---
title: "A Chinese Data Center Now Sells AI Answers Abroad. The Chips Stay in Shantou."
date: 2026-11-20
category: Chips & Compute
site: NeuroAI
canonical: https://neuroai.site/a/na-chips-shantou-token-export-cluster
language: en
---

# A Chinese Data Center Now Sells AI Answers Abroad. The Chips Stay in Shantou.

> A cluster built on the Ascend 950DT from Huawei (华为) began serving overseas customers in Shantou (汕头), turning 'token export' into a working business model — and raising new questions about how AI compute is traded.

A user in Southeast Asia opens an app and asks a question. The answer arrives in seconds. The computing that produced it happened on the Chinese coast, thousands of kilometres away. The only thing that crossed the border was the answer itself.

That is the idea behind a cluster that switched on in Shantou, Guangdong, on 13 November 2026. Officials describe it as the first AI computing cluster built around Huawei's Ascend 950DT (昇腾950DT) chip specifically to serve customers outside China.

## What switched on, and who built it

The launch ceremony took place at the Yuedong Information Building in Shantou. It was organized by the Shantou branch of China Telecom (中国电信) together with Guangdong Tianyima Information Industry, a Shenzhen-listed IT services company, and Huawei Technologies. The Shantou municipal government confirmed the event on its official site.

The division of labour is explicit. China Telecom supplies the network links and machine rooms, Huawei supplies the computing hardware, and Tianyima invests in and operates the facility and sells the capacity. A Huawei project manager was quoted as describing the cluster as a key step for Chinese-designed computing going abroad.

The site is not arbitrary. It sits at China Telecom's international submarine cable landing station in Shantou, one node in a local system of one landing-station intelligence centre plus four regional ones. The arrangement gives the project a direct path from computing in Shantou to markets in Southeast Asia.

## What "token export" actually means

The phrase behind the project is token export (词元出海): a term for selling the output of AI computation across a border while keeping the computation and the data at home.

Here is the sequence. An overseas user sends a request to an AI model. The model runs in a Chinese data centre. The result travels back to the user, billed in tokens — the units in which AI services are priced. China Telecom has described the principle as compute going out while data does not cross the border. The company also talks about making compute as metered, schedulable and tradable as water or electricity.

This distinction matters in a market where access to advanced accelerators is politically contested. Exporting a chip is a visible shipment that can be restricted. Selling the result of a chip's work is a different category of transaction, and regulators in other countries have only begun to consider how to treat it.

## The hardware and the price tag

The Ascend 950DT is a new-generation accelerator aimed at both inference and training. According to the Shantou government release and company material, a single card delivers one petaflop of FP8 compute, two petaflops of FP4, carries 144GB of memory and offers four terabytes per second of bandwidth. Large memory and high bandwidth are what make it suitable for running big models and generating tokens at scale.

Tianyima calls this the first phase of its "No. 1 token factory," with a planned investment of RMB 1 billion (≈ US$141M / HK$1.09B) and 2,000 petaflops of dedicated computing power. A second factory has been signed; a third and fourth are planned as joint ventures.

At one petaflop per card in FP8, 2,000 petaflops would imply roughly 2,000 chips — or about 1,000 if the FP4 figure is used. The release does not say which measure applies, and an earlier Sina Finance report said Tianyima had installed just 16 chips by the end of October, powered on 19 October, with service expected by the end of November. The gap between the planned figure and the confirmed installed hardware is the single most important thing to keep in mind when reading the announcements.

## The traffic numbers, and where they come from

The zone around the landing station has been used as a test bed for selling AI output abroad for months. According to reporting from People's Daily carried by Xinhua in May, daily token calls in the zone reached about 100 million in late April and passed 10 billion less than a month later. By a token-export conference in September, that figure had passed 100 billion.

The first paid transaction took place on 18 May, when a user in Singapore talked to an AI toy whose requests were processed in Shantou, billed at RMB 2 per million tokens (≈ US$0.28 / HK$2.17). By September, 132 companies had been approved under the zone's "data coming in for processing" pilot, which allows foreign data to be processed locally and the results returned.

Every one of these figures comes from local officials, state media or the companies involved. None is an independently audited business disclosure. They describe traffic and approvals, not profit, and the registered overseas demand cited by the local release — more than three trillion a day, without a stated unit — should be treated with particular caution.

## Key takeaways

- A cluster built on Huawei's Ascend 950DT (昇腾950DT) began operating in Shantou on 13 November 2026, aimed at overseas customers.

- China Telecom, Huawei and Guangdong Tianyima built it; the site uses an international submarine cable landing station to reach Southeast Asia.

- "Token export" (词元出海) means running the AI computation in China and sending only the billed output abroad.

- The first phase is described as a RMB 1 billion (≈ US$141M / HK$1.09B) project with 2,000 petaflops of planned capacity.

- Confirmed installed hardware appears far smaller than the plan, and all traffic figures come from officials and companies.

## Honest limitations

How much hardware is actually running is unclear. Public reporting says 16 chips were installed by the end of October, while the project description speaks of 2,000 petaflops of planned capacity. I found no disclosure of customer names, revenue or utilization rates.

The chip specifications come from the Shantou government release and company material, not from an independent test. The "first cluster" claim is the government's framing, though multiple outlets reported the same event.

The traffic figures are from local officials and state media, and they count calls and approvals rather than settled revenue. The three-trillion-a-day demand figure has no unit attached and is not used as a fact in this article. Export rules for compute delivered as a service are still evolving, so the model's regulatory future is uncertain.

## What readers can do now

- **If you buy AI services, ask where the computation happens.** A low price may reflect a cross-border arrangement with different data and compliance implications.

- **If you follow the chip race, watch service exports as well as shipments.** Selling inference across a border is a different route around restrictions than selling accelerators.

- **Separate announced capacity from installed hardware.** The 2,000-petaflop figure is a plan; the confirmed installation was far smaller.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-shantou-token-export-cluster
Free to quote with attribution and a link to the original.
