On November 13, China’s first Ascend 950DT computing cluster for exporting tokens went into operation in Shantou. The cluster sits next to China Telecom’s Shantou international submarine cable landing station, which the original report describes as a national-level cable interconnection hub serving the Asia-Pacific region.

Officials describe the project as a three-stage chain: submarine cable landing, local deployment of computing power, and token exports. The business model fails if any one of the three is missing. Without the cable, computing power cannot reach overseas markets; without computing power, the cable is merely a pipe; and without tokens as a billable product, neither of the first two can generate revenue.

What makes the news worth watching, then, is not the claim that this is a national first. It is that the project puts all three elements in the same physical location.

Three Pillars: Chips, Power and Connectivity

The first pillar is chips. Huawei’s Ascend 950DT is a new generation of high-end computing chips designed, according to the original report, for large-model training and high-throughput decoding. It provides the computing power needed for token production at scale and the commercial deployment of AI models. The phrase “high-throughput decoding” matters: decoding is the most throughput-intensive stage of inference, and it is where tokens are generated one by one.

The second pillar is electricity. The original report attributes Shantou’s advantage to three overlapping strengths: a cross-border data compliance pilot in the Overseas Chinese Economic and Cultural Cooperation Pilot Zone, a national-level international submarine cable hub, and green offshore wind power. The first two provide eligibility; the third determines cost.

The third pillar is connectivity, the most easily overlooked part of the business. Submarine cables provide physical reach, while the compliance pilot provides legal reach. Fiber alone is not enough for data to cross borders; the required procedures must also be in place. The original report treats these as separate issues, but they are two prerequisites for the same thing.

Three Partners, Each Responsible for One Link

The division of labor is clear: China Telecom provides network communications resources and the underlying data center infrastructure; Huawei supplies high-end, high-performance computing hardware; and Tianyima handles the project’s overall investment, construction and commercial operations.

The key to this arrangement is that the three types of assets have separate owners: the carrier owns the network, the equipment vendor supplies the chips and systems, and a local listed company takes responsibility for investment and operations. Tianyima is a Shantou-based, publicly traded IT company that, according to the original report, has focused in recent years on token factory construction, intelligent computing services and real-world AI deployment.

Why divide the work this way? Submarine cables and data center facilities are capital-intensive, long-cycle assets tied to licensing requirements, so only a telecom operator can provide them. The chips are Huawei products, but someone still needs to take responsibility for the racks, power supply, cooling and operations as a complete package. The final link carries the most risk: the operator must fund the project upfront and recover the investment gradually through token sales. That role usually falls to a local company with both government relationships and execution capabilities.

Liu Xiangmin said the three parties worked closely together and pooled their technical resources to overcome multiple deployment challenges, allowing the project to be powered on and put into service first in Shantou. The emphasis on being “first” is about the construction schedule, not the technology. The main difficulties in projects like this typically lie in civil works, power supply and acceptance testing, rather than in the chips themselves.

One Plus Four: One Hub and Four Hosting Sites

China Telecom’s Shantou branch describes its computing infrastructure as a “1+4” system: the international submarine cable station connects to global network routes; the eastern Guangdong Shantou data center aggregates computing power resources; and four AI data centers at the Eastern Guangdong Information Building, Tuopu, Zhuchi and Chaoshan provide capacity for elastic scheduling. Together, they create an end-to-end chain spanning cross-border transmission, aggregation and hosting.

Each of the three tiers has a distinct job. The cross-border tier determines where traffic goes; the aggregation tier consolidates and redistributes it; and the hosting tier runs the actual workloads. The original report describes the role of the four AI data centers as “elastic scheduling,” meaning workloads can shift among the four locations according to demand. Tasks can be routed wherever power is more readily available or racks have spare capacity.

This structure also explains why token export operations need to be based in a city with a submarine cable station rather than wherever electricity is cheapest. What is being sold overseas is a deliverable service, and the latency and compliance status of every hop in the network must be quantifiable in the contract.

Thousandfold Growth in Six Months, Yet Demand Is 30 Times Higher

The original report provides a set of figures. In April this year, Shantou completed an end-to-end, city-level token export loop, with daily usage at the time of about 100 million tokens. Six months later, average daily usage had climbed to 100 billion tokens.

That is a thousandfold increase. Such growth would qualify as explosive in any business, but what it really shows is that the project has moved from zero to one, not that it has already reached scale. The starting point of 100 million tokens was so low that it proved little more than that the chain worked.

The figure that follows is more revealing: overseas demand for AI computing power registered for deployment in Shantou now exceeds 3 trillion tokens per day. Based on the two figures in the original report, current daily usage amounts to only about one-thirtieth of registered demand. The demand is there, but capacity has yet to catch up.

The two figures need to be read together. Thousandfold growth shows that the route is operational; the thirtyfold shortfall shows that it remains narrow. Citing only the former is promotion, while focusing only on the latter is unduly pessimistic. Together, they capture the cluster’s actual position today: it does not lack customers; it lacks racks.

Chinese Chips, Models and Power: A Bundled Offering

Lin Shusheng said the project aims to create a new model for taking “Chinese chips, Chinese models and Chinese power” overseas in a coordinated way, while retaining the digital industry’s high added value. The most substantive part of that statement is the final clause.

What did local projects like this sell in the past? Land, electricity, racks and bandwidth. All four have thin margins, and pricing power lies elsewhere. Electricity, bandwidth and rack capacity each have market rates, so orders go to whoever offers the cheapest buildings and power. Tokens are different. They are priced by usage, with the cost reflecting model capabilities, scheduling efficiency and compliance.

“Retaining high added value” means shifting the pricing anchor from how much electricity was consumed or how many racks were occupied to how much deliverable intelligence was produced. Once that anchor moves, the revenue generated by the same data center can rise to an entirely different level. The risk shifts as well, of course, because customers pay only for tokens that are actually produced.

Can This Model Be Replicated?

The original report presents Shantou’s position as the combination of four conditions: Chinese-made chips, offshore green power, international submarine cables and a compliant cross-border channel. Of the four, chips can be purchased and green power can be built. Submarine cables and compliance channels cannot be replicated so easily: the former requires national-level interconnection-hub status, while the latter requires eligibility for a pilot program.

That is why China’s “first” emerged here. Token exports may sound like a software business, but in practice they require a license, a fiber-optic cable and a power plant. Without any one of them, the model cannot work. Only when all three are in place do computing power and models enter the equation.

The real significance of this news, therefore, does not lie in the chip specifications. The Ascend 950DT gives this route something to carry, but the route itself, comprising submarine cables, compliance and green power, is what Shantou is actually bringing to the table. Only a handful of Chinese cities can assemble the same combination.