Gf Securities has released a research report indicating that the core driver of the current AI industry chain comes from the enhanced agentic coding capabilities of models. Coding tools represented by Claude Code and Codex are transitioning from auxiliary completion to autonomous collaboration, thereby transmitting multi-layered investment opportunities upstream and downstream.
Looking upstream, the consumption of computing power by Agentic Workloads has significantly increased, driving high growth in cloud vendors' IaaS and MaaS businesses. This reduces market concerns about over-investment by cloud vendors leading to an AI bubble, and sensitivity to declines in Free Cash Flow (FCF) Margins is consequently weakened. Attention should be paid to the easing of CSP compute constraints and the pace of order growth.
For model companies themselves, Agentic tasks are driving a surge in Token consumption and accelerating the annual recurring revenue (ARR) growth of leading models. The short-term trading focus lies on marginal changes such as model updates, ARR data, Token prices, and compute supply.
Looking downstream to applications, Coding Agents have been the first to validate product-market fit (PMF). Attention should be paid to the "Claude Code for X" style of scenario replication and the expansion space brought about by improvements in Token economics.
Key Points from Gf Securities:
AI Model Horizontal Comparison: Domestic Models Show Significant Capability Progress.
Recently, multiple domestic and international AI vendors have launched their new-generation flagship models. Following Anthropic's release of Claude Fable 5 and OpenAI's launch of GPT-5.6, Moonshot AI released the Kimi K3 model on July 16. Gf Securities believes that leading flagship models have all crossed the usability threshold for research scenarios. Based on their testing, the functional implementation of the four models fully met the standards, and the gap between models has shifted from "whether it can be completed" to "quality and density of completion." Moreover, the task capability of domestic models has shown significant progress, further narrowing the gap with leading models.
AI Data Tracking: OpenRouter Large Model Token Call Volume Has Increased.
On a weekly basis, the token volume called via OpenRouter from July 6, 2026, to July 12, 2026, was 52.6T, an increase from 46.7T during the period of June 29, 2026, to July 5, 2026. Model call volume is expected to continue rising. Among the models called via OpenRouter, Tencent's newly launched Hy3 (free) on July 6 climbed to the top of the ranking chart driven by its free strategy, MiMo V2.5 maintained the second position, and DeepSeek V4 Flash ranked third.
AI Hot Topic Tracking: Moonshot AI Releases Kimi K3.
Moonshot AI released Kimi K3, announcing that the model is simultaneously available on Kimi, Kimi Work, Kimi Code, and Kimi API. Kimi K3 is a 2.8-trillion-parameter MoE model built on the Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) architecture, natively supporting visual understanding and a 1-million-token context window. The company stated that Kimi K3 is the first open-source model in the 3-trillion-parameter class (2.8 trillion parameters). The complete model weights will be open-sourced by July 27, along with the release of the technical report.
Risks: Model iteration results may fall short of expectations; commercialization and application implementation may fail to meet expectations; risks related to copyright, ethics, and content quality; risks of incomplete evaluation due to sample size limitations and access permission restrictions.