Moonshot AI released Kimi K3, and Nomura believes it demonstrates China's AI model R&D capabilities and may reshape the industry's competitive landscape

Overview

This report is an AI industry update from Nomura Securities, focusing on Kimi K3 released by Moonshot AI. The report believes the significance of Kimi K3 lies not only in its coding ability approaching that of models from leading U.S. AI labs, but also in its hybrid architecture combining Transformer and recurrent models, which demonstrates a new technical path for long-context, long-horizon reasoning, and agentic coding.

Core views

The core view is that Kimi K3 provides a new technical path for AI model development, which is favorable for long-term growth in the AI industry and strengthens evidence of China's capabilities in frontier AI model R&D. However, intensified short-term competition may suppress or delay the commercialization of U.S. closed-source models, thereby affecting capital flows among AI labs, cloud operators, and electronic equipment manufacturers such as memory and storage suppliers.

Analysis framework

The report mainly analyzes the impact of Kimi K3 from five angles: model architecture, inference efficiency, long-context capability, industry competitive landscape, and capital flows in the supply chain. On the technical side, it focuses on explaining how KDA reduces dependence on the KV cache through fixed-length states and how MLA complements fine-grained token-level retrieval capability; on the industry side, it evaluates the potential impact on closed-source models, cloud vendors, and the hardware supply chain.

Methodology notes

• Model architecture analysisMoE and hybrid attention architecture analysis

  Kimi K3 combines MoE, MLA, and KDA to enhance long-context agentic coding capabilities.

  The report describes Kimi K3 as a 2.8 trillion-parameter MoE model that activates a subset of experts by token during inference; its hybrid architecture uses both Transformer's MLA and the recurrent model KDA, aiming to reduce the KV cache burden while retaining fine-grained retrieval capability.

• Supply chain impact analysisAI lab–cloud vendor–hardware supply chain capital flow analysis

  Model competition can affect the pace of closed-source model commercialization and downstream demand for computing power, storage, and electronic equipment.

  The report believes that if open-source or high-performance models intensify competition, the monetization of U.S. closed-source models may be delayed, creating a risk of temporary weakening in capital flows from AI labs to cloud operators and to hardware suppliers such as memory and storage vendors.

Asset mapping & comparison

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

• Moonshot AI

  Core subject of the report; the publisher of Kimi K3.

  Strengths

  Kimi K3's coding performance is described as comparable to that of models from leading U.S. AI labs, and it uses an innovative hybrid architecture.

  Weaknesses

  The report does not provide data on commercialization, revenue, profit, or valuation.

  Comparison

  The report says its frontier agentic coding model challenges the dominance of U.S. closed-source models.

  Risks

  Competition from open-source models may create commercialization uncertainty, and the report does not cover the company's financials or capital market tradability.

• U.S. closed-source AI labs

  Potentially affected parties.

  Strengths

  They remain at the center of frontier model competition.

  Weaknesses

  High-performance models such as Kimi K3 may intensify competition and delay the monetization of closed-source models.

  Comparison

  The report believes Kimi K3 poses a challenge to the dominance of U.S. closed-source models.

  Risks

  Delayed commercialization, rising pricing pressure, and slower capital flows into the downstream hardware chain.

• Cloud operators

  Downstream recipients of spending by AI labs.

  Strengths

  Long-term growth in the AI industry may still bring demand for computing power.

  Weaknesses

  If AI lab commercialization comes under pressure in the short term, the pace of cloud service procurement or expansion may be affected.

  Comparison

  Compared with model companies, cloud operators are more dependent on the sustained release of AI capital expenditure and inference/training demand.

  Risks

  A temporary weakening of capital flows from AI labs to cloud operators.

• Memory, storage, and electronic equipment manufacturers

  Beneficiaries in the AI computing power and cloud infrastructure supply chain.

  Strengths

  Long-term AI model development still supports hardware demand.

  Weaknesses

  If upstream closed-source model commercialization is delayed, the realization pace of hardware demand may slow.

  Comparison

  The report's author covers Japanese electronic components, and therefore specifically highlights the impact of capital flows on the electronic equipment chain.

  Risks

  Slower cloud capital expenditure and downward revisions to expected demand for memory and storage.

Key data

• Release date2026-07-16The report states that Moonshot AI's Kimi K3 was released on this date.
• Model parameter scale2.8trn parametersThe report describes Kimi K3 as a 2.8 trillion-parameter MoE model.
• Number of experts896 expertsDuring inference, the router selects and activates 16 experts for each token.
• Experts activated per token16 experts per tokenReflects the sparse activation mechanism during MoE inference.
• Maximum context length1mn tokensThe report states that the model is aimed at long-horizon agentic coding, with a maximum context length of 1 million tokens.
• Nomura rating distribution date2026-06-30The disclosure appendix shows Nomura's global equity research rating distribution as of this date.
• Nomura Buy rating proportion58%Disclosure appendix information, not a rating on Moonshot AI or related stocks in this report.
• Nomura Neutral rating proportion39%Disclosure appendix information, not this report's investment recommendation.
• Nomura Reduce rating proportion3%Disclosure appendix information, not this report's investment recommendation.

Impact & implications

For investment research, the release of Kimi K3 has increased market attention on China's AI model capabilities and the competitiveness of open-source large models. In the long run, it may push AI model architectures to evolve further from pure Transformer models toward hybrid or post-Transformer directions; in the short run, it may intensify pricing and commercialization pressure on closed-source models and affect expectations for the cloud computing, memory, storage, and electronic components chain.

Risks

• The commercialization process of U.S. closed-source AI models may be delayed due to intensified competition.
• Capital flows from AI labs to cloud operators and then to electronic equipment manufacturers such as memory and storage suppliers may weaken in the short term.
• Although long-context and long-horizon reasoning capabilities are technically attractive, the report does not validate their commercialization effectiveness in real production scenarios.
• The report does not constitute an individual stock rating or target price for Moonshot AI or related listed companies.

What to watch

• Developer adoption of Kimi K3 after open-sourcing and the expansion of the agentic coding ecosystem.
• Whether the KDA and MLA hybrid architecture is adopted by other frontier models.
• Changes in pricing, API revenue, and commercialization pace at U.S. closed-source AI labs.
• Changes in cloud vendors' AI capital expenditure, GPU/storage/memory orders, and demand for electronic components.
• Public benchmark performance of Chinese AI models in long-context reasoning, code generation, and agentic tasks.