The Zhitong Finance App learned that Orient Securities released a research report saying that as AI enters the era of inference, token factories can rely on computing power clusters and inference scheduling capabilities to achieve standardized token supply, adapt to the growing trend of large-scale inference demand, have the potential to reduce costs on a large scale, and are optimistic about its long-term business development space as the core carrier for commercializing AI computing power.
Orient Securities's main views are as follows:
The AI industry's focus is shifting to reasoning, and the token market is expanding rapidly
Currently, the AI industry has entered a new stage dominated by reasoning. The industry's focus is shifting from model training to inference implementation, and tokens (terms) are changing from a technical term to a core measurement and pricing unit for AI services. The world's leading model manufacturers have all set up token billing systems to differentiate pricing for input and output tokens, while cloud vendors or professional computing power service providers transform computing power resources into efficient token services, and a complete service chain for the industry is rapidly forming. In terms of token consumption, the number of domestic token calls showed an exponential increase of nearly 5,000 times in two and a half years. Overseas leading models and application vendor token consumption are expanding at the same high rate. With the advent of the Agentic AI era, the number of active agents and total token consumption will maintain a high compound growth rate, providing a good foundation for the implementation of the token factory business model.
Token factory restructures computing power services, three types of players lay out the AI inference industry
Nvidia proposed the transformation of a data center into a token production plant. The essence of the data center is a paradigm shift from resource infrastructure to an intelligent service unit. The Token factory opens up the computing power to the model service link. Relying on the four major links of computing power base, inference optimization, platform scheduling, and token delivery, the computing power is encapsulated as a standardized large-scale model API that is billed per token. Downstream enterprises do not need to build their own computing power deployment models. There is an essential difference between token factories and computing power leasing: computing power leases deliver hardware, long-term billing, and low technical threshold; Token factories output inference services and settlement based on token volume, and rely on full-link reasoning optimization and scheduling technology to form higher barriers and have a better level of gross profit. Track participants are divided into three categories: cloud vendors, advanced computing power assets, and inference technology platforms, and rely on their respective resource endowments to seize industry growth opportunities.
Domestic models, policy drivers, and infrastructure resonance, Token overseas is expected to expand on a large scale
The performance of large domestic models continues to catch up with leading overseas models. At the same time, token pricing has a remarkable cost performance ratio. The proportion of domestic models called on overseas aggregation platforms has increased rapidly, laying the foundation for tokens to go overseas. At the policy level, the country introduced top-level data cross-border rules, and implemented pilot digital processing projects in many places to form a compliance framework where “the original data does not leave the designated area, only exports tokens”; high-capacity international submarine cables have been put into operation one after another to improve cross-border transmission bandwidth and consolidate the physical foundation for tokens to go overseas. Currently, token overseas is divided into two types of models: direct API cross-border connection and data processing. The former allows overseas customers to directly call domestic APIs and is subject to cross-border data supervision; the latter relies on quarantine bonded zones to complete reasoning, and the closed loop of compliance is more complete. With the support of multiple benefits, Token's overseas business is expected to expand on a large scale.
Various domestic and foreign companies are actively doing token factory business
Overseas token factories represent Amazon AWS and Nebius. AWSBedrock uses a tiered billing model based on token and relies on inference optimization technology to improve GPU utilization. Compared with Microsoft and Google's hourly billing models, the token business is more profitable, and the token business significantly drives the gross profit of the AWS cloud business; Nebius has built a four-tier business structure, and the Token factory is the core value-added sector, continuously expanding computing power and acquiring inference technology companies. Its self-developed Token Factory has already landed on many overseas customers. Domestic manufacturers are also actively laying out racetracks. Cloud vendors rely on public cloud to launch token services; Runjian Co., Ltd. built the first commercial token factory in China; Paradigm Intelligence and Yakang Co., Ltd. have expanded their token business with computing power optimization technology and industrial joint venture cooperation methods, and different types of enterprises have relied on their own resource endowments to enter the Token factory circuit.
Risk Alerts
Race track competition increases risk; risk of technology development and transformation falling short of expectations; risk of dynamic adjustment of regulatory policies.