Huawei Ascend 950DT Computing Cluster Debuts in Shantou, Targeting Overseas Computing Power Export
nashnova research
China's first commercial compute cluster built on Huawei's Ascend 950DT chip went live in Shantou, Guangdong — running AI inference domestically and exporting only the generated tokens. This means China is testing a pathway to sell AI computing power abroad without sending data across the border.
How does this cluster actually work?
The cluster sits at China Telecom's Shantou international submarine-cable landing station. China Telecom provides the network, Huawei supplies the compute hardware, and Tianyi Ma (天意马) handles investment, construction, and commercial operations.
The core model: all AI inference runs inside China. Only the generated tokens — the text, code, or other output units a large model produces — are transmitted to overseas clients.
This means → the data itself never leaves China; only the finished output crosses the border, sidestepping current data-export controls.
How much money, and how much capacity?
Tianyi Ma chairwoman Lin Mingling said this is the "No. 1 Token Factory, Phase 1," with a planned total investment of RMB 1 billion and dedicated compute capacity of 2,000 petaflops.
In plain terms = 2,000P is a mid-sized AI inference data center, purpose-built to serve overseas orders.
Lin added that Factory No. 2 has already been signed, and Nos. 3 and 4 will proceed as joint ventures — an aggressive expansion pace.
What do the Ascend 950DT specs tell us?
Per official figures: 1 petaflop/s at FP8 precision, 2 petaflops/s at FP4, with 144 GB of memory and 4 TB/s bandwidth.
This reflects that Huawei now has commercially viable single-card inference performance. The doubling at FP4 targets large-model inference, not training.
This means → the Ascend 950DT's positioning is clear: don't compete with Nvidia on training; own the inference-export lane.
How fast is Shantou's token demand growing?
Shantou government data: daily token calls grew from roughly 100 million in April to roughly 100 billion six months later. Registered overseas demand exceeds 3 trillion tokens per day.
IT Home flagged an important caveat: registered demand differs from actual daily usage — the registered figure reflects intent, not real consumption.
In plain terms = the demand numbers look striking, but they need a discount. The real test is how much paying usage actually materializes.
With the U.S. tightening export controls, can this model survive?
Context: the U.S. Commerce Department is moving to restrict Chinese companies' access to advanced chips through offshore data centers, further squeezing Chinese AI firms' external compute channels.
This reflects a forced pivot — from "buy overseas compute" to "serve overseas clients with domestic compute." The Shantou cluster is a direct product of that shift.
The South China Morning Post reported that Chinese AI developers' overseas demand has grown notably this year, driven by price advantages. This means → whether "compute at home, deliver tokens abroad" can keep scaling its overseas client base is the key test of this pathway's commercial viability.
市场有风险,内容仅供研究参考,不构成投资建议。
