China's Three Major Telecom Operators Bet on AI Computing Token Factories
Nashnova编辑部
China Mobile, China Telecom, and China Unicom all flagged AI compute and token output as core growth drivers in their first-half results — smart-compute revenue surged as much as 95% year-on-year — as telcos shift from selling infrastructure to selling metered intelligence.
What are the telcos actually selling now?
Until recently the business was data centers and server leases — landlord economics, essentially.
Now all three are rolling out token bundles and token marketplaces, charging users per unit of AI output.
This means → the revenue model is moving from "sell compute" to "sell measurable intelligent output" — closer to the end user, with more pricing power.
How do the three scorecards compare?
China Telecom's smart-compute revenue jumped 95% year-on-year, lifting its broader smart-services line to RMB 31.1 billion; cloud revenue rose 7.8% to RMB 61.8 billion.
China Unicom's total compute revenue hit RMB 41.9 billion, up 13%, spanning data centers, compute leasing, and cloud-based AI services.
In plain terms = Telecom is growing fastest off a smaller base; Unicom has greater scale but gentler momentum — neither path is proven yet.
Token platforms — the telcos' AI app store?
China Telecom's Starchen TokenHub hosts 142 large language models and over 420 industry-specific AI applications, with token bundles now available to its national mobile subscriber base.
China Unicom's UniAI platform has onboarded 200-plus large models and accumulated over 500 terabytes of curated data; its "token supermarket" covers the full cycle from token creation to consumption.
This means → the telcos are not just selling compute — they are competing for the AI-application distribution gateway. Whichever ecosystem activates first gains an extra layer of take-rate.
Why do chip compatibility and the software layer matter?
China Telecom upgraded its compute-orchestration software "Xirang" — a scheduling system that coordinates different chip architectures — to support more than 20 chip types.
This reflects a supply-constrained chip environment where the ability to harness multiple domestic chips is itself a competitive moat.
Put simply = if rivals can run one or two chip types, and you can marshal twenty-plus into a single pool, your compute capacity is inherently larger and more flexible.
Can the "token factory" model sustain itself?
All three telcos now list token volume alongside compute and data as a core operating metric — the metering framework is set, giving commercialization an anchor.
Whether token revenue keeps growing depends on whether downstream enterprises and consumers are willing to keep paying for AI output.
This means → the key proof point in coming earnings is not how much compute capacity expands, but whether token pricing and volume can both hold.
Content is for reference only, not financial advice.