At the G20, Jensen Huang discussed "AI infrastructure" again: What kind of "AI factories" does China need when computing power becomes infrastructure?

On September 2, local time, during the G20 Innovation Ministerial Meeting held in North Carolina, USA, Nvidia CEO Jensen Huang once again elevated AI to a more macro position. Ultimately, every country must recognize that AI is an infrastructure, akin to water, roads, electricity, and the internet, Huang stated. He emphasized that each country needs to build its own AI infrastructure to provide digital intelligence capabilities for its researchers, students, industries, and startups. This is not the first time Huang has discussed the concept of the AI factory. Traditional data centers primarily serve as storage and processing locations for information, while data centers in the age of AI are beginning to assume a new production role: power enters, goes through computing, networking, storage, and software systems, ultimately producing tokens and intelligent services. Therefore, what needs to be built in the AI era is not just more server rooms or more computing devices, but a system that can continuously and stably produce intelligence. What is this system composed of? Who is building it? What sustains its operation? These are the three questions that need to be answered sequentially to understand the AI factory. The competition for computing power is shifting from isolated resources to systemic capabilities. An obvious characteristic of this round of global AI infrastructure expansion is its increasing scale and the growing number of involved components. GPUs are just one part of the equation. Building a large-scale AI cluster requires stable electricity, high-density data centers, liquid cooling systems, high-speed networks, storage, as well as cluster networking, scheduling, and operational capabilities. Recently, global data center investment has continued to surge, and companies supplying power and cooling equipment have become significant beneficiaries of this round of AI capital expenditure. The market is increasingly clear: behind the computing power shortages lies a fundamental constraint on a whole set of infrastructure supply capabilities. This also explains why today it is challenging to evaluate an AI infrastructure company by looking at just one number. While the amount of computing power is certainly important, equally crucial are factors like where this computing power is located, what type of electricity is used, how quickly it can go online, whether it can operate stably, and how many effective tokens can ultimately be produced per unit of computing power. From this perspective, the AI factory is primarily a system engineering project. Who is building the "AI factory": two samples, two paths. The direction of system engineering is easy to articulate, but execution is more challenging. Power, land, equipment, networking, scheduling, and even computing power services are dispersed across various segments of the industrial chain, and there are not many companies capable of organizing them into a factory. In the past, competition among computing power companies was largely concentrated on single resources such as server rooms, electricity, or equipment, but now it is beginning to extend towards a more comprehensive service chain. Based on the currently disclosed business structures, Range Intelligent Computing Technology Group (300442.SZ) and GBA AI COMP (01396) are representative samples, corresponding to two distinct paths: one growing upwards based on an existing foundation, and the other assembling dispersed elements into a complete factory through integrated delivery. Range Intelligent Computing Technology Group represents the bottom-up approach. Years of IDC operations have provided it with foundational resources such as data centers, power, cooling, and operational capabilitiesthese heavy assets and long cycles are precisely the areas where it is most challenging to start from scratch in the AI factory. As the industry transitions into an era of high-density computing power, this existing foundation has become a scarce asset: it can be upgraded to support AI computing power without starting over. Data supports the speed of this transformation: in the first half of 2026, its AIDC business revenue reached 1.995 billion yuan, a year-on-year increase of 126.24%, accounting for over half of total revenue. Range is no longer just a cabinet provider; it is systematically transforming its data center foundation into productive infrastructure for the AI era. GBA AI COMP represents an alternative pathnot one that grows from existing resources, but one that constructs an entire factory in an engineered manner and delivers it. As of the mid-term performance announcement date, the company has delivered and is operating over 50,000 P of FP16 dense computing power, forming a delivery system that integrates facilities, equipment, and technology: it addresses IDC, power, bandwidth, and operations, while also organizing hardware such as servers, storage, and networks, completing networking, scheduling, and optimization for large-scale clusters, and further extending into heterogeneous computing scheduling, model deployment, and inference services, clearly defining its direction as a token factory. One company is growing upwards from foundational resources, while the other directly reaches the entire chain via integrated deliverydespite differing paths, both aim towards the same goal: organizing elements dispersed across various stages of the industrial chain into a sustainable AI factory. More important than how much to build is how much to produce. If the AI industry addressed the question of whether there is computing power in the past two years, the next phase that will truly determine industry efficiency may be how to use computing power. With the same data center and the same scale of equipment, if the cost of electricity, cluster utilization rate, and inference efficiency differ, the number of tokens produced and their costs can vary drastically. Thus, the AI factory ultimately needs to make an economic calculation. Can capital expenditure stable revenue? Can a high utilization rate be maintained throughout the equipment's lifecycle? Can electricity and operational costs continue to decline? Can technical optimizations improve token output per unit of computing power? These questions determine whether AI infrastructure is merely a straightforward capital expansion or a genuinely sustainable business model. Jensen Huang's continuous emphasis on AI infrastructure actually hints at a larger industrial shift: the competition in artificial intelligence is extending from algorithms and models to infrastructure. For China, the truly needed AI factory may not be a simple replication of any overseas model nor tied to any specific chip, but rather the establishment of a system that can adapt to diverse computing power, scale delivery, continuously optimize efficiency, and ultimately stably produce intelligence. As the industry shifts its focus from scale of computing power to efficiency of computing power output, competition in AI infrastructure will also transition from a comparison of resource reserves to a new phase of rivalry based on token output, energy efficiency, and operational capabilities.