Huang Renxun G20 talks about “AI infrastructure” again: When computing power becomes infrastructure, what kind of “AI factory” does China need?

The Zhitong Finance App learned that on September 2, local time, at the G20 Innovation Ministerial Meeting held in North Carolina, Nvidia CEO Huang Renxun once again brought AI to a broader position. “Ultimately, every country must recognize that AI is an infrastructure, just like water, roads, electricity, and the internet.” Hwang In-hoon said that every country needs to build its own AI infrastructure to provide “digital intelligence capabilities” for domestic researchers, students, industries and startups.

This isn't the first time Hwang In-hoon talks about “AI factories.” Traditional data centers are more like places for storing and processing information, while data centers in the AI era are beginning to assume a new type of production function: electricity enters, goes through computing, networks, storage, and software systems, and eventually produces tokens and intelligent services. Therefore, what needs to be built in the AI era is not only more computer rooms, not only the procurement of more computing equipment, but also an infrastructure system that can continuously stabilize “production intelligence”.

What makes up this system? Who's building it? What does it rely on to keep running? These are the three questions that need to be answered in order to understand the “AI factory”.

Computational power competition is moving from single point resources to system capabilities

One obvious characteristic of this round of global AI infrastructure expansion is that it is getting larger and larger, and at the same time, it involves more and more links. GPUs are just one part of it. To actually build a large-scale AI cluster, it also requires stable power, high-density data centers, liquid cooling systems, high-speed networks, storage, and cluster networking, scheduling, and operation and maintenance capabilities. Recently, global data center investment has continued to heat up, and power and cooling equipment companies have also become important beneficiaries of this round of AI capital expenditure. The market is seeing more and more clearly that behind the shortage of computing power, there is essentially a constraint on the supply capacity of a complete set of infrastructure.

This also explains why today's evaluation of an AI infrastructure company makes it difficult to look at just one number. How much computing power is important, but it is also important: where is this computing power placed, what electricity is used, how long it takes to go online, whether it can run stably, and how many effective tokens can be generated per unit of computing power. From this perspective, an “AI factory” is first and foremost a system project.

Who is building an “AI factory”: two samples, two paths

The direction of systems engineering is easy to explain; the difficult part is implementation. Electricity, land, equipment, networks, scheduling, and even computing power services are scattered across different links in the industrial chain, and there aren't many companies that can organize them into a “factory.” In the past, most of the competition among computing power companies focused on a single resource such as computer rooms, electricity, or equipment, but now it is beginning to extend to a more complete service chain. Looking at the current publicly disclosed business structure, Runze Technology (300442.SZ) and Guangdong-Hong Kong Bay Intelligent Computing (01396) are two representative samples, which correspond to two paths: one relies on an existing base to grow upward, and the other assembles scattered elements into a complete “factory” through integrated delivery.

Runze Technology represents a “bottom-up” path. Years of IDC's business have enabled it to master basic resources such as data centers, electricity, cooling, operation and maintenance — these “heavy assets, long cycle” capabilities are the hardest part of an AI factory to replicate from scratch. As the industry enters the era of high-density computing power, this stock base instead becomes a scarce asset: it can be upgraded to an infrastructure carrying AI computing power without having to be torn down and redone. The data confirms the speed of this transformation: in the first half of 2026, its AIDC business revenue was 1,995 billion yuan, up 126.24% year over year, accounting for more than half of total revenue. Runze is not just a “company that provides cabinets,” but is systematically transforming data center bases into productive infrastructure in the AI era.

Guangdong-Hong Kong Bay Intelligent Computing represents another path — not growing from existing resources, but by engineering to build and hand over an entire “factory”. As of the announcement date of the interim results, the company has delivered and operated FP16 with more than 50,000 P of dense computing power, forming a three-in-one delivery system of “facilities, equipment, and technology”: it not only solves IDC, electricity, bandwidth, and operation and maintenance, but also organizes hardware such as servers, storage, and networks to complete the networking, scheduling and tuning of large-scale clusters, and further extends services to heterogeneous computing power scheduling, model deployment, and inference services, clearly proposing the direction of a “token factory”.

One grows from a resource base, and one directly reaches the entire chain with integrated delivery — different paths and the same direction: organizing elements scattered across all links of the industrial chain into an “AI factory” that operates sustainably.

What is more important than “how much to build” is “how much to produce”

If the AI industry solved the problem of “whether there is computing power” in the past two years, then what really determines the efficiency of the industry in the next stage is probably “how to use computing power.” For the same data center and equipment of the same scale, if electricity costs are different, cluster utilization rates are different, and inference efficiency is different, the final number of tokens and costs generated may be completely different.

Therefore, in the end, AI factories still have to calculate an economic account. Can capital expenditure be converted into stable income? Can high utilization rates be maintained throughout the life of the equipment? Can electricity and operation and maintenance costs continue to drop? Can technical optimization increase the output per unit of computing power? These issues determine whether AI infrastructure is a simple capital expansion or a business that can actually continue to operate.

Hwang In-hoon's continuous emphasis on AI infrastructure actually suggests a larger industrial change: AI competition is expanding in depth from algorithms and models to infrastructure. As far as China is concerned, the “AI factory” it really needs may not simply replicate an overseas model, or bind to a specific type of chip, but rather establish an infrastructure system that can adapt to multiple computing power, large-scale delivery, continuously optimize efficiency, and ultimately stabilize production intelligence.

As the industry's focus shifts further from “computing power scale” to “computing power output efficiency,” AI infrastructure competition will also move from competing for resource reserves to a new stage of competing for token output, energy efficiency, and operational capabilities.