Orient: AI inference drives rapid expansion of the Token market; overseas expansion of domestic model Tokens is expected to scale up.

Orient released a research report stating that as AI enters the inference era, Token factories can rely on computing power clusters and inference scheduling capabilities to achieve standardized Token supply, aligning with the continued growth in large model inference demand and possessing potential for large-scale cost reduction. The firm is optimistic about the long-term business development space of Token factories as core carriers for AI computing power commercialization. Orient's main points are as follows: The focus of the AI industry is shifting toward inference, and the Token market is expanding rapidly The AI industry is currently entering a new stage dominated by inference, with the industry focus shifting from model training to inference deployment. Tokens have transformed from a technical term into the core metering and pricing unit for AI services. Leading global model vendors have all established Token billing systems with differentiated pricing for input and output Tokens, while cloud vendors or specialized computing power service providers convert computing power resources into efficient Token services, rapidly forming a complete industry service chain. In terms of Token consumption, domestic Token call volumes have shown exponential growth, increasing nearly 5,000-fold over two and a half years. Token consumption by overseas leading model and application vendors is also expanding rapidly. With the advent of the Agentic AI era, the number of active Agents and total Token consumption will maintain high compound growth rates, providing a solid foundation for the commercialization of the Token factory business model. Token factories restructure computing power services; three types of players are positioning themselves in the AI inference industry NVIDIA has proposed transforming data centers into Token production factories, which essentially represents a paradigm restructuring of data centers from resource infrastructure to intelligent service units. Token factories connect the computing power-to-model service chain, relying on four key segmentscomputing power foundation, inference optimization, platform scheduling, and Token deliveryto package computing power into standardized large model APIs billed by Token, eliminating the need for downstream enterprises to build their own computing power and deploy models. Token factories differ fundamentally from computing power leasing: computing power leasing delivers hardware and bills by time duration with a relatively low technical threshold, while Token factories output inference services and settle by Token volume, forming higher barriers and better gross margins through full-chain inference optimization and scheduling technology. Track participants fall into three categoriescloud vendors, computing-asset-first players, and inference technology platformseach leveraging their respective resource endowments to seize industry growth opportunities. Domestic models, policy drivers, and infrastructure resonate; Token going overseas is expected to scale up Domestic large models continue to catch up with overseas leading models in performance, while Token pricing offers significant cost-effectiveness. The proportion of domestic models called on overseas aggregation platforms is rising rapidly, laying the foundation for Tokens going overseas. At the policy level, the state has introduced top-level rules for cross-border data, and multiple localities have implemented data processing pilot programs, forming a compliance framework where "raw data does not leave the dedicated zone, and only Tokens are exported." Large-capacity international submarine cables have been gradually commissioned, improving cross-border transmission bandwidth and consolidating the physical foundation for Tokens going overseas. Currently, Token going overseas is divided into two models: cross-border direct API connection, where overseas customers directly call domestic APIs subject to cross-border data regulation; and data processing, which relies on isolated bonded zones to complete inference with a more complete compliance loop. With multiple favorable factors, the Token overseas business is expected to usher in large-scale expansion. Various enterprises at home and abroad are actively pursuing Token factory business Representative overseas Token factory enterprises are Amazon AWS and Nebius. AWS Bedrock adopts a tiered Token billing model, leveraging inference optimization technology to improve GPU utilization, achieving stronger profitability compared to Microsoft's and Google's computing-duration-based billing models, with Token business significantly boosting AWS cloud business gross margins. Nebius has built a four-layer business architecture, with Token factories as the core value-added segment, continuously expanding computing power and acquiring inference technology companies, and its self-developed Token Factory has been deployed with multiple overseas customers. Domestic vendors are also actively positioning themselves in the track: cloud vendors are launching Token services based on public clouds; RunJian Co.,Ltd.'s Wuxiang Cloud Valley has built China's first commercial Token factory; PHANCY and Beijing Asiacom Information Technology are expanding Token business through computing power optimization technology and industrial joint venture cooperation, respectively. Different types of enterprises are entering the Token factory track by leveraging their own resource endowments. Risk warnings Risk of intensified competition in the track; risk of technology development and conversion falling short of expectations; risk of dynamic adjustments to regulatory policies.