Founder: Control over large model platforms and valuation restructuring, with key focus on the real TCO of domestic accelerators, etc.

Founder released a research report stating that the large model industry is moving from "stronger models" to "cheaper, more reliable, and more callable intelligence supply." Capability remains the foundation, but the determinants of industry value are extending from the models themselves to task economics, workflow control rights, proprietary feedback loops, and capital efficiency. Foundation models still have platform value, but not all model companies can become platforms; in the Chinese market, key focus should be placed on the real TCO of domestic accelerators, overseas Token usage structure, the pace at which APIs replace private deployment, and the feasibility of Coding/Agent expanding into platform capabilities. For tech giants, key observation areas include compute utilization corresponding to AI capital expenditure, revenue growth, and free cash flow returns. Founder's main views are as follows: From the perspective of technological evolution, competition in large models is no longer merely about scaling parameter size, but rather systematic iteration around capability, efficiency, and validation in real environments. Transformer remains the current backbone architecture, while technologies such as MoE, sparsification, long context, post-training, inference-time compute, and tools and retrieval continue to improve models' computational efficiency and task capabilities, with multimodality further evolving toward omni-modality. At the same time, the production of model capabilities is gradually shifting from "compute data algorithms" to "compute data algorithms engineering closed loop," and validation, feedback, and iteration in real tasks have become important sources for continuous capability improvement. Capability improvements and cost declines are further changing the commercialization of large models, and industry value is beginning to migrate from model output itself to task results and workflows. Tokens remain the basic unit of measurement for intelligent service supply, but the key to measuring model economics is gradually shifting to cost per successful task and the corresponding customer value. As commercialization evolves from APIs to MaaS, Agents, and task-result billing, the integration of models with enterprise workflows, cloud platforms, and terminal entry points continues to deepen, and long-term value is gradually shifting from simply providing model capabilities to obtaining stable production workloads, workflow control rights, and default invocation rights. On this basis, the large model industries in China and the United States have formed different competitive paths, and a company's long-term value depends on whether it can convert its own resource endowments into sustained capability advantages and commercial advantages. Major U.S. companies rely on compute, cloud, developer ecosystems, and global entry points to form differentiated layouts in frontier models, enterprise workflows, infrastructure, and open-source ecosystems; Chinese vendors, under supply constraints, place greater emphasis on cost efficiency, industrial scenarios, localized services, and adaptation to domestic compute. Therefore, judging the competitiveness of a model company cannot rely only on model capabilities, but also requires a comprehensive assessment combining customers, ecosystem, entry points, and capital investment capacity. Therefore, the valuation logic for large model companies also needs to shift further from "technological leadership" to "value capture," with the core being to judge whether capability advantages can be continuously converted into revenue, workflow control rights, and capital returns. Based on this, this report establishes a seven-dimensional valuation framework covering capability sustainability, unit task economics, revenue quality, workflow control rights, data closed loops, ecosystem entry points, and capital efficiency, and further establishes a dynamic tracking system that incorporates model capabilities, production workloads, default invocation rights, and capital efficiency into continuous observation. Risk disclosure: risks from changes in the macroeconomic environment and regulatory policies; risks that AI industry development and commercialization fall short of expectations; risks that competition among relevant companies and returns on capital investment fall short of expectations.