 47.3 52.0 53.1 Total (DRAM+NAND) 33.2 32.6 32.0 34.0 41.9 42.8 45.0 46.8 49.3 66.4 79.4 131.9 176.4 187.2 218.5 224.5 QoQ/YoY in capex DRAM 63% -4% -4% 4% 26% 4% 8% 4% -17% 51% 31% 70% 37% 5% 19% 3% NAND 29% 4% 4% 12% 16% -4% -4% 4% -30% 11% -3% 57% 25% 10% 10% 2% Total (DRAM+NAND) 53% -2% -2% 6% 23% 2% 5% 4% -23% 35% 20% 66% 34% 6% 17% 3% Source: Companies' reports, BofA Global Research estimates BofA GLOBAL RESEARCH Exhibit 3: Our higher 2026-28E NAND revenue estimates (up $3 - 4\%$ ) are mostly based on stronger 3Q26 ASP assumptions (up $20\%$ QoQ vs $15\%$ previously); nil revisions to our DRAM forecast as we already assumed solid ASP increase in 2H26 Global memory forecast revisions – top-down analysis 1Q26 2Q26 3Q26E 4Q26E 1Q27E 2Q27E 3Q27E 4Q27E 2021 2022 2023 2024 2025 2026E 2027E 2028E DRAM (8Gb equiv) Sales (US$bn) New 83.4 130.6 171.0 188.0 202.7 218.0 240.9 255.5 92.1 77.7 47.3 87.9 133.8 572.9 917.1 1,024.8 Old 83.4 130.6 171.0 188.0 202.7 218.0 240.9 255.5 92.1 77.7 47.3 87.9 133.8 572.9 917.1 1,024.8 Diff nm nm 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% nm nm nm nm nm 0.0% 0.0% 0.0% Shipments (bn units) New 10.3 11.0 11.7 12.0 12.3 12.8 13.9 14.6 24.1 24.2 26.7 30.6 36.0 45.1 53.7 63.3 Old 10.3 11.0 11.7 12.0 12.3 12.8 13.9 14.6 24.1 24.2 26.7 30.6 36.0 45.1 53.7 63.3 Diff nm nm 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% nm nm nm nm nm 0.0% 0.0% 0.0% Bit growth - sequential New 2.2% 7.0% 6.4% 2.1% 2.7% 4.2% 8.3% 5.3% 23.3% 0.3% 10.3% 14.8% 17.6% 25.2% 19.1% 17.8% Old 2.2% 7.0% 6.4% 2.1% 2.7% 4.2% 8.3% 5.3% 23.3% 0.3% 10.3% 14.8% 17.6% 25.2% 19.1% 17.8% Diff nm nm 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% nm nm nm nm nm 0.0% 0.0% 0.0% ASP chg - sequential New 73.2% 46.3% 23.1% 7.7% 5.0% 3.2% 2.0% 0.7% 14.2% -16.0% -44.8% 61.8% 29.4% 241.9% 34.5% -5.2% Old 73.2% 46.3% 23.1% 7.7% 5.0% 3.2% 2.0% 0.7% 14.2% -16.0% -44.8% 61.8% 29.4% 241.9% 34.5% -5.2% Diff nm nm 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% nm nm nm nm nm 0.0% 0.0% 0.0% Blended ASP (US$) New 8.1 11.8 14.6 15.7 16.5 17.0 17.3 17.4 3.8 3.2 1.8 2.9 3.7 12.7 17.1 16.2 Old 8.1 11.8 14.6 15.7 16.5 17.0 17.3 17.4 3.8 3.2 1.8 2.9 3.7 12.7 17.1 16.2 Diff nm nm 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% nm nm nm nm nm 0.0% 0.0% 0.0% NAND (256Gb equiv) Sales (US$bn) New 48.4 85.7 111.5 121.5 121.8 121.7 126.0 128.3 73.3 67.3 42.4 78.1 81.0 367.2 497.8 513.3 Old 48.4 85.7 106.9 116.5 116.7 116.6 120.7 123.0 73.3 67.3 42.4 78.1 81.0 357.5 477.0 491.9 Diff nm nm 4.3% 4.3% 4.3% 4.3% 4.3% 4.3% nm nm nm nm nm 2.7% 4.3% 4.3% Shipments (bn units) New 9.0 9.6 10.4 11.0 11.0 11.4 12.4 13.0 21.3 22.6 26.6 29.5 33.2 40.0 47.7 56.3 Old 9.0 9.6 10.4 11.0 11.0 11.4 12.4 13.0 21.3 22.6 26.6 29.5 33.2 40.0 47.7 56.3 Diff nm nm 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% nm nm nm nm nm 0.0% 0.0% 0.0% Bit growth - sequential New -2.2% 6.3% 8.4% 5.8% 0.1% 3.1% 9.0% 5.0% 37.8% 6.1% 17.6% 11.2% 12.5% 20.4% 19.3% 18.0% Old -2.2% 6.3% 8.4% 5.8% 0.1% 3.1% 9.0% 5.0% 37.8% 6.1% 17.6% 11.2% 12.5% 20.4% 19.3% 18.0% Diff nm nm 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% nm nm nm nm nm 0.0% 0.0% 0.0% ASP chg - sequential New 86.6% 66.8% 20.0% 3.0% 0.1% -3.1% -5.0% -3.0% -9.1% -13.5% -46.4% 65.5% -7.8% 276.5% 13.6% -12.6% Old 86.6% 66.8% 15.0% 3.0% 0.1% -3.1% -5.0% -3.0% -9.1% -13.5% -46.4% 65.5% -7.8% 266.5% 11.9% -12.6% Diff nm nm 5.0% 0.0% 0.0% 0.0% 0.0% 0.0% nm nm nm nm nm 10.0% 1.8% 0.0% Blended ASP (US$) New 5.4 8.9 10.7 11.1 11.1 10.7 10.2 9.9 3.4 3.0 1.6 2.6 2.4 9.2 10.4 9.1 Old 5.4 8.9 10.3 10.6 10.6 10.3 9.8 9.5 3.4 3.0 1.6 2.6 2.4 8.9 10.0 8.7 Diff nm nm 4.3% 4.3% 4.3% 4.3% 4.3% 4.3% nm nm nm nm nm 2.7% 4.3% 4.3% Source: BofA Global Research estimates, industry data. Chg = change. Equiv = equivalent BofA GLOBAL RESEARCH ## Rubin Ultra and Vera Rubin NVL72 despec tables Exhibit 4: NVIDIA's original Rubin Ultra roadmap contemplated 16 HBM4 stacks with roughly 1TB of memory capacity. However, ongoing HBM supply constraints have shifted industry expectations toward lower-memory configurations, including 8-stack HBM4E 12-Hi (\~288GB) or 8-stack HBM4E 8-Hi (\~192GB) designs. We believe the lower-memory configurations should improve affordability and ultimately enable higher shipment volumes. Rubin Ultra's new spec shows 192GB/288GB vs original roadmap of 1TB Stage HBM type # of HBM units Stack height DRAM die density HBM capacity Remarks Source Original roadmap (2025) HBM4E 16 16-Hi ~32Gb ~1TB Initial Rubin Ultra concept targeting maximum memory capacity and performance. NVIDIA GTC 2025 keynote Baseline design (1H26) HBM4E 8 12-Hi ~24Gb ~288GB NVIDIA's primary Rubin Ultra design through 1H26. TrendForce (Aug 2026) Current leading alternative HBM4E 8 8-Hi ~24Gb ~192GB Lower-memory option being evaluated due to HBM supply constraints and HBM4E qualification risks. SemiAnalysis; corroborated by TrendForce Alternative under evaluation HBM4 8 12-Hi ~24Gb ~288GB Maintains capacity by using HBM4 rather than HBM4E. TrendForce (Aug 2026) Alternative under evaluation HBM4 8 8-Hi ~24Gb ~192GB Lowest-memory option currently under consideration. TrendForce (Aug 2026) Current Vera Rubin HBM4 8 12-Hi ~24Gb 288GB Current-generation reference platform. NVIDIA GTC 2025 roadmap Source: Company, media reports, BofA Global Research BofA GLOBAL RESEARCH Exhibit 5: According to recent media reports, Vera Rubin NVL72's HBM content remains unchanged at approximately 20.7TB per rack. The main revision relates to CPU-attached DRAM (SOCAMM), with expectations reduced to \~27TB per rack from an initial \~55TB. Despite the lower configuration, total DRAM capacity would still be roughly 50% higher than Blackwell GB300 NVL72 (\~18TB). Rubin NVL72 rack server spec comparison original spec vs street assumption Parameter GB300 NVL72 Rubin NVL72 (Old NVIDIA spec) Rubin NVL72 (Street speculation) System type Giant superchip Giant superchip Giant superchip Use case Training & inference Training & inference Training & inference Launch 2H25 (exp) 2H26 (exp) 2H26 (exp) GPU count 72 Blackwell GPUs 72 Rubin GPUs 72 Rubin GPUs CPU count 36 Grace CPUs 36 Vera CPUs 36 Vera CPUs HBM (per GPU) 288GB HBM3e 288GB HBM4 288GB HBM4 Total HBM per rack ~20TB ~20-20.7TB ~20-20.7TB (unchanged) DRAM per CPU ~500GB LPDDR5X (implied) 1.5TB LPDDR5X ~750GB LPDDR5X Total DRAM per rack ~18TB ~54-55TB ~27-28TB DRAM vs GB300 Base ~3x higher (initial plan) ~50% higher (revised) Source: Company, media reports, BofA Global Research BofA GLOBAL RESEARCH ## Taiwan sales and Korea exports Exhibit 7: Exceptionally strong sales growth continued in 3Q26 (NT\$134bn; up 62% QoQ and 612% YoY) Nanya Tech – Monthly sales trend ![](images/0a2680a84874ae91b5571f0e86eae9afc6eebc79a4334424eecd439b89465ad1.jpg) Source: Company BofA GLOBAL RESEARCH Exhibit 9: Broadly strong MoM recovery with notable growth from Pegatron (+143%), Hon Hai (+26%), Phison (+8%), Wiwynn (+20%) Taiwan tech companies' MoM monthly sales (Sep-2026) ![](images/e37c9c80df880a1c398021da699690d92f78663380f73fb4255a1f8f29affd4e.jpg) \*Pegatron up 143%; axis scaled to show more relative comparison Source: Companies BofA GLOBAL RESEARCH Exhibit 10: Significant MoM growth in Sept (+29%), resulting in a new record-high in Sep (US\$60bn), which is around 4x higher than month-average in 2025; mostly driven by HBM4, DDR5 and rush order Korea semis exports - US\$bn/month ![](images/20a7ddbe7f171b338b1741e7968f15ff05a784cd77aff9ccedff36689f4a688c.jpg) Exhibit 8: Also robust growth see in 3Q26 (NT\$86bn; up 27% QoQ and 375% YoY) BofA GLOBAL RESEARCH Phison – Monthly sales trend Source: MoTIR ![](images/5d587e1b2087b5a20f22db4e26bcb9b569d7fb04941f0cb2152c859a8c927e71.jpg) Source: Company BofA GLOBAL RESEARCH Source: Companies Exhibit 10: Robust YoY rebound across the board – TSMC up 55%, Nanya Tech up 577%, Phison up 370%, Unimicron up 65%, Quanta up 77% ![](images/ba01a7250eca52716c167b8f8350cd8958640ce4206b8739d6d86ae05918a966.jpg) Taiwan tech companies' YoY monthly sales (Sep-2026) BofA GLOBAL RESEARCH Exhibit 12: Robust YoY growth in Sept (+263%) and strong trend should likely continue throughout 4Q26 Korea semis exports - YoY change in monthly US\$bn ![](images/42eaf7adc4639eed9f37ad9941793e6876b4cd73979704db677c58f1418af494.jpg) Source: MoTIR BofA GLOBAL RESEARCH ## BofA Memory Indicator Exhibit 12: The BofA Memory Indicator still presents super-cycle – much higher than 2017-18 peak level; of course, it slightly declined to 179 in Aug-26 (due to lower YoY growth) after hitting peak of 190 in Mar/Apr-26, driven by exceptionally strong DRAM/NAND spot pricing, rising ASPs and billings, along with Korea's \~200%+ export growth. BofA Memory Indicator – back to upturn level since Oct-2025 and hit highest peak in Mar/Apr-26 (back-tested) ![](images/40fa3340aa0e98a7d03f1f89b0bb6186fa18c4ac8dc29c20733786d5519d97ff.jpg) Note: Our indicator had previously been capped at \~140, but the recent unprecedented surge—driven by memory spot prices, ASPs, and billings—has led us to raise the ceiling (up to 240 levels) to reflect a more accurate relative comparison. This exceptional strength is further validated by strong earnings from memory chipmakers. Source: DRAMeXchange, WSTS, MoTIR Korea, BofA Global Research \*The shaded area represents back-tested results from January 1991 to March 2021. The unshaded area represents actual performance since April 2021. This performance is back-tested up to March-2021, and does not represent the actual performance of any account or fund. Back-tested performance depicts the theoretical (not actual) performance of a particular strategy over the time period indicated. No representation is being made that any actual portfolio is likely to have achieved returns similar to those shown herein. Disclaimer: The BofA Memory Indicator is intended to be an indicative metric only and may not be used for reference purposes or as a measure of performance for any financial instrument or contract, or otherwise relied upon by third parties for any other purpose, without the prior written consent of BofA Global Research. This indicator was not created to act as a benchmark. For methodology, see the 5 Aug 2024 Global Memory Tech report. BofA GLOBAL RESEARCH Exhibit 13: Sharp correction in memory companies' stock prices in Jul-Aug due to memory peak-out concerns, global uncertainty and AI concerns, but steady recovery seen in Sep-early Oct BofA Memory Indicator is highly correlated with stock performance (back-tested) ![](images/ab933b59fa0598aa34d5d007f386c146ecfe167ed6b2d5fe221fc5eba4cc8c10.jpg) Note: Memory companies share price is average of Samsung, Hynix, Micron, and Nanya share price YoY changes Source: Bloomberg, BofA Global Research The light green shaded area represents back-tested results from January 1991 to March 2021. The blue shaded area represents actual data since April 2021. This performance is back-tested and does not represent the actual performance of any account or fund. Back-tested performance depicts the theoretical (not actual) performance of a particular strategy over the time period indicated. No representation is being made that any actual portfolio is likely to have achieved returns similar to those shown herein. For methodology, see the 5 Aug 2024 Global Memory Tech report. ## Key 6 memory price charts Exhibit 15: DDR5 spot prices increased slightly this week to a fresh all-time high of \$58, with pricing now up nearly 100% YTD, further extending the robust rally that started in 4Q25 16Gb DDR5 spot price – daily, Jan '26 - Oct '26 ![](images/4051bdb2b49c5f29acc62c597502b9f9f73337eb0a4377f06df6abe9c960234d.jpg) Source: DRAMeXchange BofA GLOBAL RESEARCH Exhibit 15: 8Gb DDR4 spot prices remained stable at around \$45, with no signs of a downturn yet Exhibit 16: NAND spot prices softened through Sep and early-Oct after recovering from July through August. ![](images/521956623b6009cbeffa712f973035c37ef90a3010cf3111341a68afc0b494bb.jpg) 1Tb NAND wafer spot price – weekly, Jan '26 - Oct '26 Source: DRAMeXchange BofA GLOBAL RESEARCH 8Gb DDR4 spot price trend, Jan '26 - Oct '26 Source: DRAMeXchange BofA GLOBAL RESEARCH ![](images/b5e2682c0b3fa885c7a8200570034ab7de607124e80ca89589cfea9f873d379b.jpg) Exhibit 17: In contrast, low-density NAND wafer contract price rebounded sharply through Aug-Sep and early-Oct as well 256Gb NAND wafer spot price – weekly, Jan '26 - Oct '26 ![](images/9c82e8fbaa926a487a5325142c978b9e5c56f2344ecab3cc6003d6503ecd0ec5.jpg) ![](images/8bb35f6c792881858372243be7ec02c5bd02f609392bc5a19e307adb034dd507.jpg) Source: DRAMeXchange BofA GL…

Upward revision of NAND ASP assumptions (4% for 2H26 and 2027-28) based on stronger-than-expected Micron and Korean/Japanese chipmaker contract pricing.；3Q global NAND QoQ ASP change forecast raised to +20% from +15% (+5ppt).

October 11, 2026 09:42 PM GMT Tech Bytes | Asia Pacific # Memory – Cycles vs. Super Cycles Memory investing is less about pinpointing inflections or identifying data points – but it's about the market rewarding those who priced it first. AI is a compute problem, a memory problem, hence structurally different. In the later stages, however, still cyclical and reward is harder to come. Fundamentals look bulletproof but a more cautious tale for stocks. The AI memory super cycle is here, yet the market is consciously refusing to pay a full multiple for unprecedented earnings. Just like past cycles, it suspects that may prove temporary – memory stocks have historically looked cheapest right at the top of the cycle, because analysts extrapolate boom prices into the future. The case for a Supercycle rests on the idea that this time is structurally different, which looks obvious by now with trillion dollar market cap. But separating the asset from the price of the asset (share price) is paramount following a 9x run in share prices. ## Structurally higher AI secular growth riding on top of a cycle... is still a cycle. What's different is that this one is steeper, much steeper (nearly 7x higher \$/Gb today vs. past peaks). Markets can certainly be wrong, but market prices themselves are part of the evidence set for how a memory cycle is perceived. The 'lazy' answer is: (1) nothing being different this time; or (2) the magnitude of the next downturn may be much smaller and the floor higher than in 2023. Both are conflicting views, but both can be right, as we have come a long way already – the memory supercycle is real but it is also not the end of cyclicality. Hence, a structurally better memory cycle and attractive memory stocks prices do not have to go hand-in-hand. In other words, we think the easy money is past. ## Navigating memory cycles – they often repeat previous cycles...but they rarely run like a train schedule: What ends this memory supercycle may not be AI demand softening but supply catching-up and/or a demand digestion. That may still be distant as memory prices stay firm through 2027, despite deceleration. Trying to figure out the exact start and end dates of cycle inflections is unrealistic. But our view is that one should build portfolios that are durable enough against the inevitability of cycle inflections with the understanding that they are a natural extension of markets, even if we don't know when they will occur. In the current market, we are reminded that history does not always repeat itself (this time it's AI inference, last time it was cloud) – but it often rhymes. By studying what drives the previous cycle inflections, one can apply the remedies used to devise new ones from the similarity of the events. From this, we believe that: \- It is better to be generous than conservative on valuations on the way up but MORGAN STANLEY & CO. INTERNATIONAL PLC+ Shawn Kim Equity Analyst Shawn.Kim@morganstanley.com +44 20 7677-1018 Cindy Huang Equity Analyst Cindy.Huang@morganstanley.com +44 20 7425-2915 MORGAN STANLEY & CO. INTERNATIONAL PLC, SEOUL BRANCH+ Ryan Kim Equity Analyst Ryan.G.Kim@morganstanley.com +82 2 399-4939 MORGAN STANLEY ASIA LIMITED+ Kathy Zhu Research Associate Kathy.Zhu@morganstanley.com +852 3963-2618 ## S. KOREA TECHNOLOGY Asia Pacific Industry View Attractive Morgan Stanley does and seeks to do business with companies covered in Morgan Stanley Research. As a result, investors should be aware that the firm may have a conflict of interest that could affect the objectivity of Morgan Stanley Research. Investors should consider Morgan Stanley Research as only a single factor in making their investment decision. For analyst certification and other important disclosures, refer to the Disclosure Section, located at the end of this report. += Analysts employed by non-U.S. affiliates are not registered with FINRA, may not be associated persons of the member and may not be subject to FINRA restrictions on communications with a subject company, public appearances and trading securities held by a research analyst account. the opposite is true on the way down. \- It is better to be optimistic rather than cautious on earnings – price hikes tend to be greatly underestimated, especially in super cycles. The opposite is also true. \- It is better to be contrarian when looking for exit points well ahead of cycle inflections. ## Super Cycles – Not The End of Cyclicality Memory fundamentals look bulletproof but a more cautious tale for stocks. Samsung reported the highest earnings among all companies worldwide with US\$80bn in operating profits in CY3Q26 which is the largest ever by a technology company, displacing NVDA US \$64bn (2QFY27). For a company that sells a commodity and the single priciest component inside an AI server (not the logic chip), it is remarkable. Yet, both Samsung and Hynix shares lag 35% below peak, trade below historical peak P/E of 4x on peak earnings (currently at 3-4x NTM P/E) implying that Mr. Market seems less convinced that this time is different. Despite significant LTA commitments and market expectations of more muted cyclicality, memory makers are being modeled as if the old boom-bust cycle still governs. In this demand-led super cycle, we think super profits will eventually bring on new supply, with inevitable consequences for returns and ultimately for share prices. Exhibit 1: Samsung world No.1 profit... ![](images/b7f35e8bd0b7f477ccbe8a3af4d74d452a4eeb91dbe5d0cc0c5ecc7f33c9c55f.jpg) Source: FactSet consensus (mean), Morgan Stanley Research. Exhibit 2: ... but trading well below ![](images/e2931fffe51c66e9163a99c716e11ed906199d53195d9d395d201893d55d238e.jpg) Source: FactSet consensus (mean), Morgan Stanley Research. ## The good, the bad and the realistic This time is truly different... The structural story remains compelling for memory – AI changes everything. The demand outlook is structural as AI data centres spending is the main driver of revenues, on-device AI is the next wave, physical AI, humanoid robots, self-driving cars, and even the AI data centres opportunity in space. Industry capacity is sold out years ahead, the majority under long-term contracts with price floors that never existed before and security cash deposits for some LTAs. The HBM market is forecast to triple to US\$100 bn by 2028. This is what separates a Supercycle from an ordinary one. \- Demand from AI data center construction is outstripping what the industry can supply. • Structurally higher margins. \- LTA changes the risk of capacity allocation – with locked-in customer commitments. • HBM physically crowds out commodity DRAM. \- Every new model generation needs more memory per chip, and there are more chips being made. … or is this the top… But new capacity will come eventually, and quarterly memory WFE shipment showed how numbers keep climbing higher on equipment pull-ins and shipment all year. Compute gets cheaper, competition itself is the mechanism that drives the cost curve down and mean-reversion has never missed a beat on past memory cycles. Producers are all racing to deliver more intelligence per dollar per watt, and the philosophical disagreement about how to get there does not change where they are all going. 1. China is a major swing factor (Tech Bytes: China Memory – Reshaping Global Competition (26 Aug 2026)), where state-backed CXMT is expanding fast in DRAM and HBM3 while YMTC already a legitimate global competitor in NAND. \- YMTC is already at scale and globally competitive. It holds roughly 12–13% of the global NAND market – most of it sold domestically – on a 294-layer X-stacking process, running around 200k wafers per month with \~45% domestic-tool adoption. \- CXMT is at scale and shipping DDR5 and HBM3. It's main HBM3 customers are from China CSPs and AI chip design houses and we estimate about 20% HBM market share in China. 2. AI data center rack content could be another factor. GPUs currently account for roughly half of total Vera Rubin cluster cost, memory about 29% with the remaining 23% is networking, power, cooling, and facilities. As data centers shift to multi-racks (scale out) and next generation iterations, networking content value increases significantly at the expense of memory content while GPUs stay relatively unchanged. Exhibit 3: AI data center rack content ![](images/5e45eaf75fdb60743ff44b4a316fcfcbe25dc2a91bc16c46ca5e47163e5d2066.jpg) Source: Morgan Stanley Research ... But both can be right. We have come a long way already. The easy money is past, stocks are up 9 fold since the start of the cycle (September 2025), all within USD\$1tr market cap, and not pricing a normal outlook. The memory super-cycle is real but it is also not the end of cyclicality. The cycle amplitude is steeper and duration may be much longer this time, but the path to a trillion market cap is already here and has priced-in a great deal of that run. Hence, a structurally better memory cycle and attractive memory stocks prices do not have to go hand-in-hand. ## What comes next? Historically, memory companies have been poor at predicting their own inflection points. In this context, management guidance tends to confirm trends that the market already knows. What ends a memory cycle is never demand softening but supply catching up and/or demand digestion – reaching a point where there is enough HBM supply, ABF, CoWoS packaging capacity, enough fab output to flip deficit to surplus, resulting in hyperscaler buying pauses to absorb what it already bought. This happened in 2000, 2016, 2019, and 2022. This is unlikely to happen in 2026 and most of 2027 given lead times for new capacity additions. It is reasonable to say that every smart person in the AI industry is looking for ways to alleviate the 'memory wall' from every angle given that memory reached nearly half the cost of the most important chip in the world. Putting the current memory group stock consolidation in the context of previous memory super cycles, the cycle upturn may be reaching a relatively mature stage relative to market expectations that may be more advanced due to the positives in AI capex spend to meet insatiable demand trends. While it appears reasonable to assume that a lot of good news is discounted, with the \~180% average performance for memory stocks YTD, a sustained cyclical upturn would require the memory group's contract pricing to rise another c. 70% QoQ from current levels in DRAM (vs. c. 700% YoY peak in 3Q26 and 160% YoY in 2017's past super cycle peak). This seems out of reach and the market is now left with incoming massive capital returns. Based on the average earnings upgrade for Asian memory stocks of 432% this year, it should be clear that consensus is "erring on the side of an unprecedented cycle turn" consistent with our base case assumption and market expectations, which is a sharp upturn and durable/elongated cycle with visibility into 2027 (Samsung) and beyond (Micron, SK hynix). Exhibit 4: Memory EPS revisions breadth (3M avg) – DRAM turning down first ![](images/a5410f270125194d7bede52b55063ea58edb7a3f26904322275d06861f30d188.jpg) Source: Trendforce, Morgan Stanley Research Exhibit 5: DRAM YoY vs. PB ![](images/5cce71c09e718d8e67876ba9c27aa5f5e3142b9ed1e22f00ad24f5df14fc3d04.jpg) Source: Trendforce, Morgan Stanley Research. Can memory shares underperform before upgrades have peaked? In short, we think yes: In Exhibit 2, we illustrate some of the historical data with an overlay of share price and consensus earnings trajectory. Over the long term, there is a close correlation between share price and earnings, although there is an offset, with shares having historically fallen before earnings peaked – i.e., the shares no longer outperform even if the upgrade cycle has not ended. \- Memory shares outperformed four months before earnings peaked in the 2008 cycle, four months in 2012, and eight months in 2018. \- With decelerating revenue growth and margin improvement moderating, we think earnings estimate revisions have limited upside – sell-side estimates have become a lagging indicator at this point. \- Staying with stocks that have a margin of safety on valuation seems prudent before growth peaks. ## Anatomy of previous super cycles \- Low valuations at cycle peak. In 4Q17, the MSCI Asia Semi sector 12-month forward P/E was 13.3x at peak – this fell to c.10.0x by 4Q18. • Stocks looking “cheap” does not stop forced selling. \- The sector declined much sooner than earnings estimates – at the time EPS peaked, it traded at c. 10x on average. \- Current P/E valuations are 34% (semis) and 39% (hardware) above the 2018/19 peak, but are 14% (semis) and 25% (hardware) below their most recent peak in the 2023/24 cycle. \- The earnings upgrade cycle is uneven and stays robust for the AI computing segment while decelerating for memory – this could take several months. In general terms, companies are showing robust financial health – and are structurally stronger than in 2018/19. Although we can say that these are materially better companies now (in terms of balance sheet and return on invested capital), there are some reasons why returns are already higher and could remain so: \- Demand trends have been unusually strong in some important new emerging segments, especially AI-led (HBM, SOCAMM, CXL, SRAM). \- We don't know where AI architectures are heading. Transformers dominate today, but the field moves fast and the adoption of AI agents only began with 'Open Claw' in January 30, 2026. \- Managements generally appear focused on profitability (lessons learned from the 2022 crash) but have been more focused on bringing supply to keep up with insatiable compute demand. \- However, there is very little inventory build from over-ordering to meet a pull-forward of demand. The fact remains that sector ROE and margins are well above long-term averages, while our leading earnings model is pointing to significant increases. The inconsistency of prosperity. Market behavior can be difficult to predict – i.e., regardless of the ultimate outcome, markets may move to price in a peak followed by a new downturn without waiting for any confirmation. This creates opportunities to benefit from such divergence, and is the basis of our stock recommendations. \- Valuations are not a good predictor of future returns. Risk-reward is likely to improve marginally from here with industry and consensus upgrades likely only just moderating. \- Management of memory companies is generally bullish and rarely show less optimism. Relying on companies to call a downturn is ineffective. \- Semiconductors and memory is always cyclical. \- It is always good to take a step back and reassess when stocks stock repeatedly go down on good news. ## Stock implications Favored: We continue to favor AI computing and networking supply chain over memory on a relative basis. Our preference is mainly earnings revisions, content growth (new specs) or product cycle/unit growth: • High-quality, idiosyncratic factors \- Early-cycle commodity plays (i.e., accelerating YoY) \- Content growth – HV/DC ; and \- Areas offering higher pricing leverage on earnings – CCL, ABF, power semis. See Exhibit 6 – stock preference table. Least favored: Downstream hardware stocks where risk of overbuild and margin pressure is greatest owing to higher input costs (memory pricing) and valuations are perceived to be highest. Exhibit 6: Three AI buckets – compute first, networking second, memory selective ![](images/28f78d981607e97c04a0270dac9f296cd24d5833af2c3b268041f5896bd9b52b.jpg) Source: Morgan Stanley Research. ## Memory channel checks ## HBM \- HBM pricing moves sharply higher in 2027 as expected. Industry forecasts put blended HBM ASP to grow \~121% YoY. HBM4 is tracking at US\$3.0–5.0/Gb and HBM4e at US\$5.0/Gb or above. \- Rubin HBM is ramping slower than planned. VR200 NVL72 rack shipments in 2026 may come in below 10K, vs. an original target of about 15K, with meaningful volume in 4Q26-1Q27. Blackwell is expected to account for \~78% of NVIDIA's high-end GPU shipments in 2026, with Rubin becoming mainstream after 1H27. \- Yields remain the binding constraint. HBM4 production continues to face technical and yield challenges across suppliers, which limits AI chip shipments. 8-high yields run about 5 percentage points above 12-high with a shorter cycle time, which is part of the appeal. \- 8-high to become the 2027 mainstream configuration, driven by supply and cost rather than weaker demand. 8-high carries a 10-20% per-Gb premium over 12-high, as each stack still needs one base die. Lower HBM content per chip should support RDIMM and SOCAMM demand, as part of the inference KV cache moves to host memory. Exhibit 7: HBM price forecast by generation (US\$/Gb) 2025 2026F 2027F HBM3e 1.80 1.3–2.0 2.3–2.5 HBM4 2.05 2.05+ 3.0–5.0 HBM4e - - 5.0+ HBM blended ASP 1.77 1.74 3.84 Blended ASP YoY 18% (1%) 121% Source: TrendForce ## DRAM \- Pricing momentum improves into 4Q26. Industry data points to conventional DRAM contract prices up 10-15% QoQ, revised up from 3-8% previously. PC DRAM is tracking up 13-18% and server DRAM up 10-15%, while mobile DRAM slows to 0-5%. Consumer DRAM (DDR3/DDR4) remains the most undersupplied segment. \- Server demand remains the key driver. US hyperscalers released additional RDIMM orders after reviewing their 2027 build-out plans. Suppliers continue to prioritize server allocation, constrained by both front-end and back-end capacity. \- Consumer demand stays weak. Smartphone brands remain cautious given high memory costs. PC OEMs continue to pull in orders, aiming to build inventory above 10 weeks on concern about 2027 shortages. \- PC DRAM supply set to shrink in 2027. Korean suppliers' PC DRAM supply could fall more than 20% YoY in 2027 as capacity moves to server and HBM. Suppliers expect to reallocate any capacity PC OEMs do not take to server DRAM. \- Legacy DDR4 remains the tightest. DDR5 modules now sell at a 10% discount to DDR4, versus 6% in 2Q26. DDR4 supply is expected to fall about 20% in 2027 as major suppliers exit legacy nodes, while secondary supply shrinks as data centres extend server life and buy used RDIMMs. \- Capex accelerating, but new output arrives late. Industry data puts DRAM capex at about US\$98bn in 2026 (+82% YoY) and US\$150–160bn in 2027 (+\~60%), mainly for new fabs. Meaningful bit output only starts in 2H27, so supply should stay short through end-2027. ## NAND \- Pricing also revised up 15-20% QoQ in 4Q26, from 0-5%. eSSD is the only category still accelerating, at 23-28% QoQ (from 20–25% in 3Q26). Client SSD and eMMC/UFS slow to 5-10%, and wafers to 3-8%. \- QLC eSSD adoption is broadening. Several large US cloud have shifted from TLC SSD plus HDD to QLC SSD to store AI inference KV cache. QLC costs about one third of TLC for the same capacity, and agentic workloads are write-once, read-many, which limits the impact of QLC's shorter lifespan. \- Divergence set to widen in 2027. eSSD demand keeps growing, while consumer NAND may face downward price pressure as OEMs downgrade specifications and Chinese supply increases. \- Consumer inventory is building and moving downstream. Consumer-grade wafer inventory is rising at suppliers, who are clearing it through bundled sales. Module houses are taking on consumer stock in exchange for allocations of scarce components. Exhibit 8: DRAM earnings revisions breadth coming off recently... ![](images/35a370f556b0495277ad656cedf4ee3a044d91fe7df984b65f7ba8df58d64c18.jpg) Source: Morgan Stanley Research. Exhibit 9: ...driven by all three major players ![](images/7e0203342946cb8b0dfba9bb261d335f96e08ab3f0c7e6832eb766081231a7ab.jpg) Source: Morgan Stanley Research. Exhibit 10: Samsung: EPS Revisions Breadth vs Share Price Performance Y/Y ![](images/42531d26c02a25f6cb0a19df0c49f3b9d9e9f108766a5f5d02ae53edfc237b82.jpg) Source: Morgan Stanley Research. Exhibit 11: SK hynix: EPS Revisio…

Samsung's US$80bn operating profits in 3Q26 set a record as the largest ever by a technology company, surpassing NVIDIA.；Morgan Stanley frames the memory supercycle as 'real but not the end of cyclicality' – a structurally better memory cycle does not necessarily mean attractive stock prices.

# Memory Semis Monthly October '26 Edition: Is the Wall of Worry justified? ## CXMT's possible 4F2 ramp not a concern for industry supply According to the press (link), CXMT's President indicated "launching" 4F2 Vertical Channel Access Transistor (VCAT) DRAM before year end; it is unclear what "launching" means, but we suspect R&D or sample as 6F2 G5 (1a nm) manufacturing ramps in 2027. As we wrote before (link; link), 4F2 VCAT allows to increase density (all things being equal) by $50\%$ vs. 6F2 whilst not changing the lithography half pitch; or less if relaxing it. Ultimately though, subsequent process nodes will need lower half pitch and more advanced litho. VCAT also has its own sets of challenges around deposition (ALD) and selective epi in particular. We note that Samsung is likely targeting 4F2 VCAT for 0a nm by late-2028, after 1d nm 6F2 ramp in 2027. Lastly, moving on a distinct path for DRAM process tech usually requires at least 3 process nodes iterations to maximize efficiency and performance. This may mean a later than otherwise move to 3D DRAM, which would, ultimately, remove the need for EUV litho for DRAM manufacturing. This, to us, would be the more material fork in the road. ## HBM de-specing may continue amidst stronger server DDR/SSD demand Our checks indicate that DRAM makers continue to see upsized orders into 2027 for server DDR5, boosted by agentic AI inferencing deployment. In addition, we believe that SOCAMM orders from Nvidia (for Vera CPUs) have gotten stronger over the past 2 months, now exceeding 50bn Gb in 2027. This is larger than total PC+GDDR DRAM bit consumption. This, in turn, has pressured down HBM shipments, which we estimate would likely only grow 49% YoY. Hence, HBM de-specing was largely unavoidable. We continue to expect Nvidia VR300 to adopt HBM4 8-Hi (then HBM4E). Feynman may adopt HBM4E 8 or 12-Hi, as well as TPU v9. Some inference-centric ASICs may be at 4-Hi. One consequence of HBM density de-specing for VR300 is an increase in demand for NAND (KV Cache; CMX for Nvidia), which our checks indicate has recently picked up. ## Samsung may accelerate the cadence of capacity ramp further We recently increased, again, our capex forecasts for Samsung (link). We expect Samsung to add 120k wpm of DRAM capacity 4Q26>4Q27 from 700k wpm to 820k wpm. This is on the back of P4 finishing its ramp and P5-1 starting to get equipment in July 2027. After this, in 2028, it is possible that Samsung would aim to start ramping up one clean room in P5-1 whilst the precedent one is still getting in equipment, and/or debottleneck P2/P3. This could allow Samsung to add >150k wpm in DRAM capacity in 2028, even assuming some NAND capacity addition in P5. This may not materially impact supply / demand before late 2028. ## Correction in memory stocks may present an attractive opportunity Memory stocks are now down 19% from the more recent peak in September, almost back to the July trough. We think this should push memory companies to get going on shareholder returns. At current average '27-30E FCF yields for our coverage (ex-CXMT) of 29%, we remain positive on the group. Our top pick remains Key Call Buy Samsung Electronics, whilst we maintain Buy ratings on all stocks in our coverage universe. ## Equities Global Semiconductors Nicolas Gaudois Analyst nicolas.gaudois@ubs.com +65-6495 5148 Kenji Yasui Analyst kenji.yasui@ubs.com +81-3-5208 6211 Jimmy Yu Analyst S1460517080002 jimmy.yu@ubs.com +86-21-3866 8880 Jimmy Yoon Analyst jimmy.yoon@ubs.com +65-6495 4617 Luke Yoo Analyst luke.yoo@ubs.com +82-2-3702 8132 Atsuhiro Kinoshita Analyst atsuhiro.kinoshita@ubs.com +81-3-5208 6768 Sunny Lin Analyst sunny.lin@ubs.com +886-2-8722 7346 Francois-Xavier Bouvignies Analyst francois.bouvignies@ubs.com +44-20-7568 7105 Randy Abrams Analyst randy.abrams@ubs.com +886-2-8722 7338 ## KEY CHARTS Figure 1: AI server GPU/accelerator unit forecasts ('000 units) HBM Gen. HBM Content (GB) 2025 1Q26 2Q26 3Q26E 4Q26E 2026E 1Q27E 2Q27E 3Q27E 4Q27E 2027E NVIDIA 2025E 2026E 2027E L40S GDDR6 48 38 H200 HBM3E 141 593 H20 / H20E HBM3 / 3E 96 441 B200 (8Hi) HBM3E 192 1,441 GB200 (2x GPU - 8Hi) HBM3E 384 384 1,080 103 103 B300 (12Hi) HBM3E 288 288 288 587 267 300 200 100 867 50 50 GB300 (2x GPU - 12Hi) HBM3E 576 576 576 778 658 670 580 450 2,358 360 270 180 90 900 R200 (12Hi) HBM4 288 288 20 200 320 540 360 400 320 240 1,320 RV200 (2x GPU - 12Hi) HBM4 576 576 40 250 450 740 600 720 690 630 2,640 R300 (8Hi) HBM4 192 250 250 R300 (8Hi) HBM4E 256 400 400 RV300 (2x GPU - 8Hi) HBM4 384 250 250 RV300 (2x GPU - 8Hi) HBM4E 512 550 550 *Other GPU/Accel HBM3 / 3E 288 384 576 107 60 80 80 80 300 70 70 80 80 300 Total Units 5,066 1,088 1,110 1,310 1,400 4,908 1,440 1,460 1,770 1,990 6,660 % YoY 3% -17% 0% 3% 3% -3% 32% 32% 35% 42% 36% % QoQ -20% 2% 18% 7% 3% 1% 21% 12% Total SiPs 6,924 1,849 1,820 2,140 2,300 8,109 2,400 2,450 2,890 3,260 11,000 % YoY 38% 27% 23% 15% 8% 17% 30% 35% 35% 42% 36% % QoQ -13% -2% 18% 7% 4% 2% 18% 13% AMD MI300 HBM3 192 71 MI325X (12Hi) HBM3E 256 256 224 18 18 MI355X (12Hi) HBM3E 288 288 221 170 150 80 70 470 MI308X HBM3 192 192 96 29 28 5 62 MI450 (12Hi) HBM4 432 432 45 230 275 350 450 470 380 1,650 MI5XX (16Hi) HBM4E 1024 60 160 220 *Other GPU/Accel HBM3 192 192 288 4 1 1 1 1 6 2 2 2 2 8 Total Units 616 218 179 131 301 831 352 452 532 542 1,878 % YoY 41% 81% 86% -26% 37% 35% 61% 152% 305% 80% 126% % QoQ -1% -18% -27% 129% 17% 28% 18% 2% Google TPU v5e HBM2E 32 134 TPU v5p HBM2E 96 1,916 TPU v6e HBM3E 32 32 426 205 224 141 70 640 TPU v7 HBM3E 192 192 192 90 520 720 800 750 2,790 600 450 300 150 1,500 TPU v8i (Dual die AX) HBM3E 288 288 25 225 250 700 1,000 1,300 1,000 4,000 TPU v8t (Single die X) HBM3E 216 216 70 380 450 700 1,200 1,500 1,000 4,400 *Other TPU/Accel HBM3 96 144 144 77 7 9 10 10 37 25 25 25 25 100 Total Units 2,643 732 953 1,046 1,435 4,167 2,025 2,675 3,125 2,175 10,000 % YoY 19% 14% 66% 82% 68% 58% 177% 181% 199% 52% 140% % QoQ -14% 30% 10% 37% 41% 32% 17% -30% Amazon Inferentia 2 HBM2E 32 88 Trainium 2 / 2.5 HBM3E 96 144 144 496 130 220 140 100 590 50 10 60 Inferentia 3 HBM3 96 96 96 162 30 28 20 18 96 15 10 25 Trainium 3 HBM3E 144 144 320 620 840 1,780 840 780 640 540 2,800 Trainium 4 HBM4 256 40 240 280 *Other Accel HBM3 144 144 144 37 5 6 8 10 28 8 8 7 7 30 Total Units 783 165 574 788 968 2,494 913 808 687 787 3,195 % YoY -8% -17% 180% 302% 424% 218% 454% 41% -13% -19% 28% % QoQ -11% 248% 37% 23% -6% -12% -15% 15% Intel / Habana Gaudi 3 HBM2 / 3E 288 288 188 30 18 8 56 Gaudi 4 HBM3E 288 288 10 25 35 25 20 15 10 70 *Others HBM2 / 3E 192 192 192 25 6 4 4 5 18 4 4 4 4 16 Total Units 213 36 22 22 30 109 29 24 19 14 86 % YoY -32% -20% -60% -63% -46% -49% -19% 11% -12% -53% -21% % QoQ -36% -40% 0% 39% -3% -17% -21% -26% Microsoft Maia 100 HBM3 120 120 34 3 1 4 Maia 200 HBM3E 288 288 2 6 10 14 33 14 10 7 4 35 Maia 300 HBM4 288 8 14 22 *Others HBM3 120 192 192 8 2 2 2 2 8 2 2 2 2 8 Total Units 42 7 9 12 16 45 16 12 17 20 65 % YoY 22% -27% -19% 4% 84% 7% 119% 35% 36% 24% 45% % QoQ -17% 22% 40% 29% 0% -25% 42% 18% Meta MTIA v1 LPDDR5 64 64 421 20 5 25 MTIA v2 LPDDR5 128 128 30 20 8 2 30 MTIA v3 HBM3E 192 192 192 5 15 20 35 40 110 40 30 20 15 105 MTIA v4 HBM4 432 40 80 100 120 340 *Others LPDDR5 64 128 128 28 3 1 1 1 6 5 5 5 5 20 Total Units 484 58 34 38 41 171 85 115 125 140 465 % YoY 19% -43% -71% -71% -68% -65% 47% 235% 228% 240% 171% % QoQ -56% -41% 11% 8% 106% 35% 9% 12% Tesla Dojo HBM3 64 64 64 62 17 18 15 14 64 15 15 18 15 63 *Others HBM3 64 64 64 44 7 7 6 6 26 6 6 5 5 22 Total Units 106 24 25 21 20 90 21 21 23 20 85 % YoY -39% -12% -9% -24% -16% -15% -12% -17% 10% 2% -5% % QoQ 2% 6% -17% -7% 7% 0% 10% -13% Others Startups HBM3 192 288 288 657 150 150 155 155 610 170 180 185 190 725 Others HBM2E /3 192 288 288 361 30 30 39 39 138 60 63 65 67 254 Total Units 1,019 180 180 194 194 748 230 243 250 257 979 % YoY 59% -28% -32% -26% -20% -27% 28% 35% 29% 32% 31% % QoQ -26% 0% 8% 0% 18% 6% 3% 3% Source: Company data, UBS estimates Figure 2: HBM demand forecasts by GPU/accelerator HBM Gen. HBM Content (GB) 2025 1Q26 2Q26 3Q26E 4Q26E 2026E 1Q27E 2Q27E 3Q27E 4Q27E 2027E NVIDIA 2025E 2026E 2027E H200 HBM3E 141 83,571 H20 / H20E HBM3 / 3E 96 42,336 B200 (8Hi) HBM3E 192 276,649 GB200 (2x GPU - 8Hi) HBM3E 384 384 414,872 39,504 39,504 B300A (12Hi) HBM3E 144 144 0 0 0 0 0 0 B300 (12Hi) HBM3E 288 288 288 169,176 76,896 86,400 57,600 28,800 249,696 14,400 14,400 GB300 (2x GPU - 12Hi) HBM3E 576 576 576 448,259 379,008 385,920 334,080 259,200 1,358,208 207,360 155,520 103,680 51,840 518,400 R200 (12Hi) HBM4 288 288 5,760 57,600 92,160 155,520 103,680 115,200 92,160 69,120 380,160 RV200 (2x GPU - 12Hi) HBM4 576 576 23,040 144,000 259,200 426,240 345,600 414,720 397,440 362,880 1,520,640 R300 (12Hi) HBM4 192 48,000 48,000 R300 (12Hi) HBM4E 256 102,400 102,400 RV300 (4x GPU - 12Hi) HBM4 384 96,000 96,000 RV300 (4x GPU - 12Hi) HBM4E 512 281,600 281,600 *Other GPU/Accel HBM3 / 3E 288 384 576 30,787 23,040 30,720 30,720 30,720 115,200 40,320 40,320 46,080 46,080 172,800 Total ('000 GB) 1,465,650 518,448 531,840 624,000 670,080 2,344,368 711,360 725,760 783,360 913,920 3,134,400 Total (bn Gb) 11.73 4.15 4.25 4.99 5.36 18.75 5.69 5.81 6.27 7.31 25.08 AMD MI300 HBM3 192 13,685 MI325X (12Hi) HBM3E 256 256 57,294 4,608 4,608 MI355X (12Hi) HBM3E 288 288 63,569 48,960 43,200 23,040 20,160 135,360 MI308X HBM3 192 192 18,394 5,568 5,376 960 11,904 MI450 (12Hi) HBM4 432 432 19,440 99,360 118,800 151,200 194,400 203,040 164,160 712,800 MI5XX (16Hi) HBM4E 1,024 61,440 163,840 225,280 *Other GPU/Accel HBM3 192 192 288 812 274 285 282 281 1,123 576 576 576 576 2,304 Total ('000 GB) 153,754 59,410 48,861 43,722 119,801 271,795 151,776 194,976 265,056 328,576 940,384 Total (bn Gb) 1.23 0.48 0.39 0.35 0.96 2.17 1.21 1.56 2.12 2.63 7.52 Google TPU v5e HBM2E 32 4,282 TPU v5p HBM2E 96 183,913 TPU v6e HBM3E 32 32 13,632 6,554 7,168 4,506 2,253 20,480 TPU v7 HBM3E 192 192 192 17,280 99,840 138,240 153,600 144,000 535,680 115,200 86,400 57,600 28,800 288,000 TPU v8i (Dual die AX) HBM3E 288 288 7,200 64,800 72,000 201,600 288,000 374,400 288,000 1,152,000 TPU v8t (Single die X) HBM3E 216 216 15,120 82,080 97,200 151,200 259,200 324,000 216,000 950,400 *Other TPU/Accel HBM3 96 144 144 7,389 1,044 1,359 1,492 1,440 5,335 3,600 3,600 3,600 3,600 14,400 Total ('000 GB) 226,495 107,437 146,767 181,917 294,573 730,695 471,600 637,200 759,600 536,400 2,404,800 Total (bn Gb) 1.81 0.86 1.17 1.46 2.36 5.85 3.77 5.10 6.08 4.29 19.24 Amazon Inferentia 2 HBM2E 32 2,823 Trainium 2 / 2.5 HBM3E 96 144 144 47,574 18,720 31,680 20,160 14,400 84,960 7,200 1,440 8,640 Inferentia 3 HBM3 96 96 96 15,563 2,880 2,688 1,920 1,728 9,216 1,440 960 2,400 Trainium 3 HBM3E 144 144 46,080 89,280 120,960 256,320 120,960 112,320 92,160 77,760 403,200 Trainium 4 HBM4 256 10,240 61,440 71,680 *Other Accel HBM3 144 144 144 5,375 691 818 1,123 1,380 4,012 1,152 1,152 1,008 1,008 4,320 Total ('000 GB) 71,336 22,291 81,266 112,483 138,468 354,508 130,752 115,872 103,408 140,208 490,240 Total (bn Gb) 0.57 0.18 0.65 0.90 1.11 2.84 1.05 0.93 0.83 1.12 3.92 Intel Gaudi 3 HBM2 / 3E 288 288 54,212 8,640 5,184 2,304 16,128 Gaudi 4 HBM3E 288 288 2,880 7,200 10,080 7,200 5,760 4,320 2,880 20,160 *Others HBM2 / 3E 192 192 192 4,789 1,152 691 691 960 3,494 768 768 768 768 3,072 Total ('000 GB) 59,001 9,792 5,875 5,875 8,160 29,702 7,968 6,528 5,088 3,648 23,232 Total (bn Gb) 0.47 0.08 0.05 0.05 0.07 0.24 0.06 0.05 0.04 0.03 0.19 Microsoft Maia 100 HBM3 120 120 4,050 360 120 480 Maia 200 HBM3E 288 288 660 1,696 3,016 4,052 9,424 4,032 2,880 2,016 1,152 10,080 Maia 300 HBM4 288 2,304 4,032 6,336 *Others HBM3 120 192 192 960 384 384 384 384 1,536 384 384 384 384 1,536 Total ('000 GB) 5,010 1,404 2,200 3,400 4,436 11,440 4,416 3,264 4,704 5,568 17,952 Total (bn Gb) 0.04 0.01 0.02 0.03 0.04 0.09 0.04 0.03 0.04 0.04 0.14 Meta MTIA v3 HBM3E 192 192 192 960 2,880 3,840 6,720 7,680 21,120 7,680 5,760 3,840 2,880 20,160 MTIA v4 HBM4 432 17,280 34,560 43,200 51,840 146,880 Total ('000 GB) 960 2,880 3,840 6,720 7,680 21,120 24,960 40,320 47,040 54,720 167,040 Total (bn Gb) 0.01 0.02 0.03 0.05 0.06 0.17 0.20 0.32 0.38 0.44 1.34 Others Tesla HBM3 64 64 6,780 1,523 1,613 1,344 1,255 5,735 1,344 1,344 1,472 1,280 5,440 Startups / Etc. HBM2E / 3 192 288 195,559 51,840 51,840 55,800 55,800 215,280 66,096 69,984 71,928 73,872 281,880 Total ('000 GB) 202,339 53,363 53,453 57,144 57,055 221,015 67,440 71,328 73,400 75,152 287,320 Total (bn Gb) 1.62 0.43 0.43 0.46 0.46 1.77 0.54 0.57 0.59 0.60 2.30 Total HBM Demand ('000 GB) 2,184,545 775,025 874,103 1,039,582 1,332,654 4,021,364 1,604,832 1,831,968 2,098,536 2,154,672 7,690,008 Total HBM Demand (bn Gb) 17.48 6.20 6.99 8.32 10.66 32.17 12.84 14.66 16.79 17.24 61.52 % YoY 103% 97% 103% 72% 77% 84% 107% 110% 102% 62% 91% % QoQ 3% 13% 19% 28% 20% 14% 15% 3% Source: Company data, UBS estimates Figure 3: HBM vendors' bit shipments (M 1Gb) 2025 Q126 Q226 Q326E Q426E 2026E Q127E Q227E Q327E Q427E 2027E SK Hynix 12,628 3,803 3,804 4,559 5,273 17,438 4,611 4,812 5,058 5,587 20,068 % YoY 86% 31% 20% 41% 59% 38% 21% 26% 11% 6% 15% % QoQ 15% 0% 20% 16% -13% 4% 5% 10% Samsung 4,327 1,519 2,061 3,384 4,452 11,416 4,891 5,281 5,864 6,356 22,392 % YoY -15% 161% 170% 140% 183% 164% 222% 156% 73% 43% 96% % QoQ -3% 36% 64% 32% 10% 8% 11% 8% Micron 4,584 1,612 1,821 1,985 2,243 7,662 2,468 2,714 3,067 3,466 11,716 % YoY 417% 106% 60% 57% 60% 67% 53% 49% 55% 55% 53% % QoQ 15% 13% 9% 13% 10% 10% 13% 13% CXMT 53 21 26 31 44 122 53 74 111 165 403 % YoY 168% 142% 93% 139% 130% 149% 191% 264% 272% 232% % QoQ 15% 20% 20% 44% 20% 40% 50% 48% Total Supply 21,591 6,955 7,712 9,959 12,012 36,638 12,022 12,882 14,100 15,574 54,579 % YoY 69% 63% 52% 68% 90% 70% 73% 67% 42% 30% 49% % QoQ 10% 11% 29% 21% 0% 7% 9% 10% Source: Company data, UBS estimates Figure 4: Adjusted HBM supply vs demand ![](images/ca98f3553f2ff31321b541126b93ab41a53be52413846a211a39947a8a34cffc.jpg) Source: Company data, UBS estimates Figure 5: HBM as % of total DRAM revenues ![](images/a123afaa68d895e3552e8e74aa0a6375db7ed93ba9b2f01e50eb21b925cb8fa1.jpg) Source: UBS estimates Figure 6: HBM as % of total DRAM bits ![](images/3aeafe50812cbf505f1a473cab65e09b5fd22425cd639dbd958c9720566a6f5c.jpg) Source: UBS estimate Figure 7: HBM ASP forecasts (US\$/1Gb) ![](images/bf71484fea5ed6a7110b104880ebf34a2348d4f9829aae151fd203f32a39df08.jpg) Source: UBS estimates Figure 8: HBM bit share by vendor ![](images/66507cdb8a981712767ae698e4e4fa2382efb442d1781093e9a287302c6c6447.jpg) Source: Company data, UBS estimates Figure 9: DRAM industry capacity (k wpm) breakdown between DDR and HBM ![](images/27f6afbfbe87906efa70244edec75beed05d42ab18f23be5ad60fd17158eed0b.jpg) Source: Company data, UBS estimates Figure 10: DRAM capacity (k wpm) breakdown between DDR and HBM supplier ![](images/d04714db5b4847fbff8e68f3c8203b31676041301042cb933f14aa4853ca624b.jpg) Source: Company data, UBS estimates Figure 11: HBM demand by type (2026E) ![](images/71f368f9c7ddee4b1171022481e23345053b823d5062f2dd65717db90bda0845.jpg) Source: UBS estimates Figure 12: HBM demand by type (2027E) ![](images/cc7b48c6f548e3a371f44238883e64f8b6c13532714ea1343e195ede3de66482.jpg) Source: UBS estimates Figure 13: HBM demand by customer (2026E) ![](images/05a808335ab04736a043f89bce5d5ab54adb2d594238d2058a1a7f51c40e46fb.jpg) Source: UBS estimates Figure 14: HBM demand by customer (2027E) ![](images/509e18a37e16ae2a33bdb2ec6efd06d0b4ee744c4c473b18076c5270034f020f.jpg) Source: UBS estimates Figure 15: DDR cycles in context ![](images/af1d6e163badb5aa9b7335a82d0a8483eefbef63d30491240c4eaf1f508913bb.jpg) Source: Company data, UBS estimates Figure 16: NAND cycles in context ![](images/d1fac2c4819aded5f947cacd2db652ab2a8a8cb2af25ec2c5291a01a44316d1c.jpg) Source: Company data, UBS estimates Figure 17: DRAM bit demand per application (2026E) ![](images/b0c4a6dfdf86f2b965b7c1e85cdf48f9fb08f13f6b7ce795be6273b3876fa7c6.jpg) Source: Gartner, UBS estimates Figure 18: DRAM bit demand per application (2027E) ![](images/0b976208c526fa7bb0971fe63ae203eddb4ee82f4f871e925a03acf8f9361cbf.jpg) Source: Gartner, UBS estimates Figure 19: NAND bit demand per application (2026E) ![](images/7045b6278948c78b7fafd9dad7374e2bfdb933dcbdee556491803102ae5c8561.jpg) Source: Gartner, UBS estimates Figure 20: NAND bit demand per application (2027E) ![](images/08671b7b1e19e3d3abd22df88cb9eb53a4cf2297521a64681a5dd9fed973e6d9.jpg) Source: Gartner, UBS estimates Figure 21: Memory as % of flagship/premium smartphone BOM costs, 1Q22-4Q27E ![](images/9d5c0edc4331eb22f474ac8c65e80e42be71078ce0aa32a7a2d305ab77d38d09.jpg) Source: Counterpoint, Fomalhaut, UBS estimates Figure 22: Memory as % of mid-range/lower-end smartphones, 1Q22-4Q27E ![](images/94646acaf3678076bc540c11957bd5ff7b34c5fa83dc0a9704d38d0dc908cede.jpg) Source: UBS estimates Figure 23: DRAM industry wafer starts and YoY growth ![](images/4adb8d169b5f0b78e3ab297a0bf8abfa7d62d28e5678459ec1abd82650f7fa22.jpg) Source: Company data, UBS estimates Figure 24: NAND industry wafer starts and YoY growth ![](images/6cdb8905f564f53715493e8b601100b607bda7d5cc803684e0766fd7ba280975.jpg) Source: Company data, UBS estimates Figure 25: DRAM industry bit supply and YoY growth ![](images/adf2baf01af25211fad95ec7885e158b4d6ecfa5704b106876e98203ff908928.jpg) Source: Company data, UBS estimates Figure 26: NAND industry bit supply and YoY growth ![](images/e276b686c832212c743bbb6e3d3a59f0b3ca44e17be7e59c89af71ccf7f83852.jpg) Source: Company data, UBS estimates Figure 27: CXMT wafer capacity and as % of DRAM industry ![](images/0ba74b124b7449759625054870b41bd735a9fe5c29f40a3214be1a60704ad649.jpg) Source: UBS estimates Figure 28: CXMT bit shipments and as % of DRAM industry ![](images/fd881344555b43e23180693f8e60d86d3235633e87f278ea596ff6b89fffc0f6.jpg) Source: UBS estimates Figure 29: UBS DRAM supply/demand summary Q125 Q225 Q325 Q425 2025 Revenues ($ mil) 26,415 31,776 39,234 52,958 150,384 % QoQ -7.3% 20.3% 23.5% 35.0% % YoY 46.4% 39.0% 54.3% 85.8% 58.6% Supply (Shipments Based) (M 1Gb) 64,317 75,938 85,576 87,400 313,229 % QoQ -2.6% 18.1% 12.7% 2.1% % YoY 3.4% 14.2% 28.9% 32.3% 19.9% Demand (End Consumption) (M 1Gb) 69,879 73,066 87,229 94,120 324,295 % QoQ 0.2% 4.6% 19.4% 7.9% % YoY 17.5% 21.0% 35.1% 34.9% 27.6% DDR Demand (M 1Gb) 306,809 % YoY 24.9% Inventory adjustment -6,150 3,250 200 -700 -3,400 Inventory Adj. Demand (M 1Gb) 63,729 76,316 87,429 93,420 320,895 % QoQ -5.8% 19.8% 14.6% 6.9% % YoY 1.6% 10.5% 28.7% 38.1% 20.0% Wafer out (k wpm, 12" equi.) 1,716 1,764 1,804 1,854 1,784 % QoQ 2.5% 2.8% 2.3% 2.8% % YoY 17.9% 16.5% 13.4% 10.8% 14.5% % Sufficiency Ratio 0.9% -0.5% -2.1% -6.4% -2.4% Blended ASP incl. HBM ($/Gb) 0.41 0.42 0.46 0.61 0.48 % QoQ -4.8% 1.9% 9.6% 32.2% % YoY 41.5% 21.8% 19.7% 40.4% 32.2% DDR ASP ($/Gb) 0.33 0.35 0.38 0.53 0.40 % QoQ -6.2% 7.0% 7.0% 42.0% % YoY 20.3% 10.6% 11.8% 52.5% 26.4% HBM ASP ($/Gb) 1.42 1.55 1.56 1.52 1.52 % QoQ 2.7% 9.4% 0.4% -2.5% % YoY 25.6% 28.3% 16.9% 10.0% 15.6% Source: Company data, UBS estimates Q126 Q226 Q326E Q426E 98,124 146,329 193,867 227,347 85.3% 49.1% 32.5% 17.3% 271.5% 360.5% 394.1% 329.3% 89,789 96,748 103,476 109,808 2.7% 7.8% 7.0% 6.1% 39.6% 27.4% 20.9% 25.6% 89,600 91,657 100,847 109,294 -4.8% 2.3% 10.0% 8.4% 28.2% 25.4% 15.6% 16.1% 17,300 13,200 6,700 4,000 106,900 104,857 107,547 113,294 14.4% -1.9% 2.6% 5.3% 67.7% 37.4% 23.0% 21.3% 1,889 1,907 1,929 1,964 1.9% 1.0% 1.2% 1.8% 10.1% 8.1% 7.0% 5.9% -16.0% -7.7% -3.8% -3.1% 1.09 1.51 1.87 2.07 80.4% 38.4% 23.9% 10.5% 166.1% 261.4% 308.6% 241.7% 1.05 1.51 1.88 2.09 97.0% 43.0% 25.0% 11.0% 220.3% 328.0% 400.0% 290.9% 1.57 1.59 1.79 1.91 3.0% 1.4% 12.7% 7.0% 10.3% 2.3% 14.8% 26.0% 2026E Q127E Q227E Q327E Q427E 2027E 665,667 258,641 288,546 316,883 337,869 1,201,939 13.8% 11.6% 9.8% 6.6% 342.6% 163.6% 97.2% 63.5% 48.6% 80.6% 399,821 114,091 121,028 128,428 133,830 497,376 3.9% 6.1% 6.1% 4.2% 27.6% 27.1% 25.1% 24.1% 21.9% 24.4% 391,398 111,919 124,474 143,094 164,161 543,648 2.…

UBS raised Samsung capex forecasts again；SOCAMM orders for Nvidia Vera CPUs now exceed 50bn Gb in 2027

## U.S. Software # C3Q Networking Security Preview ## Summary Into 3Q, we remain Neutral on both FTNT and CHKP, despite very different setups. For FTNT, our checks remain constructive across firewall and SASE. However, following significant upward revisions to both product and subs expectations this year, we think continued outperformance is increasingly expected, particularly with shares trading at \~38x our CY28E FCF. For CHKP, expectations remain much lower, with shares down \~26% YTD versus the BUG up \~60% despite a strong firewall backdrop. While that creates a more forgiving setup, our checks provide little evidence of an underlying inflection for CHKP, with growth still driven largely by renewals and continued challenges around net-new business and broader platform adoption. ## Summary of checks Firewall continued to sound healthy in 3Q, supported by ongoing refresh cycles and strong NGFW demand. That said, we heard some evidence this Q that unit growth in firewalls is starting to slow, with revenue growth still buoyed by cumulative 30-40% price increases. Into 4Q and 2027, we see a tougher setup as the group starts to face harder comps. Our firewall industry model suggests that growth should decelerate in 3Q, even assuming typical 3Q seasonality. We think continued acceleration would require sustained demand for firewalls tied to AI adoption (partners were mixed on how SASE demand impacts growth) and particularly opportunities with neoclouds and AI factories. Our preliminary work here suggests firewall density within AI data centers isn't necessarily increasing and the main opportunity lies in north-south throughput. Within SASE, partners generally described a healthy but increasingly saturated market, with fewer large enterprise greenfield opportunities but continued opportunity across the mid-market. Among the group, our checks were more constructive on Fortinet, where partners continue to highlight strength in NGFW, traction in FortiSASE, and growing interest in opportunities such as OT. Partners were more cautious on Check Point, as it continues to face challenges driving net-new business and broader platform adoption beyond areas such as Harmony. We include commentary below. ## FTNT - reports 10/28 AMC We remain constructive on underlying trends heading into 3Q, with our checks continuing to suggest healthy firewall demand, but prefer to remain on the sidelines given elevated expectations and valuation at \~38x CY28E EV/FCF, which we think is underwriting 20%+ growth in CY27. We model 29% product growth in 3Q, which would imply \~10% growth in 4Q to reach \~31% FY guidance. On subs, we model another acceleration to \~14% growth in 3Q, with current subs billings accelerating toward \~20% exiting the year, supported by FortiSASE traction. With both product and subs expectations having moved meaningfully higher throughout the year, we think continued outperformance is increasingly expected and reiterate our Neutral rating. ## CHKP - reports 11/2 BMO Expectations remain low following the reset earlier this year, and we remain cautious heading into 3Q earnings. CHKP is down \~26% YTD versus the BUG up \~60%, despite a strong firewall refresh cycle that CHKP has struggled to capture. That said, guidance calling for product revenue to decline \~19% y/y in 3Q, before growing \~15% in 4Q against an easier comp, looks very conservative and would represent the single worst 3Q q/q product growth performance in the company's history. While the 3Q total billings comp is tough, the model gets easier into 4Q/CY27. Better execution would make this a more interesting setup; however, our checks provided little evidence of growth, and it's increasingly evident in our view that a lack of SASE traction is costing CHKP mindshare (PANW/FTNT currently growing SASE \~30%/20%). ## Equities Americas Software Roger Boyd Analyst roger.boyd@ubs.com +1-212-713 1256 Nicholas Hodge Associate Analyst nicholas.hodge@ubs.com +1-212-882 5954 ## Firewall market We spoke with a handful of partners, experts, and customers, highlighting a handful of dynamics across the firewall market. In digesting commentary, we'd argue checks continue to sound healthy through 3Q, with refresh activity remaining solid and some evidence that the current cycle still has runway, but some early cautionary signals that parts of the firewall market could start to slow. Several partners described 3Q refresh momentum as continuing or accelerating, and one partner was optimistic on hardware appliance demand to remain elevated into CY27. On the other hand, one partner characterized firewall unit growth as roughly flat across the first three quarters and indicated that pull-forward conversations are shifting to push-outs, as customers look to wait out pricing surcharges. Looking longer term, checks continue to debate how durable traditional firewalls will be as network architectures shift toward SASE. \- Refresh dynamics: Partners continue to see healthy firewall refresh activity, with one describing the pace of refreshes as largely consistent with 2Q, while another saw hardware firewall refreshes accelerate throughout 3Q and expects the resurgence to persist for another 2-3 quarters. Another partner described demand for next-generation firewall equipment as tied to data center migrations. One customer noted that they are expecting a 20-30% increase that will require them to explore all options over the next 3-6 months. \- Inventory dynamics: Commentary suggests inventory availability itself is generally manageable, although checks were mixed. Two partners indicated previous hardware inventory challenges have improved and/or they don't see firewalls suffering from an inventory issue at the moment, while another continues to see inventory constraints across next-generation firewall appliances amid strong demand. One partner suggested that distributor level inventory is no longer increasing. \- SASE vs firewall debate: Checks offered contrasting views on the longer-term durability of traditional firewalls as enterprises evolve their network architectures. One partner sees a significant shift away from traditional firewalls outside the data center, with SASE ultimately taking a greater role across campus and remote-office environments. Conversely, another partner is seeing enterprises reconsider the shift from hardware toward software-defined architectures, as customers recognize that moving everything to the SASE/SSE layer may not provide the security, segmentation, and control they want, contributing to a resurgence in hardware firewall refreshes. Interestingly, the partner arguing a broader shift toward SASE cited customers continuing to deploy significant quantities of firewalls across remote offices even with SASE already in their environments. Partner #1: I would say that from a unit perspective in firewall, directionally across the first three quarters of the year I would characterize as flat. However, with some of these effective price increases that we have seen in the ranges of 30-40%s sometimes, overall sales and gross revenue are up. In terms of inventory, I don't think anyone is stocking up anymore with these surge prices. From what I have heard, it looks like 12 to 18 months before the memory supply shortage kind of straightens out. So, down the road, this all might become a bigger pressure on inventory, but today I haven't really seen any problems. There has not been an inventory problem. Partner #2: I see the momentum that we are seeing in firewall with refreshes continuing as vendors are focused on proposing renewals in advance or with new technologies that will have an increased list price. So I would say that the pace of the refresh has largely been consistent vs 2Q. In the past, we saw some challenges in terms of inventory for firewalls but more generally about hardware technologies. At the moment, the situation has definitely improved, and I wouldn't say that firewalls are suffering from any negative trend in this area. Partner #3: For the entire third quarter, we have seen a resurgence on the hardware side of the business as firewall refreshes have accelerated. This has been consistent right through the end of September and will likely stay for the next 2 to 3 quarters in my view. We really saw this with Palo in July, and to some extent with Fortinet. I think the reason for this is a couple of factors. One part is that the hardware refresh, where customers are recognizing that the preconceived notion of converting all hardware infrastructure into software-defined may not serve the ultimate purpose of becoming more secure and have more control from a segmentation standpoint. So more and more enterprises want to retain control of the hardware layer rather than disposing of everything to the SASE or SSE layer. Another is the overall concerns that are being raised on mass with all these AI attacks and development. So customers, especially large enterprises, are willing to invest more in control. Partner #4: The campus layer, remote offices, etc., we are seeing a big shift away from the traditional firewalls with most people going the SASE route. I think the traditional firewall market outside the data center is going to continue to trend down unless it can be better interwoven into the entire security strategy of an organization. I just don't see hardware devices persisting outside the data center level. That said, I have very forward-thinking customers who are leveraging AI, agentic employees, etc., but can't wrap their heads around deploying SASE across their enterprise. So, instead they have SASE in the environment but bought so many Palo firewalls for remote offices that it was kind of shocking. So, as much as I have said how everything is going in the SASE direction, there are still companies that don't have the talent to make the switch so they'll keep doing what worked in the past despite potential performance gains, savings, and improvements to their general architecture. Partner #6: High-level, we still have seen quite a bit of inventory constraints and that is not just the case for Fortinet and Palo, I think it is across the board for their next-gen firewall solutions and appliances. I think part of it is still issues around chip costs and manufacturing. But this combined with huge demand doesn't help. We have data center customers that have huge demand for next-generation firewall equipment because it is more efficient, higher bandwidth, more capacity, etc. And so we have seen this as we help customers do data center migrations and orders for brand-new equipment. We have still seen demand at an extremely high level. Customer #1: We haven't refreshed our firewalls. After the whole memory prices going through the roof, we haven't had a refresh cycle come through, but we are close. I think it will be end of next year that we will have done it, which means in the next three to six months we are going to start looking and comparing. We are a little bit prepared for some kind of uptick, but I don't know if we are prepared for a doubling or a 20% or 30% increase. In our Firewall market model (see below), 3Q guidance for the four vendors suggests that the product revenue cycle peaked in 2Q26. Even if we assume typical q/q seasonality (median of the last 4 q/q growth) for all four vendors in 3Q and 4Q, it still implies that the product market peaked in 2Q after experiencing higher than typical seasonality (impacted by the bulk of price increases done so far). Overall, this feels fairly consistent with cycles seen in 2018 and the 2021-2022 supercycle. That said, industry forecasts (IDC, Gartner) suggest memory supply/pricing could remain elevated through 2027, with additional fab capacity beginning to ease supply in 2028 as AI demand continues to redirect capacity toward HBM and other higher-value memory products. On a multiyear basis, we would argue that overall product revenue growth is in the low-double-digits, with a slight downward trajectory towards high-single-digits over the past 10 years. Figure 1: Firewall Market ![](images/e221ca5c82b50388badbd5208d825ff30d33eb0c4a8bfb99f4373371b3a339e4.jpg) Source: FactSet, company data ## SASE market We also spoke with partners on the SASE market, with checks suggesting demand continues to grow despite a market that is becoming increasingly mature and competitive. However, checks continue to see incremental opportunity, particularly across the mid-market and lower end of enterprise, while another partner characterized the SASE/SD-WAN category as accelerating despite longer sales cycles. At the same time, commentary highlighted the increasingly interconnected nature of the firewall and SASE markets, as customers weigh cloud-delivered security against the control, infrastructure, and economics of retaining hardware. \- Competitive environment: Palo Alto and Zscaler continue to be viewed as leaders across enterprise and large enterprise, although partners pointed to increasing competition from a broader group of vendors. Netskope was cited as gaining share by one partner, particularly against Zscaler, while others also highlighted Cato and Versa as expanding within SMB and Check Point as improving its position. Competition appears particularly broad in the mid-market, where partners see both a greater number of vendors and more remaining opportunity. \- Firewall/SASE architectures: The strength in firewall demand also appears relevant to the evolution of SASE architectures, with one partner highlighting continued hybrid environments as companies weigh cloud costs, infrastructure requirements, and the level of control they want across different layers of the stack. \- Market opportunities: Partners see a SASE market that is more mature than a year ago, particularly within large enterprise, where there are fewer greenfield opportunities and more competitors. However, this has not hindered growth, with one partner seeing particular opportunity among enterprises with 2K-5K employees, while another described the SASE/SD-WAN market as accelerating, despite longer sales cycles than other security segments. Overall, checks suggest the opportunity is increasingly shifting toward the mid-market and lower end of enterprise as penetration within larger organizations matures. Partner #2: The SASE market from our perspective is definitely more mature today than last year. But it is interesting to see that it has continued to grow. It seems that the opportunities in large enterprise are less than a year ago, and there are more players in the market, but there is still space for new opportunities. I see this especially in the mid-market and enterprises, with companies in the 2K to 5K employee range. We see Palo Alto and Zscaler as the leaders in the enterprise and large enterprise. I would say Netskope is growing and gaining some market share, especially against Zscaler. The mid-market is totally different, as we see more players, but more opportunities. I think Zscaler is suffering the competition more than others there. Also, Palo doesn't fit well for small businesses like a Cato Networks would. Partner #5: Overall, the SASE and SD-WAN category is moving relatively well, but there are more competitors. In the past, when we were talking about SASE, it was basically Palo Alto, Fortinet, and Cisco. Now Cisco is probably losing some ground, and there are others like Zscaler or even Check Point that are doing better. There are other emerging players that are more agnostic, like Versa or Cato, that are expanding specifically within the small- and medium-business market. But I would describe the SASE market as one that is accelerating but it is lower than other segments because the projects have longer sales cycles. Partner #6: I think that the resurgence or strength in firewall influences the market for SASE. I think when you look at how much data and throughput you actually need leveraging tokenization usage, the bandwidth needed when self-hosting a lot of models, highly-regulated companies require more infra, you are expanding the blast radius. We are still seeing a lot of hybrid environments where companies are bringing things back onshore or are starting to do a lot of self-hosting. So as cloud costs get exponentially more expensive, it ultimately depends on how much control you want at what layer of the stack. ## Fortinet (FTNT) - reports 10/28 AMC We remain constructive on the underlying trends heading into 3Q earnings, but prefer to remain on the sidelines given elevated expectations and valuation at \~38x our CY28E FCF. Our firewall checks remained healthy, with continued refresh activity and some evidence of runway into coming quarters, while FTNT-specific checks were broadly constructive across NGFW, FortiSASE, AI data centers, and emerging areas such as OT security. However, following two quarters of upward revisions to FY expectations and significant share price appreciation (up \~140% YTD), we think continued outperformance is at least partially reflected in the stock, leaving the debate centered on the magnitude/durability of growth from here, and durability of opportunities at the neocloud/AI factory level. On product, we model \~29% y/y growth in 3Q, which would imply \~10% growth in 4Q to reach the current FY guidance framework of \~31%. Our checks remain supportive, with partners highlighting healthy NGFW demand, new hardware refreshes, and continued demand for larger appliances. We also continue to see AI data centers/neoclouds as an important potential driver of the product cycle. Following the 8-fig neocloud win highlighted in 2Q, management indicated the opportunity set continues to build across neoclouds and enterprise AI data centers, despite it still being at an early stage. Management continues to view the broader AI-driven infrastructure cycle as relatively early, potentially providing another source of durability to product growth beyond traditional refresh activity. That said, following the significant upward revisions already made this year, we think another product beat and raise is increasingly important to supporting expectations. On subscriptions, we remain constructive on the accelerating trajectory. We model subs revenue growth continuing to accelerate to \~14.2% in 3Q, with the current FY framework implying another modest step-up to \~14.5% in 4Q on our estimates. Notably, FTNT has also raised its FY subscription growth framework in each of the last two quarters, from \~11% entering the year to \~12% and ultimately \~13.5%. We also model current subs billings accelerating to \~19%/20% in 3Q/4Q, or \~17% for FY26, while our checks suggest FortiSASE continues to perform well. Overall, we remain constructive on the fundamental backdrop, but believe the current valuation and elevated expectations appropriately keep us Neutral rated. ## PRE-QUIET CHECK-IN TAKES We caught up with Fortinet management ahead of their quiet period and our key takeaways were as follows. Management continues to see customers adopting higher-capacity firewall appliances across the portfolio, although it remains difficult to determine how much of this is being driven by AI-related traffic growth versus customers preparing for future demand. Regardless, newer platforms support capabilities and features that are not available on legacy hardware and appear to be contributing to refresh activity. On AI, management highlighted a growing pipeline of opportunities spanning both neocloud providers and enterprise AI data centers, noting that while neocloud projects are generally larger, enterprise opportunities are currently more numerous. In discussing the durability of product growth beyond FY26, management remained constructive on the outlook and reiterated its view that the industry remains in the early stages of an AI-driven infrastructure cycle. Management also reiterated that OT security remains a fast-growing area of the b…

UBS preliminary work suggests firewall density within AI data centers is not necessarily increasing, with main opportunity being north-south throughput；FTNT highlighted an 8-figure neocloud win in 2Q, with management indicating opportunity set continues to build across neoclouds and enterprise AI data centers

飞塔信息（Fortinet）三季度营收2 飞塔信息（Fortinet）三季度营收20.8亿美元，分析师预期20.7亿美元。 三季度Billings（账单收入/订单金额）为23.4亿美元，分析师预期23亿美元。 三季度调整后每股收益（EPS）为0.91美元，分析师预期0.86美元。 预计全年营收81.3亿-82.2亿美元，公司原本预计80.2亿-81.8亿美元。 预计全年服务营收51.5亿-51.7亿美元，分析师预期52亿美元。 预计全年调整后毛利润率79.7%-80.3%，公司原本预计79%-81%。 预计全年Billings将为95.7亿-97.2亿美元，公司原本预计93.5亿-95.5亿美元。 预计四季度营收21.6亿-22.5亿美元，分析师预期21.7亿美元。 飞塔信息（FTNT）美股盘后暴跌22.5%。 AI 解读 Fortinet（FTNT）Q3 复盘：一次"全面超预期却暴跌 22.5%"的估值出清 一句话结论 性质：短期利空 + 中期估值锚重定价，不是产业需求崩塌。 事实层面 Q3 三项指标全部超预期（营收 20.8 亿 vs 20.7 亿、Billings 23.4 亿 vs 23.0 亿、调整后 EPS 0.91 vs 0.86），全年营收与 Billings 双双上调； 唯一的裂缝是全年服务营收指引 51.5–51.7 亿低于市场预期的 52 亿，叠加调整后毛利率区间由 79%–81% 收窄至 79.7%–80.3%（中值不变） 。由于 EPS 并未下修、全年指引反而上调，22.5% 的跌幅几乎全部来自倍数而非盈利，属于典型的"预期与仓位出清"，市场真正在质疑的是 订阅/经常性收入的增长质量 ，而非订单。影响最直接的资产是 FTNT 自身，其次是对"硬件+防火墙"同业（CHKP、Cisco/Arista/HPE）的映射，纯订阅标的（PANW、CRWD、ZS）方向分歧。 我的倾向：不追空、也不急于抄底；对 FTNT 由中性转谨慎，对网络安全板块中性（情绪扰动大于基本面传导）。 事件背景与上下文 Fortinet 过去 14 个月已形成一条清晰的"指引失望链"：2025 年 8 月因贸易/关税不确定性给出低于预期的 Q3 销售额指引而暴跌；2025 年 11 月因 Q4 收入前景黯淡再度暴跌；2026 年 7 月 Q2 强劲、大幅上调全年（营收 80.2–81.8 亿、调整后 EPS 3.41–3.47）后情绪修复。本次是在 已被上调的高基数上继续超预期，但服务指引没有跟上 ，于是市场不再给予"指引可信度"的溢价。 同期财报季呈现同一范式：泰瑞达（营收/EPS/净利三超，盘后跌 5.8%）、博通（半导体与 AI 收入均超预期，盘后跌 4%）、英特尔（超预期但下季 EPS 指引 0 美元，跌超 5%）—— 当前市场处于"对任何指引瑕疵零容忍"的高预期环境 。相关研报未直接覆盖 FTNT，但提供了"超预期≠上涨"的宏观定价背景，以及企业 IT/算力资本开支仍是主线、安全支出被归入"可延后软件预算"的行业语境。 影响路径 传导链： 服务营收指引低于共识 + 毛利率区间收窄 → 推断营收超预期的构成偏向 低毛利、一次性的产品/硬件 （防火墙更新、关税后补库回补），而经常性收入引擎减速 → 订阅转型故事的"成长质量"被重新定价 → 目标价下修、PE 压缩 → 盘后 22.5% 下跌。 一阶影响： FTNT 自身。服务/订阅是估值溢价来源，Billings 虽超预期（测算全年中值 96.45 亿 vs 原 94.5 亿，+2.1%），但 Q3 Billings 23.4 亿 > 营收 20.8 亿形成的递延差，被市场解读为"确认节奏变慢"而非"未来弹药"。Q4 指引中值测算约 22.05 亿、高于共识 21.7 亿，但 区间下限 21.6 亿低于共识 ，保守姿态可见。 二阶影响： ①若服务疲软源于 FTNT 自身产品组合/云化迁移节奏，则 PANW、CRWD、ZS 在 ARR 份额上可能是 受益方 ；②CHKP 作为最接近的"防火墙+硬件"可比公司，将承受 read-through 抛压；③产品端超预期对网络设备链（Cisco、Arista、HPE-Juniper）与 ODM（Flex、Jabil、Celestica）偏正面；④"安全预算可延后"叙事压制软件板块整体情绪（IGV/纳指软件）。 市场预期差： 市场 低估 了 Billings 持续高于营收所隐含的 2027 年收入能见度与递延转化能力；同时 高估 了"服务指引低于共识 = 需求走弱"这一等式（服务确认受续约定价、递延节奏影响）。反过来，市场 低估 的是连续三次指引失望后，FTNT"廉价成长"信誉基础的动摇——这是倍数层面真实的、非一次性的损伤。 核心分歧 看多（错杀派）： 营收/EPS/Billings 三超 + 全年营收与 Billings 双上调，基本面没有恶化；跌 22.5% 后估值回到历史区间低位，自由现金流收益率吸引力显著上升，属于情绪踩踏。 看空（裂缝派）： FTNT 的估值建立于"订阅转型 + 高毛利 + 高增长"三件套，服务营收指引低于共识是转型故事的第一次实质裂缝；毛利率区间收窄消灭了上行期权；硬件驱动的超预期不可持续（关税/补库效应将回吐）。 最关键的分歧变量： 全年服务营收指引与实际 Billings 之间的缺口，能否在未来两个季度收敛 （即递延收入转化 + 服务/ARR 增速回升）。 我的判断： 站在"利空但被显著放大"一侧。Billings 的超预期与上调是对"需求崩塌论"最有力的反驳；但"成长溢价"的重建需要至少两个季度的服务数据支撑，短期内任何反弹都更可能是交易性而非趋势性。 受影响资产分层 层级 方向 标的/行业/商品 影响逻辑 确定性 时效 直接受损 利空 Fortinet（FTNT） 服务指引低于共识 + 毛利率区间收窄 → 订阅成长质量重定价，跌幅来自倍数 高 日内–1–3 天集中释放，1–2 周波动延续 直接受损 利空 Check Point（CHKP） 同为"硬件防火墙+订阅"模式，估值锚被下修，read-through 抛压 中 1–3 天 间接（方向分歧） 利多/中性 PANW、CRWD、ZS、Netskope 等纯订阅标的 若属 FTNT 自身迁移问题，同业 ARR 份额可能受益；但高倍数标的同受"安全支出减速"情绪拖累 低–中 1–3 天 间接受益 中性偏多 Cisco、Arista、HPE/Juniper；ODM（Flex、Jabil、Celestica） 产品/硬件端超预期，企业网络与安全硬件更新周期仍在兑现 中 1–2 周 情绪映射 利空（弱） 网络安全 ETF（如 HACK、BUGS）、软件指数 IGV 龙头"超预期仍暴跌"强化"指引零容忍"的高预期定价环境 低 日内–1 周 大宗商品/汇率/利率 中性 半导体代工/晶圆、美元、10Y 利率 无实质传导；防火墙 ASIC 自研为主，链条影响可忽略 高（指无影响） 不适用 投资结论与观察点 短期（1–3 天）： 22.5% 的跌幅已把"服务指引瑕疵"定价得相当充分， 不建议追空 ；同样不建议在递延转化数据出来前抢反弹——若出现放量下跌后连续两日无法收复，说明是主动减仓而非被动止损。 中期（1–2 个季度）： 对 FTNT 的定性由"廉价成长"下调为" 待验证的成长质量 "。核心不是营收，而是服务/ARR 增速能否回到指引隐含水平、毛利率能否重新打开上行区间。板块层面，这是一次 估值容忍度 事件，而非安全预算事件。 最值得跟踪的 3 个信号： 1. Q4 及 FY2027 的服务营收指引与 Billings 增速 ——尤其 Billings/营收剪刀差是否继续走阔（走阔=弹药，收敛且服务加速=修复）。 2. 同业交叉验证： PANW 的 ARR 增速与下季指引、CRWD 的模块附加/费率卡数据、CHKP 的 ARR——若同业同步下修，则结论由"公司特有"升级为"板块需求"。 3. 企业 IT 预算与硬件链： Gartner/IDC 安全支出预测是否被下调、Cisco/Arista 订单积压与关税/供应链表态，用于判断产品端超预期是否为一次性补库。 结论会增强的情形： 若同业（尤其 PANW/CHKP）在同窗口出现 ARR 或指引下修，则"安全支出减速"成立，FTNT 及板块的负面判断从中期扩展为趋势性；反之，若 FTNT 下一份财报显示服务营收追平共识、毛利率区间重新上移，则本次应被确认为仓位出清，"错杀"逻辑成立。 分类: 重大公司事件 原因: 飞塔信息发布季度财报并调整全年指引，属于重大公司事件，直接影响投资者预期。 重要程度: 重要 温馨提示： 请将我们的邮箱地址 "mail.lxaa.top"、"lxkj.fintechcreater.top" 添加到您的邮箱白名单，以确保您能正常接收我们的邮件。 此邮件由系统自动发送，请勿回复。 如需调整推送设置，请 登录网站 进行修改。 如果您不想再收到此类邮件，请 点击这里退订 。

Fortinet公布第三季度营收20.8亿美元、Billings 23.4亿美元和调整后每股收益0.91美元，三项均高于分析师预期；公司同时上调全年营收及全年Billings指引，但全年服务营收指引低于分析师预期，全年调整后毛利率指引区间较此前收窄。