全球内存：KVTC与Turbo Quanta是内存市场的风险吗？

Are TVTC and Turbo Quanta risks for memory?Quick Note Asia PacificTechnology CW Chung- NIHK cwchung@nomura.com+85222526075Korea Technology We have believed that that memory is becoming thekey beneficiary in scaling Al Eon Hwang-NFIK infrastructure, particularly driven by the exponential growth in KV cache demand.However,at the same time,newtechnologies suchas Nvidia's (NvDA US, Not rated)KVTC and Google's (GOOGL US, Not rated) Turbo/Quanta aim to improve memoryefficiency in orderto improve thekeybottleneck for scaling.The market has someconcerns-aswasthecaseforDeepseek(unlisted)-aboutwhetherthesenewtechnologies can be harmfulforthe memory market.We believe these solutions lowereffectivecostsandimprovesystemthroughput,ultimatelyenablingbroaderAldeployment. As seen in prior efficiency cases, improvements in resource utilization tend toexpand total workloads, suggesting that memory demand will continue togrow alongsideAl adoption.The Jevons paradox still works at the current early stage of Al, in our view.Structural imbalance between exponential demand and linear supply eon.hwang@nomura.com+822 3783 2318 In our view,thefundamental issue lies in a structural mismatch:KV cache demand is scaling exponentially,while memory supply has grown only linearly (30-60% per year)This divergence is driven by several reinforcing trends, including longer user sessions withmulti-turn interactions,rapid expansion in context length (from 4K to 128K and beyond),increasing adoption of RAG architectures, and the shift toward multimodal inputs such asimages and video.These factors collectively have driven a step-change in token intensity,with video data potentially demanding ~3.3mn times largertoken size than text.Alagentsalso use ~1k times larger KV caches and tokens than text, in our estimate. As a result, KVcache demand is increasing at multifold rates annually, far outpacing supply growth (Fig.2 ). Optimizationisastructural requirementforsustainingAlscalingWe note that memory optimization is no longer optional but essential for system viability, in our view. The industry is converging on three key approaches: compression,quantization,andaccessoptimization.Compressiontechniquescanreducedatavolumeby up to ~10-20x, while quantization delivers an additional ~4-8x reduction, and accessoptimization improves effective bandwidth utilization and latency.While these combinedapproaches materially improve efficiency, they should be viewed as enabling mechanismsthat allow Al workloads to scale further, rather than as factors that suppress underlyingmemory demand. Efficiencygainsreinforcetotaldemand,withimplicationsacrossmemorysegmentsgrowth. Instead, we believe reductions in per-unit memory requirements are likely to bemore than offset by a rapid expansion in total workloads, as lower costs enablelongercontext windows,more complex models,and broaderdeployment. (continued on the next page) Production Complete: 2026-03-26 06:40 UTC Finally, we believe that the competitive landscape is increasingly defined by memory efficiency rather than hardware capacity only. GPU vendors are integrating system-levelsolutions to mitigate memory bottlenecks, while alternative architectures retainadvantages in optimized environments.As Al infrastructure continues to scale,the abilitycompetitiveness, in our view. In this context, we believe memory efficiency is not only aconstraint but also a critical driver of long-term industry evolution. Appendix A-1 See Disclaimers for Nomura Group entity details. Analyst Certification any or all of the subject securities or issuers referred to in this Research report, (2) no part of our compensation was, is or will be directly orindirectly related to the specific recommendations or views expressed in this Research report and (3) no part of our compensation is tied to anyspecific investment banking transactions performed by Nomura Securities International, Inc., Nomura International plc or any other NomuraGroup company. Important Disclosures Online availabilityof research and conflict-of-interest disclosures Nomura Group research is available on www.nomuranow.com/research, Bloomberg, Capital IQ, Factset, LSEGImportant disclosures may be read at http://ao.nomuranow.com/research/m/Disclosures or requested from Nomura Securities International, Inc.If you have any difficulties with the website, please email grpsupport@nomura.com for help. The analysts responsible for preparing this report have received compensation based upon various factors including the firm's total revenues, anot registered/qualified as research analysts under FINRA rules, may not be associated persons of NSl, and may not be subject to FINRA Rule2241restrictions on communications with covered companies,public appearances,and trading securities heldby a research analyst account. Nomura Global Financial Products Inc. (NGFP) Nomura Derivative Products Inc. (NDP) and Nomura International plc. (Niplc) are registered withthe Commodities Fut