持有：提升年化经常性收入；Kimi K3的影响分析

InternetSoftware&Services +RaiseARRassumptions; cutTP onlowermargins inlongterm amid competition and higher share count 1,500.001,900.00 + Competition intensifies, but Z.ai can potentially catch up withhigherparametersizemodels +MaintainHold,cutTPtoHKD1,500(fromHKD1,900) Estimate, TP changes: Z.ai (Zhipu) ARR reached USD1bn in July (per BloombergNews on17 July2026),ahead of theguidance byDecember2026.We raiseARRestimates to USD2bn (from USD1bn) in December 2026 and USD6.5bn (fromUSD2.4bn) in December 2027. Though the revenue mix shift to lower-margin API willdilute margins, we expect Z.ai to turn profitable in 2027 instead of 2028. However, welowerour TP to HKD1,500 (from HKD1,900),factoring inlower FCFassumptions toreflect more intense competition in pricing, higher share count after the HKD31bnH-share follow-on offerings in July 2026, and changes in exchange rate (USD-HKD),OurTPimplies14xP/ARRinDecember2027vsAnthropic's8x(basedonHSBC'sglobal tech platforms team estimates on Anthropic's ARR and USD1.2trn valuation insecondary market transactions (Business Insider, 9 July 2026)).The premium isjustified by faster ARR growth of 220% for Z.ai vs Anthropic (50%). Maintain Hold. Implications from theKimi K3launch:(1)Scaling lawcontinues:We think K3'sintelligence leap is due to higher parameter size (rises from 1trn for K2.7 to 2.8trn)Z.aicouldpotentiallycatchupwiththenextmodelasGLM-5.2onlyhas744bnparameters.(2)Competition intensifies:K3's launch suggests that Chineseopenweight models could step up with 3trn parameter size models in 3Q26, includingMiniMax and DeepSeek.We notice the iteration frequencyfromfrontierlabs hasfurther accelerated in recent weeks, e.g., five leading US players (Anthropic, OpenAl,Google, SpaceXAl, Meta)all released model updates in July to date (exhibit 2).Thisshows that leadership in model capability has turned more temporary and frontierlabsneedtostayinvested.(3)ComputingpowermattersforARR:K3hasmuchslower token throughput than peers and paused new to-C subscription two days afterthe launch.This suggests it lacks computing power. On the other hand, Z.ai hasbegun operating a 1GW data center (several clusters, each with over 10kdomesticchips),withbreakthroughinusingdomestic chipsforinferencing(BloombergNews,21July2026)buffering computepower constraint. In addition,per Sina News on21 July 2026, Z.ai acquired XCore Sigma (a heterogeneous Al compute softwareprovider)to improve utilization of various types of chips.Webelieve these measurescansupportZ.aitogrowitsARRcontinuouslyamidrobustuserdemand Ritchie Sun*, CFAAnalyst, Internet ResearchThe Hongkong and Shanghai Banking Corporation Limitedritchie.k.h.sun@hsbc.com.hk+852 2822 4392 Charlene Liu*Head of Internet and Gaming Research, Asia PacificThe Hongkong and Shanghai Banking CorporationLimited, Singapore Branchcharlene.r.liu@hsbc.com.sg+6566580615 Peishan Wang* Analyst, Internet ResearchThe Hongkong and Shanghai Banking Corporation Limitedpeishan.wang@hsbc.com.hk+85239417008 Catalysts and risks: (1) Launch of a new GLM in 3Q26 (+ve), (2) model iterationsfrom competitors (-ve), (3) geopolitical risks related to potential ban of Chinesefrontier models in the US market (-ve), (4) lock-up expiry (-ve): c38% of Z.ai's sharescould be unlocked in early January 2027, and (5) fundraising/IPO of frontierlabscould reduce the scarcity premium. "Employed by a non-US afiliate of HSBC Securities (USA) Inc, and isnot registered/qualified pursuant to FINRA regulations No country for bears / The 24"edition ofthe EM Sentiment Surveya(A 03 xP11 Issuer of report: The Hongkong and ShanghaiBankingCorporation Limited Disclosures&Disclaimer This report must be read withthe disclosures and the analyst certifications inthe Disclosure appendix, and with the Disclaimer, which forms part of it. View HSBC Global Investment Research at:https://www.research.hsbc.com guidancebyDecember2026.WeraiseARRestimatestoUSD2bn (fromUSD1bn)inDecember2026 and USD6.5bn (fromUSD2.4bn) inDecember2027.Though the revenuemix shift tolower-margin API will dilute margins, we expect Z.ai to turn profitable in 2027 instead of 2028.OurTPcuttoHKD1,500 (fromHKD1,900)factors inlowerFCFassumptionstoreflectmoreintense competition in pricing, higher share count after the HKD31bn H-share follow-onofferings in July 2026, and changes in exchange rate (USD-HKD), Methodology: DCF Upside risks: + Resolving the computingbottleneck (i.e, limited access to high-performance Al compute)+GLMiterations+Faster-than-expected pace to break even+Endorsement from major industry playersDownside risks: HKD1,500.00 target price (from HKD1,900.00). This is based ona 10-year discounted cash flow (DCF) model, based on 11.8%WACC and 3% terminal growth rate. WACC is based on a 4.25%risk-free rate and 4.75% market risk premium (based on HSBC'sglobal equity strategy team views),a 0.5% regulatory riskpremium (due to being on the BIS Entity list), 1.45 beta, and0.3% debt-to-debt & equity ratio.We raise ARR est