DeepSeek V4系列发布：能力提升且效率更高；对云和数据中心板块利好

withhigherefficiency;+VEon cloud and DCPrice Objective Change DeepSeekV4seriesunveiled:V4-FlashandV4-Pro China (Fri), its first major model update since launch of R1 in Jan 2025. There are two MoEameters (13Bactivated)andDeepSeek- V4-Pro with1.6T parameters (49B activated), both extending the context length to 1Mtokensfrom128K of V3.2.DeepSeek-V4-Prlax(themaximumreasoning effortmodeof V4-Pro)achievestop-tielcoding benchmarksand significantlynarrows thegap with leading closed-snodelson reasoning and agentictasks(Exhibit 3). For API pricing, V4-Pro charges only 6-14% that of frontier models offeredby OpenAl, Google and Anthropic for input, and merely 2% for output (Exhibit 5). APIcost of DS V4-Pro could see significant reduction in 2H26 with the mass-shipment ofHuawei Ascend950SuperNode,according to DeepSeek.Positive read-through for public cloud & DC sectors Daley Li, CFA >Research AnalystMerrill Lynch (Hong Kong) daley.li2@bofa.comHarryZhuang>>Research AnalystMerill Lynch (Hong Kong)+85235087998 harry.zhuang@bofa.comOwenWang>Research AnalystMerill Lynch (Hong Kong) +85235086533owen.wang2@bofa.com continuous model iterations by domestic LLM developers (e.g., DeepSeek, Moonshot Al, Zhipu) and internet companies (Alibaba, ByteDance and Tencent) could further drivecomputing power demand for model training, 2) cheaper APl pricing of open-sourcemodels could boost wider Al adoption among customers, leading to rising Al inferencingdemand, and 3) rising performance of domestic Al chips could gradually ease the supplytightness of Al accelerators. Looking at the public cloud and data center segments, webelieve public cloud providers could have greater upsides to revenue growth and takeshorter time to see fundamental improvements following the surging Al demand.KeyinnovationsofV4inarchitectureandoptimization CSA: Compressed Sparse Attention HCA: Heavily Compressed Attention LLM: large language model mHC: manifold-constrained Compared to DeepSeek-V3,V4 series introduces several key innovations in architecture hyper-connectionsMoE: mixture of experts and optimization, improving long-context efficiency, modeling capability, and trainingstability.It debuts a hybrid attention architecture combining Compressed SparseAttention (CSA) and Heavily Compressed Attention (HCA), cutting single-token inferenceFLOPs to27%andKVcacheto10%forDeepSeek-V4-Proversus V3.2 in1M-tokenlong-context setting.It also incorporatesmanifold-constrained hyper-connections(mHC) that upgrade conventional residual connections for strengthening modelingcapability,as well as Muon optimizerforfaster and more stable training. According to Huawei, entire Ascend SuperNode product series are capable of supporting DeepSeek V4 series. Through fused kernels and multi-stream parallelism, Ascend 950 Alsignificantly improve inference performance,while multiple quantization algorithmsenable high-throughput and low-latency deployment of V4 models. >> Employed by a non-US affiliate of BofAS and is not registered/qualified as a research analystunder the FINRA rulesRefer to'Other Important Disclosures'forinformation on certain BofA Securities entities that takeresponsibility for the information herein in particular jurisdictions.BofA Securities does and seeks to do business with issuers covered in its researchreports.Asaresult,investorsshouldbeawarethatthefirmmayhaveaconflictofinterest that could affect the objectivity of this report. Investors should consider thisreportasonlyasinglefactorinmakingtheirinvestmentdecision.Objective Basis/Risk on page 4.12964065 Benchmark (Metric)# Shots modelsonreasoningandagentictasksComparison of DeepSeek-V4-Pro-max against frontier models Benchmark (Metric)DS-V4-ProMaxOpus-4.6Max OverviewofmajorLLMdevelopment CompanyModel Stocksmentioned Priceobjectivebasis&risk GDSHoldings (GDS/XMJFF) Our PO of US$62.0 / HKs61.0 is based on a blended average of 17x 12-month-forwardEV/EBITDA for China business (100% stake) and the international business, which is setat 20% discount to overseas peers given lower profitability expected overseas expansion, and (3) new government policies supporting the datacenter industry. Downside risks: (1) lower prices for data center services on higher competition, (2) overseas restrictions.KingsoftCloud (KCLHF/KC) We derive our PO of HKs10.8 / USs20.8, based on 3x 12mths forward P/S, which is set as the average valuation of China's independent cloud service providers listed in A-sharemarket and then applying a 50% A/H-share discount. The 50% A/H-share discount is setat the average share price discount for all companies dual-listed in both A-share and H-share market. We derive our PO on revenue multiple, because (1) Kingsoft Cloud is stillloss making in near-term, and (2) we believe P/S multiple can better reflect its growthoutlook, which investors care more about. business,2) substantial demand increase from Kingsoft and Xiaomi ecosystem, and (3)significant demand recovery from the inte