Kimi K3 is viewed as a new key milestone for China's AI models, but Nomura does not see it as a signal that computing demand has peaked
AI summary card
Kimi K3 is viewed as a new key milestone for China's AI models, but Nomura does not see it as a signal that computing demand has peaked
Nomura believes Moonshot AI's release of the 2.8T-parameter open-source Kimi K3 strengthens the global competitiveness of Chinese large models, while the large-model race and expansion of inference workloads should continue to benefit the AI infrastructure supply chain.
- Kimi K3 is an open-source large model with approximately 2.8T parameters, featuring native vision, multimodality, a 1 million-token context window, and persistent inference capabilities.
- The report says Kimi K3 ranks third to fourth on the Artificial Analysis Intelligence Index, with an estimated task cost of approximately USD0.94, lower than some leading overseas models but higher than most domestic peers.
- Nomura believes Kimi K3 could trigger concerns similar to those following the release of DeepSeek R1, but competition among large models will not weaken computing demand; instead, it will drive expansion in training, post-training, and inference workloads.
- Beneficiaries include advanced computing, AI chip testing and packaging, CCL/PCB, storage, optical modules, AI cloud platforms, and IDC.
- Uncertainty remains on the software and applications side. General-purpose software companies may face pressure, while companies with vertical-industry moats that can leverage generative AI are more likely to prevail.
Report interpretation
Overview
This report focuses on Kimi K3, released by Moonshot AI ahead of WAIC 2026, and assesses its impact on the global large-model competitive landscape, the internationalization of Chinese AI models, and AI supply chain demand. Kimi K3 is described as one of the largest open-source LLMs currently available. It adopts the Kimi Delta Attention, Attention Residuals, and an expanded MoE architecture, delivering an approximately 2.5-fold improvement in overall scaling efficiency compared with Kimi K2. The report believes Kimi K3 could remind the market of the shock to computing demand following the release of DeepSeek R1, but Nomura's core view is that competition and innovation in large models will not stop, and that infrastructure areas including computing, networking, AI cloud, and IDC still offer structural opportunities.
Core views
The core views are as follows: First, Kimi K3 marks the entry of Chinese large-model developers into a higher-end tier of global competition, targeting high-end, economical, and low-price model segments. Second, the share of Chinese models in global developer token volumes has risen significantly; the report cites OpenRouter statistics showing that Chinese models now account for more than 45%, compared with less than 2% a year ago. Third, Kimi K3's success does not imply declining training demand; rather, it reinforces the view that scaling laws for pre-training and post-training remain valid. Fourth, long-horizon coding, knowledge work, multimodal creation, and agentic tasks will drive heavier token consumption and growth in inference workloads. Fifth, software and application companies face disruption from LLMs, but vertical leaders with industry data, use cases, and customer barriers may prevail over the long term.
Analysis framework
The report uses an event-driven and industry-chain mapping approach: it first reviews Kimi K3's architecture, capabilities, pricing, and use cases, then decomposes its impact across four areas—computing, networking, IDC/cloud, and software applications—and maps these effects to ratings and investment preferences for relevant listed companies. The analysis also compares the relative capabilities and costs of leading overseas models, domestic models, and Kimi K3.
Methodology notes
Combines model intelligence scores, per-task costs, context windows, inference capabilities, and multimodal capabilities to assess Kimi K3's position within the global LLM pricing spectrum.
Kimi K3's task cost is approximately USD0.94, lower than Fable 5 and Claude Opus 4.8, and close to GPT-5.6 Sol, but higher than GLM-5.2 and DeepSeek V4 Pro, indicating that its positioning is shifting from low cost toward high cost-performance.
Assesses separately the impact of large-model advances on computing, networking, IDC/cloud, and software applications.
The report believes the computing and networking segments benefit from the expansion of training and inference, while AI cloud and IDC benefit from model hosting and computing demand. The software applications segment faces competitive pressure from lower entry barriers created by LLMs.
Uses pre-training, post-training, and inference workloads to assess whether demand for advanced computing will continue.
The report believes leading LLMs have not yet reached their performance limits, and that additional computing resources can still translate into stronger model capabilities, so demand for AI infrastructure remains robust.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- TSMC (2330 TT)Beneficiary of AI chip manufacturing and advanced computing
- Strengths
- Its advanced process technology and AI chip foundry capabilities place it at the center of AI computing expansion.
- Weaknesses
- It remains exposed to the global semiconductor cycle, customer capital expenditure, and geopolitical risks.
- Comparison
- The report classifies it as an AI chip enabler and reiterates Buy.
- Risks
- Orders and valuation could come under pressure if AI capital expenditure slows or the supply chain becomes constrained.
- ASE (3711 TT)Beneficiary of advanced packaging and the AI chip supply chain
- Strengths
- Benefits from demand for advanced packaging such as WoS and CoW.
- Weaknesses
- Cyclicality and fluctuations in packaging capacity utilization remain constraints.
- Comparison
- The report includes it in the Buy list for the computing segment.
- Risks
- AI chip demand or the pace of advanced packaging capacity expansion may fall short of expectations.
- Samsung Electronics (005930 KS)Beneficiary of AI-driven memory shortages
- Strengths
- The global memory industry is experiencing severe shortages due to AI demand, and the report identifies it as a preferred name among Korean technology analysts.
- Weaknesses
- The memory business remains subject to price-cycle fluctuations.
- Comparison
- It is a preferred Buy-rated stock in the report.
- Risks
- If AI server demand cools, memory prices and earnings recovery may fall short of expectations.
- Zhongji InnoLight (300308 CH)Beneficiary of AI networking and optical modules
- Strengths
- The increasing demand for large-scale AI factories and SuperNodes enhances the value of optical modules and interconnects.
- Weaknesses
- It is exposed to customer concentration, product iteration, and price competition.
- Comparison
- The report identifies it as a preferred name in global AI networking and assigns it a Buy rating.
- Risks
- The pace of AI cluster construction, overseas demand, or changes in technical specifications may fall short of expectations.
- Suzhou TFC (300394 CH)Beneficiary of optical communications components for AI networking
- Strengths
- Benefits from growth in AI cluster interconnect and high-speed optical communications demand.
- Weaknesses
- Industry competition and customer certification cycles may affect the realization of growth.
- Comparison
- The report lists it alongside Zhongji InnoLight as a Buy-rated networking name.
- Risks
- Capital expenditure may slow or optical communications prices may decline.
- Alibaba (BABA US)Beneficiary of the Chinese AI cloud ecosystem
- Strengths
- Its AI cloud platform can strengthen its ecosystem and pricing power through MaaS, open-source model hosting, private deployments, and compliance services.
- Weaknesses
- Cloud competition and macro internet demand may still constrain growth.
- Comparison
- The report is positive on Alibaba's position in China's AI Cloud ecosystem and assigns it a Buy rating.
- Risks
- AI cloud monetization may fall short, model-service price competition may intensify, or regulatory uncertainty may increase.
- GDS (GDS US)Beneficiary of IDC infrastructure
- Strengths
- Benefits from training and inference infrastructure demand from major AI cloud companies and frontier-model companies.
- Weaknesses
- The IDC business requires high capital expenditure and has a long return cycle.
- Comparison
- The report identifies it as a Buy-rated IDC name.
- Risks
- Customer expansion may fall short, power resources may be constrained, or financing costs may rise.
- VNET (VNET US)Beneficiary of IDC infrastructure
- Strengths
- It is positioned to benefit from data-center demand generated by Chinese AI training and inference workloads.
- Weaknesses
- Industry competition, the balance sheet, and utilization fluctuations may affect profitability.
- Comparison
- The report identifies it as a Buy-rated IDC name.
- Risks
- AI customer demand may materialize more slowly than expected, or IDC supply may become excessive.
- Kingdee (268 HK)Beneficiary of vertical software and generative AI applications
- Strengths
- It has an enterprise software and industry-use-case foundation, and effective integration of generative AI could strengthen its moat.
- Weaknesses
- Lower development barriers created by LLMs may intensify competition at the application layer.
- Comparison
- The report identifies it as a Buy-rated name in the Chinese software sector.
- Risks
- AI feature monetization may be insufficient, or customer IT spending may weaken.
- Kingsoft Office (688111 CH)Beneficiary of office software and generative AI applications
- Strengths
- Office scenarios are well suited to AI assistants and knowledge-workflow upgrades.
- Weaknesses
- General-purpose software may be challenged by LLM-native applications.
- Comparison
- The report identifies it as a Buy-rated name in the Chinese software sector.
- Risks
- AI subscription conversion may fall short of expectations, or competition may intensify.
Key data
- Kimi K3 parameter scale2.8TThe report describes it as one of the largest open-source LLMs currently available.
- Context window1M tokensKimi K3 has a 1 million-token context window.
- Improvement in scaling efficiencyApproximately 2.5xImprovement in overall scaling efficiency relative to Kimi K2.
- Artificial Analysis Intelligence Index57The report says Kimi K3 ranks third to fourth among 189 models.
- Kimi K3 pricingUSD3/mn token input；USD15/mn token outputThe input price for cache hits is USD0.30/mn token.
- Kimi K3 per-task costApproximately USD0.94Lower than approximately USD2.75 for Fable 5 and approximately USD1.80 for Opus 4.8, and close to approximately USD1.04 for GPT-5.6 Sol.
- Share of global developer token volumes from Chinese modelsMore than 45%The report cites OpenRouter statistics and says the figure was below 2% a year ago.
- Unsupervised chip-design use case48 hoursKimi K3 completed an example of an unattended chip-design workflow based on open-source EDA.
Impact & implications
From an investment perspective, the report views Kimi K3 as a catalyst for the upgrading and globalization of Chinese AI model capabilities, rather than evidence of a decline in AI infrastructure demand. In the short term, the market may worry about AI capital expenditure after Chinese models achieve breakthroughs under restricted chip conditions, but Nomura believes stronger models will expand the application scenarios for generative AI among consumers and enterprises, sustaining demand for training, post-training, inference, networking, AI cloud, and IDC. The impact on software is more differentiated: LLMs will lower barriers to application development and weaken the competitive advantages of some general-purpose software companies, but vertical leaders that use generative AI to strengthen their products and ecosystems may benefit over the long term.
Risks
- Breakthroughs by Chinese models such as Kimi K3 could trigger short-term market concerns about demand for advanced computing.
- Geopolitical risks and technological decoupling could affect AI chips, model services, and cross-border cloud ecosystems.
- If AI infrastructure capital expenditure slows, computing, storage, networking, and IDC-related names could come under pressure.
- Lower software development barriers created by LLMs could weaken the moats of some general-purpose software companies.
- Capability, pricing, and cost data for the models come from third parties or company disclosures and are subject to subsequent changes and verification uncertainty.
What to watch
- Developer adoption and ecosystem diffusion following the release of Kimi K3's complete model weights.
- Changes in the token share of Chinese models on global developer platforms such as OpenRouter.
- Technical and pricing responses to Kimi K3 from leading overseas models.
- MaaS revenue, inference token consumption, and private-deployment demand on AI cloud platforms.
- Whether capital expenditure on AI servers, advanced packaging, HBM/memory, optical modules, and IDC remains strong.
- The ability of vertical software companies to convert generative AI into paid products and customer retention.