Kimi K3 may become the “DeepSeek 2.0 moment” for China's large-model market, but Nomura believes AI infrastructure demand will not be weakened.

Overview

This report focuses on Kimi K3, released by Moonshot AI on July 16, 2026, discussing whether it constitutes a new “DeepSeek moment” for Chinese large models and its implications for the global AI supply chain, model competition landscape, cloud platforms, and software application companies. Kimi K3 is a 2.8T-parameter open-source large model, with full model weights scheduled to be released on July 27, 2026. It features native multimodality, a 1M token context window, persistent reasoning, and long-horizon agentic task capabilities.

Core views

Nomura's core view is that Kimi K3 increases the presence of Chinese large models in competition among global frontier models, but should not be interpreted as signaling a peak in or weakening of AI compute demand. On the contrary, competition in frontier models, the continuation of Scaling Laws, rising inference workloads, and the spread of sovereign AI demand will continue to support the AI infrastructure value chain. Leading large-model vendors in both China and the U.S. stand to benefit, provided they maintain leadership on the technology curve.

Analysis framework

The report analyzes Kimi K3 from five dimensions: model architecture, performance benchmarks, task cost, real-world use cases, and segmented supply-chain impact. The supply-chain section discusses compute, networking, IDC and AI cloud, and software and applications in layers, and maps them to relevant listed companies' ratings and industry opportunities.

Methodology notes

• Model performance and cost comparisonArtificial Analysis Intelligence Index and Cost per AA task

  Use an independent model capability index and cost per task to measure Kimi K3's position relative to Claude Fable 5, Opus 4.8, GPT-5.5, and GLM-5.2.

  Kimi K3 is in the global frontier tier on the intelligence index. Its cost is lower than some top closed-source models but higher than some Chinese peers, indicating that it is positioned more as a highly cost-effective frontier model rather than simply a low-cost model.

• Supply-chain layered analysisAI supply chain segment mapping

  Break down the impact of large-model innovation separately on compute, networking, IDC/cloud, and software applications.

  The report argues that model progress will reinforce training and inference demand. Compute and networking are direct beneficiaries, AI cloud and IDC benefit from model hosting and capital expenditure, while software applications face disruption but vertical leaders may prevail.

• Investment rating systemNomura Group equity research rating system

  Buy indicates expected outperformance versus the benchmark over the next 12 months, Neutral indicates performance broadly in line with the benchmark, and Reduce indicates underperformance.

  This is not a single-company report, but a thematic research report that reiterates rating preferences across multiple related supply chain and software names.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Moonshot AI / Kimi K3

  Core thematic asset, unlisted

  Strengths

  2.8T open-source model, 1M token context window, native multimodality, persistent reasoning, strong long-horizon agentic task capability, and lower cost than several overseas frontier closed-source models.

  Weaknesses

  Still lags Claude Fable 5 and GPT-5.6 Sol in performance, and costs more than some Chinese peer models.

  Comparison

  Positioned in the global frontier model tier, with cost roughly close to GPT-5.6 Sol, lower than Fable 5 and Opus 4.8, but higher than GLM-5.2 and DeepSeek V4 Pro.

  Risks

  Intense frontier model competition, advanced chip supply constraints, geopolitics, and uncertainty around open-source commercialization.

• TSMC (2330 TT)、ASE (3711 TT)、ASPEED (5274 TT)、MediaTek (2454 TT)、GWC (6488 TT)、KYEC (2449 TT)、EMC (2383 TT)、TUC (6274 TT)、ZDT (4958 TT)、Samsung Electronics (005930 KS)

  Beneficiary names in AI compute, semiconductors, memory, and the server supply chain

  Strengths

  The report believes the expansion of training and inference will support demand for advanced computing, packaging and testing, CPU/TPU, SiC, AI chip testing, CCL, PCB/HDI, and memory.

  Weaknesses

  Some segments depend on hyperscale cloud vendors' capital expenditures and the sustainability of the AI cycle.

  Comparison

  Compared with software applications, demand transmission from frontier model competition is more direct for hardware infrastructure.

  Risks

  Insufficient cloud vendor free cash flow, cyclical capex fluctuations, supply bottlenecks, and valuation pullbacks.

• Zhongji InnoLight (300308 CH)、Suzhou TFC (300394 CH)、Yuanjie Tech (688498 CH)

  Beneficiary names in AI networking and optical communications

  Strengths

  The large-scale AI factory and SuperNode trend is driving demand for high-speed interconnects, optical modules, optical devices, and optical chips.

  Weaknesses

  Yuanjie Tech is rated Neutral, indicating that its valuation or fundamental leverage may be less attractive than optical module leaders.

  Comparison

  Chinese AI chip performance still lags global top-tier chips, so SuperNode and networking solutions are becoming important levers to narrow the gap.

  Risks

  AI cluster build-out timing, price competition, customer concentration, and technology iteration risk.

• Alibaba (BABA US)、GDS (GDS US)、VNET (VNET US)

  Beneficiary names in AI cloud platforms and IDC

  Strengths

  AI cloud platforms can benefit from MaaS demand, the pricing power from hosting open-source models, and the ecosystem of cloud infrastructure/model/application services; IDC benefits from capex driven by training and inference demand.

  Weaknesses

  Requires continued infrastructure investment, which may pressure short-term profitability and cash flow.

  Comparison

  Compared with standalone model vendors, cloud platforms can simultaneously serve multi-model hosting, private deployment, and compliance needs.

  Risks

  Cloud capex volatility, model platform competition, and changes in regulatory and compliance requirements.

• Kingdee (268 HK)、Kingsoft Office (688111 CH)

  Beneficiary names with vertical moats in the software and applications layer

  Strengths

  Vertical leaders can use Gen AI to enhance product capabilities, customer stickiness, and the depth of industry solutions.

  Weaknesses

  General software and application companies face competitive pressure as LLMs lower development barriers.

  Comparison

  The report is more positive on software companies that can embed Gen AI into industry scenarios and strengthen their moats, rather than undifferentiated general software vendors.

  Risks

  Competition from AI-native startups, uncertainty in commercialization pace, customer willingness to pay, and product substitution risk.

Key data

• Kimi K3 parameter scale2.8TThe report describes it as currently the largest open-source LLM.
• Context window1M tokenSupports long-horizon knowledge work, coding, and agentic tasks.
• Artificial Analysis Intelligence Index57, about 3rd-4th/189Behind Claude Fable 5 and GPT-5.6 Sol, but broadly in the same tier as Claude Opus 4.8 and GPT-5.5.
• CostAbout $0.94/AA task; input $3 per million tokens, cache hit $0.30 per million tokens, output $15 per million tokensLower than Claude Fable 5 at about $2.75 and Opus 4.8 at about $1.80, but higher than GLM-5.2 at $0.32-$0.47.
• Architecture featuresKimi Delta Attention, Attention Residuals, Stable LatentMoE, MoE activating 16/896 expertsThe report says overall scaling efficiency improved by about 2.5x versus Kimi K2.
• Long-horizon task examples48-hour unattended chip design, an ASIC industry research site with 120+ rounds of self-improvement, gravitational wave analysis of Event 391, 3D open-world game generationUsed to illustrate that K3 is aimed at long-horizon agentic work, not just chat.
• Global developer token share of Chinese modelsMore than 45%According to OpenRouter statistics; less than 2% a year ago.

Impact & implications

The launch of Kimi K3 may briefly trigger market concerns over AI supply-chain valuations and compute demand, but the report believes its longer-term implications are more positive: competition among frontier models will push AI labs and hyperscale cloud platforms to continue investing, and demand for advanced GPUs/ASICs, memory, AI networking, IDC, and cloud platforms is likely to continue; on the software side, general software vendors face lower barriers to entry and LLM-driven disruption, but companies with industry-specific scenarios, customer bases, and vertical moats are more likely to win in the Gen AI cycle.

Risks

• Intensifying competition among frontier large models may compress pricing and profit margins for model vendors.
• Advanced chip supply constraints and geopolitical risks may affect the expansion of Chinese AI labs.
• The market may once again interpret Chinese breakthroughs in efficient models as a decline in AI compute demand, causing short-term valuation volatility in the AI supply chain.
• Ongoing capital expenditure by cloud vendors may be constrained by free cash flow, the macro environment, and the pace of demand realization.
• LLMs lower software development barriers, which may weaken the existing competitive advantages of general software and application companies.
• Sovereign AI and decoupling trends may increase demand for localized deployment, but may also lead to market fragmentation and rising compliance costs.

What to watch

• Developer adoption after Kimi K3 releases its full model weights on July 27, 2026.
• Kimi K3's actual commercialization performance in long-horizon coding, knowledge work, visual reasoning, and agentic tasks.
• Whether the global token share of Chinese models on platforms such as OpenRouter continues to rise.
• Subsequent technology iterations and pricing strategies of global leading models such as Claude and OpenAI.
• AI capital expenditure, data center construction, and free cash flow changes among hyperscale cloud vendors around 2027.
• Whether AI networking, optical modules, memory, and ASIC/GPU supply chains continue to see tight supply-demand conditions or rising prices.
• Evidence that leading vertical software companies are converting Gen AI into revenue, retention, and margin improvement.