On November 16, the launch event for Guangzhou’s domestic 10,000-chip AI cluster in the Guangdong-Hong Kong-Macao Greater Bay Area was held at China Telecom’s Innovation Incubation (Southern China) Base in Huangpu District. Guided by the Guangzhou Artificial Intelligence Industry Development Office and jointly organized by China Telecom Guangdong and Alibaba Cloud Intelligence, the event marked the official start of commercial operations for Guangzhou’s first domestic 10,000-chip AI cluster powered by Zhenwu.

First, consider the organizers: one telecom operator and one cloud provider. Nearly every aspect of the product design that follows reflects this division of labor.

Full-Stack Technological Autonomy, With Three Components From Three Providers

According to the original account, the cluster is built around the domestically developed Zhenwu series of AI chips, achieving full-stack technological autonomy across chips, servers and the scheduling platform. The three components appear in the same sentence, but they come from different providers.

Zhenwu supplies the chips. Servers are the layer that puts those chips into racks, while the scheduling platform organizes the machines to do useful work. The first two are largely matters of procurement and integration. The third is what this launch was really about: chips can be bought and servers can be assembled, but a scheduling platform cannot simply be purchased off the shelf.

That also explains why a cloud provider had to be among the organizers. The challenge of a 10,000-chip cluster is never just powering up 10,000 chips. It is getting them to run the same workload under a unified scheduler: which jobs run first, which wait for others, how communications are routed, and how a failed chip is removed before computation restarts. That expertise resides with cloud providers, not in the server room.

The 10,000 Chips Are Not the Point. Connecting to the City’s Computing Power Network Is

The original account includes another line that is easy to overlook: the cluster is “connected to the city’s computing power network.”

Those few words define not its scale, but its role. A standalone 10,000-chip cluster is an island, running at full capacity when demand is high and sitting idle when it is not. Once connected to the city’s computing power network, it becomes a node that can be centrally scheduled, called on by other nodes, and draw capacity from them in turn. For buyers, that means they do not need to build their own infrastructure. For sellers, it means lower idle capacity.

The statement that the cluster will “serve local industries in Guangzhou while extending across the Guangdong-Hong Kong-Macao Greater Bay Area” makes the same point. That reach depends on network connectivity. Without it, the cluster can serve only its immediate neighbors.

TokenHub: Selling Computing Power by the Token

The event also introduced the next-generation TokenHub platform. According to an expert from China Telecom’s Guangzhou branch, the platform draws on a Token Factory and city-level computing power scheduling infrastructure to establish a token-level operating model. It connects the full chain from computing power supply and platform scheduling to application services, creating a closed-loop service system.

“Token-level operations” is the key phrase. Computing power has traditionally been sold by chip-hour, rack or kilowatt, requiring customers to estimate how many chips they need and for how long. Overestimate and they waste money; underestimate and capacity falls short. Token-based pricing works differently: instead of asking how many chips a customer wants, it charges for the volume of model usage. Customers pay only for what they use, lowering the barrier to entry. The burden then shifts to the platform, from “leasing out machines” to “ensuring every request succeeds.”

The platform also supports private deployment and custom development. Both are somewhat counterintuitive in the context of a conventional cloud business, which depends on standardization and scale, while private deployment works against scale. But they make sense in government and healthcare, where keeping data within a designated domain is a hard requirement. Scale may have to be sacrificed to satisfy that condition first.

Deployment Across Government, Industry and Healthcare

The agreements signed at the event covered three levels. The first was government: China Telecom’s Guangzhou branch signed a strategic cooperation agreement with the Huangpu District government to advance the local AI data center industry. The second was industry, with group signings involving companies in intelligent manufacturing, the internet, energy and intelligent connected vehicles. The third was healthcare. Experts from the Guangzhou National Laboratory discussed using domestic 10,000-chip computing power and AI for Science to reshape drug development and make target discovery “computable and predictable.” The Guangzhou Healthcare AI Computing Power Node was also inaugurated, alongside cooperation agreements involving several hospitals and research institutions.

Healthcare is the most consequential of the three. Drug development workloads are unusually demanding: individual jobs are compute-intensive and long-running, accuracy is critical, and failures far outnumber successes. What matters most for this type of workload is not peak performance, but the ability to run continuously for months without disruption. The original account describes the process as “computable and predictable.” Here, “predictable” is fundamentally about stability.

Three Terms, Three Layers

China Telecom Guangdong summed up the strategy this way: the 10,000-chip cluster serves as the computing power foundation, token-based operations as the scheduling link, and industry deployment as the central focus. The three correspond neatly to three layers: hardware, platform and use cases.

This structure helps explain why a telecom operator would build another cluster in a city that already has numerous AI data centers. Most remain stuck at the first layer, selling racks and electricity. Moving to the second depends on having a platform that can bill by the token and schedule workloads across nodes. Once that layer is in place, the same server room can sell a fundamentally different product.

Whether the model works will ultimately depend not on how many companies signed agreements at the launch, but on two metrics: the volume of requests actually running through TokenHub, and how many customers beyond the Huangpu District government both signed contracts and paid. The original account lays out the vision. The answer will come next quarter.