Networks & InfrastructureHuawei Ascend 950dtChina TelecomAI InfrastructureSubmarine Cable Stations
China Telecom hosts first Huawei Ascend 950DT AI cluster at Shantou cable station
China's first Huawei Ascend 950DT AI cluster has gone live at China Telecom's Shantou submarine cable landing station, a $150 million site processing over three trillion tokens daily for foreign users under a 'token export' model.
Networks & InfrastructureWhy it matters
- China's first Huawei Ascend 950DT AI cluster launched at China Telecom's Shantou submarine cable landing station with about $150 million invested.
- The site processes more than three trillion tokens daily for foreign users under a 'token export' business model.
- Ascend 950DT offers up to 2,007 TFLOPS (MXFP4), 144GB HiZQ 2.0 HBM at 4Tbps, and scales to clusters above 128,000 cards.
- A 1,024-card Ascend cluster with liquid cooling went live in Shanghai in September; DeepSeek plans Ascend use in Inner Mongolia.
- China Telecom also launched a 10,000-chip Alibaba Zhenwu cluster in the Greater Bay Area, planned to scale to 100,000 chips.
The story
A data center running China's first AI cluster of Huawei Ascend 950DT chips has begun operating at China Telecom's submarine cable landing station in Shantou, backed by roughly $150 million in investment, the South China Morning Post reported. The site already processes more than three trillion tokens daily for foreign users.
The project marks one of the clearest commercial deployments yet of Huawei's newest data center accelerator, a 2026 chip built on the Da Vinci v3 architecture for decode-stage inference and training. IT services firm Guangdong Tianyima Information Industry operates the facility, which launched last week. Neither the IT capacity nor the number of chips deployed has been disclosed.
What is the Ascend 950DT?
The specifications are the selling point. The chip delivers up to 2,007 TFLOPS at MXFP4 precision and 1,034 TFLOPS at FP8 or MXFP8, paired with 144GB (or 96GB) of HiZQ 2.0 high-bandwidth memory running at 4Tbps. Huawei says the silicon scales to clusters exceeding 128,000 cards.
The Shantou deployment sits far below that ceiling for now. The undisclosed card count makes it impossible to gauge the site's true throughput against the three-trillion-token daily figure, and vendor scaling claims remain untested at the sizes Huawei describes.
Why a submarine cable station?
Location is the strategy. The Shantou International Submarine Cable Landing Station, built between 1997 and 2001, is one of China Telecom's three mainland landing stations. For a business built on serving foreign AI queries, direct access to international submarine capacity reduces latency and transit complexity.
The site anchors what operator-linked firm TokenHeart describes as a 1+4 computing layout. Shantou aggregates compute power resources, while four satellite data centers — the Eastern Guangdong Information Building, Tuopu, Zhuchi and Chaoshan — provide capacity for elastic scheduling when demand spikes.
What does the 'token export' model change?
The practical question is what this changes for China's AI infrastructure economics. Instead of selling cloud access or model licenses, the operators sell inference output itself — tokens generated for overseas customers, routed through existing telecom infrastructure. Foreign demand at the site now exceeds three trillion tokens per day, according to the report, suggesting early traction for the model.
It also gives Huawei a reference deployment for a chip it positions against foreign accelerators at a moment when US export controls limit Chinese buyers' access to top-tier Nvidia hardware.
Other Ascend deployments stack up
The Shantou launch is not an isolated case. A separate cluster of 1,024 Ascend cards with liquid cooling went live in Shanghai in September, SCMP reported. DeepSeek, the Chinese AI lab behind widely used open models, is said to be planning to use the AI infrastructure at a data center it intends to build in Inner Mongolia.
China Telecom is also hedging across domestic chip suppliers. This month the operator held a launch event for a 10,000-chip AI cluster in the Guangdong-Hong Kong-Macao Greater Bay Area powered by Alibaba's Zhenwu AI chips, with an ultimate target of 100,000 chips at that site.
Taken together, the deployments show China Telecom assembling multi-vendor domestic AI compute at carrier-grade facilities, using its submarine cable assets as an entry point for cross-border inference revenue. How far foreign token demand grows, and whether Huawei's untested cluster scaling holds at five-figure card counts, will determine whether the Shantou model expands beyond this single landing station.
Also reported
Source: datacenterdynamics.com
More from James Calloway
Show full bio
Staff writer covering companies, deals and strategy in Carrier business, 5G and 6G, fiber and broadband at Telecom Gazette.
560 articles