On Sep 30, 2026, IDC officially carved out "AI cloud" as a distinct market and published its first annual sizing: China's AI cloud market reached RMB 68.4 billion in 2025. IDC defines AI cloud as "a complete infrastructure and platform service system delivered as cloud services that supports the full lifecycle of large models" — not a simple upgrade of traditional IaaS, but a new market with a different value anchor.
1. Inference demand drove the market: token calls up 16x
2025 marked the accelerated inflection from training to inference. Inference's share of spending rose sharply as LLM applications spread across internet, finance and manufacturing — in role-play, marketing, search, data processing and document work. The numbers tell the story:
- Enterprise MaaS token calls: 1,994 trillion for the year, up ~16x YoY;
- Daily token consumption: from 1.6 trillion at the start of the year to 9.6 trillion by year-end — 6x in 12 months;
- Market size: RMB 68.4 billion, driven by the surge in inference demand.
2. The pricing anchor has switched: from resources to outcomes
IDC's research shows buyers evaluate AI infrastructure through a clear progression: resources are the threshold, platform is the value, trust is the gate. The anchor has moved from core-hours and GB-months to "scaled delivery of intelligence" — answer quality/model performance, performance, and security & compliance are the top three factors. Infrastructure built purely on selling raw compute no longer matches enterprise needs; market structure and barriers to entry are being rewritten.
3. Market landscape: leaders built comprehensive advantages
| Vendor | 2025 AI cloud share | Note |
|---|---|---|
| Alibaba | 29.7% | "Qwen + Bailian" dual engines, No. 1 overall |
| ByteDance | 14.7% | Volcano Engine, leads MaaS model services |
| Baidu | 8.2% | Search + AI, deep in government & enterprise |
| Huawei | 7.9% | Ascend full-stack domestic compute stack |
IDC stresses that AI cloud competition has moved from "who can provide resources" to "who can turn resources into deliverable, trustworthy outcomes": models are the entry ticket with a shrinking window, inference optimization is the long-term moat, and AI security has become a premium rather than a cost item.
What this report means for you
A 16x jump in enterprise MaaS token calls says "running LLMs on the cloud" is no longer experimental — it's mainstream production. The same logic applies to individual developers, indie teams and SMBs: instead of buying your own compute, put model inference and elastic GPU capacity on the cloud — pay per use, scale in minutes, zero ops, and let compute follow the business.
The dimensions IDC highlights — model quality, performance, security & compliance — are exactly what cloud vendors compete on. Opening cloud GPU / LLM services now is the lowest-effort way to ride this trend.
4. Why move AI workloads to the cloud now
- Cost: self-hosted GPUs cost thousands upfront and churn yearly; cloud is on-demand with pay-per-use;
- Elasticity: inference demand is spiky; scale up/down in minutes instead of buying for peaks;
- Compliance & security: major clouds provide governance, data-residency and audit capabilities — the "trust layer" IDC emphasizes;
- Ecosystem: model services, vector databases and AI-native toolchains are ready out of the box.
Ride the RMB 68.4B wave: put your AI app on the cloud
New-user promo pricing — GPU cloud servers and LLM services on demand, up and running in minutes.