Introduction
Cloud computing has transitioned from general-purpose computing to intelligent computing, and will further evolve from intelligent computing to Token computing. In the era of Token computing, the entire computing power industry chain will inevitably undergo tremendous changes.
From chips to equipment, from intelligent computing centers to AI factories, and from model Tokens to AI agents. In today's article, we aim to spark discussion by briefly exploring the Token industry chain.
1 Overview of the Token Industry Chain
The Token industry chain encompasses the following segments:
Chips: Primarily GPUs and dedicated AI chips.
Server Equipment: GPU servers or other servers equipped with dedicated AI chips.
IDC Data Centers: Mainly involving power, data center facilities, and cabinets for intelligent computing centers, with the external output being data center cabinets.
Intelligent Computing Centers: Equipment + Network + Data Center Cabinets = Intelligent Computing Center, outputting computing power externally. Computing power serves as the raw material for Tokens.
Large Models: Models act as the production lines for Tokens, capable of processing computing power into Tokens.
AI Factories: The actual production carriers for Tokens. Using models as production lines, they convert the computing power (raw materials) from intelligent computing centers into model Tokens (final products) for large-scale external sales.
AI Stores: Token aggregation and sales platforms responsible for converting wholesale to retail, allowing users to make one-stop purchases.
AI Agents and AI Tools: Responsible for the practical implementation of specific AI businesses, consuming massive amounts of Tokens.
AI Users: The end enterprises or individuals using AI, providing relevant scenarios and requirements.
Technical Services: Each segment can be developed in-house or through collaboration with third-party technical service companies. Technical service providers offer technology and other related services for one or more segments of the Token industry chain.
2 Token-Native Chips and Models
Chips and models represent the most core segments of the entire Token industry chain. The most prominent company in the chip sector is NVIDIA, while domestic players include Huawei, Muxi, and Moore Threads. In the model sector, the leading companies in the US are OpenAI, Anthropic, and Google, whereas domestic leaders include ByteDance, Alibaba, Zhipu AI, Kimi, and Minimax.
Given the rise of Tokens, future efforts must focus on developing Token-native chips tailored to the specific characteristics of Tokens. For domestic chips, the Token boom presents: overall benefits with localized drawbacks. The primary reasons are:
Tokens involve inference, which has relatively low requirements for scale-up capabilities, benefiting domestic chips.
Token services are a type of Serverless service. Architectural and ecosystem issues, which are disadvantages for domestic chips, are resolved on the platform side, sparing end developers from worrying about ecosystem details. The rise of Tokens is advantageous for domestic chips.
The Token model entails fiercer competition. Traditionally, different chip manufacturers strive independently across various industries and scenarios with relatively less competition. However, the Token model brings everyone onto the same track, putting their capabilities to the ultimate test, which makes the competition even more brutal.
The core competitiveness of Token services lies in pricing, leading to greater competitive pressure.
The Token service market will experience an order-of-magnitude explosion.
We anticipate a massive explosion of domestic chips in the Token market in 2 to 3 years.
3 Token-Native IDC Data Centers, Equipment, and Intelligent Computing Centers
Tokens are a relatively standardized product with massive scale; therefore, the entire related infrastructure inevitably requires full-stack optimization:
Token-dedicated server equipment is required. It needs to be custom-built for higher efficiency and lower costs, without the need to accommodate traditional intelligent computing requirements.
More cost-effective IDC data centers. For instance, traditional IDCs require dual municipal power feeds, generators, UPS, and other power supply safeguards. However, Token systems significantly reduce the reliability requirements for infrastructure, meaning the power supply system can even be downgraded to a single municipal power feed.
There are two directional choices for the construction of intelligent computing centers:
One is ultra-large-scale centralized construction, which is more conducive to cost reduction. Such intelligent computing centers are typically built in western cities rich in green power resources.
The other is small-scale, elastic edge intelligent computing centers, which can fully utilize small-scale idle power, space, and other resources in eastern cities. These edge intelligent computing centers can not only serve ultra-low latency business scenario requirements but also leverage resources to reduce costs.
In the Token era, where cost sensitivity is extreme, both types of intelligent computing centers require further optimization of construction, scheduling, and other related costs.
4 Production and Sales of Tokens
Within the Token industry chain: from a technical perspective, chips and models are core; from a business perspective, the production and sales of Tokens are core.
For the production and sales of Tokens, we analyze both using four aspects: technology, content, input, and output.
Regarding Token production in AI factories:
Technology. This primarily refers to model deployment and performance optimization. In Token production, technology is extremely critical. Token production is not merely simple model deployment. Tokens are standardized products and are price-sensitive, necessitating relentless efforts in performance optimization. Optimization requires "continuous improvement and pursuit of excellence." It must evolve from software optimization to deeper hardware-software co-optimization, and even chip architecture optimization.
Content. Content refers to models. For many AI factories, currently, they can only produce Tokens for open-source models. In the future, they can further collaborate with model providers to deploy closed-source models, or even go a step further to develop proprietary models.
Input. The input is computing power. This computing power can be self-built or leased. From the perspective of the industry chain, intelligent computing centers and AI factories are two independent business forms, meaning the decoupling and separation of intelligent computing center business and AI factory business. For AI factories, once mature and stable, the core competitiveness will become "how to obtain lower-priced and larger-scale computing power from upstream?".
Output. Its output product is Tokens, an exceptionally massive volume of Tokens. A small portion of Tokens from AI factories can be retailed independently, while the majority must be sold wholesale through sales channels such as aggregation platforms.
Regarding AI stores (Token aggregation and sales platforms):
Technology. The technology mainly involves the model aggregation platform. The technical development difficulty is relatively low; it requires more long-term post-launch operations to connect with more users, as well as AI agents and tools.
Content. The primary content of the aggregation platform is model Tokens. In the long run, it can also include AI agents.
Input. The input for AI stores is a massive volume of Tokens. They wholesale Tokens from upstream for retail.
Output. The output consists of retail Tokens, which may also include AI agents and the like. Contrary to AI factories, for AI stores, once stabilized, their core competitiveness lies in sales capabilities and the quantity and quality of connected customers. Therefore, AI stores need to invest more effort in sales and innovate with more diverse strategies.
Finally, the most crucial point to emphasize is "Token business, scale first." Only by scaling up can it be possible to amortize technology and other investment costs, negotiate with upstream providers to obtain lower-cost "raw materials," and achieve a larger market scale and higher profits.
5 AI Agents and User Business Scenarios
AI agents are model-based AI applications that cater to business scenarios to meet user requirements, serving as the ultimate carriers for business implementation.
Ultimately, it is the combination of "models + AI agents" that fulfills user requirements.
From the perspective of Tokens, AI factories produce Tokens, and AI stores sell Tokens; thus, AI agents and various AI tools (AI tools are also AI agents) consume Tokens.
It is essential to focus not only on model capabilities and Token prices but, more importantly, on business scenario requirements and AI agent applications.
6 Token Technical Services
Market competition is particularly fierce. As the saying goes, "use the best steel on the edge of the blade." Every company needs to focus on its most core value competitiveness while seeking external cooperation and support for common "reinventing the wheel" tasks as much as possible. The division of labor and cooperation among enterprises collectively form a complex industry chain.
Based on the "Pareto Principle" (80/20 rule), Tokens have two characteristics: a small number of products, accounting for only about 5-10% of the total intelligent computing cloud products; and a massive market scale, representing 80-90% of the entire intelligent computing cloud market size.
Such business characteristics enable Token-related technologies to serve as common technologies, allowing for unified R&D and amortization of R&D costs. Consequently, the business model of Token technical services has emerged.
Token technical services, along with all segments—especially the most closely related raw materials (intelligent computing centers), production (AI factories), sales (AI stores), and consumption (AI agents and AI tools)—collectively form an open, collaborative, and healthily developing Token industry chain.
(End of Article)
Juxiang Technology is an AI industry chain technical service enterprise focusing on the production and sales of Tokens. It has successfully implemented AI factory and AI store projects with numerous partners both domestically and internationally. Friends interested in the Token business are welcome to communicate, discuss, and jointly explore business opportunities.