A report from the China Telecom Research Institute projects that AI agents will drive a tenfold annual increase in computing demand, with inference tasks expected to account for 80% of the market by 2029. The shift toward agentic AI is prompting Chinese telecom operators to transition from model-building to token-based billing models for intelligent services.
The transition toward agentic AI signifies a fundamental shift in the streaming and tech infrastructure stack from one-time training costs to continuous operational expenses. As China Telecom, China Mobile, and China Unicom move toward token-based billing models, AI is being repositioned as a metered utility similar to mobile data. This shift places immense pressure on domestic hardware providers like SMIC and Huawei to scale production of inference-optimized chips despite ongoing HBM shortages and export restrictions. Watch for whether these token-based pricing tiers become the global standard for consumer-facing AI services as Western cloud providers face similar hyperscaler spend for consumer AI agents as Western cloud providers face similar margin pressures.
Huawei's Ascend chip line has become the focal point of China's domestic AI compute strategy as telecom operators scale agentic workloads. In August 2025, Huawei unveiled its Ascend 910C processor to select Chinese AI firms as a competitor to Nvidia's H100, positioning the chip for large-scale inference deployments across cloud and telecom infrastructure. China Telecom's Xingchen TokenHub platform, which provides token-based billing for AI agent services, relies on Ascend-series accelerators for inference workloads, tying the operator's commercial model directly to Huawei's silicon roadmap. China Mobile and China Unicom have similarly integrated Ascend hardware into their respective AI platforms, creating a domestic compute ecosystem that bypasses Western chip supply chains.
The regulatory and trade environment continues to shape how quickly these operators can scale. In March 2025, the U.S. Commerce Department expanded export controls to cover additional AI accelerator and HBM-related technologies destined for China, tightening restrictions on advanced memory and interconnect components that Chinese fabs like SMIC and CXMT need for high-bandwidth inference chips. SMIC has responded by accelerating its 7nm-class N+2 process node for AI accelerator production, though yields remain below those of TSMC's equivalent process. China's Ministry of Industry and Information Technology issued guidance in June 2025 encouraging telecom operators to prioritize domestically produced AI chips for government and public-sector workloads, a policy that directly benefits Huawei Ascend AI chips line and reinforces the token-based billing models that China Telecom and its peers are building.
On the technical side, independent benchmarking of Ascend 910C against Nvidia's H100 shows mixed results for inference-specific tasks. SemiAnalysis published a comparative evaluation in July 2025 finding that Ascend 910C achieved roughly 70% of H100 throughput on Llama-3 70B inference workloads when using Huawei's CANN software stack, with the gap narrowing to 85% on smaller models optimized for Huawei's MindSpore framework. For telecom operators running agentic AI media workflows at scale, where latency and cost-per-token matter more than peak FLOPS, these figures suggest Ascend hardware is viable for production inference but still trails Nvidia on efficiency for the largest frontier models. China Telecom's projection of 80% inference share by 2029 implies that even modest per-chip efficiency gains across millions of deployed tokens per day will compound into significant cost advantages for operators who optimize their software stacks early.
China Telecom projects a tenfold increase in annual compute demand over the next three years, driven by the rise of agentic AI. This shift marks a transition from one-time training costs to continuous operational expenses, with inference tasks expected to account for 80% of total compute power by 2029.
The surge is driven by the adoption of agentic AI, which requires continuous operational compute power for inference tasks rather than just one-time model training.
Inference computing is projected to overtake training to reach an 80% market share by 2029.
Operators like China Telecom are moving toward token-based billing models, repositioning AI as a metered utility similar to mobile data.
China Telecom's Xingchen TokenHub platform relies on Huawei's Ascend-series accelerators for its inference workloads.
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source