Market Opportunity
Trillion-scale compute + MaaS demand meets massive supply-side idleness
Macro backdrop
China's IDC industry has grown from a ~2.16 billion RMB market in 2006 into a trillion-scale business that now spans AI training, inference and cloud compute — not just colocation.
| Metric | Figure | Source |
|---|---|---|
| 2025 China smart-compute capacity | ~1037.3 EFLOPS | IDC |
| 2026 projected capacity | ~1180 EFLOPS | IDC |
| Smart-compute growth | 3×+ faster than general compute | IDC |
| 2025 smart-compute services market | >130 billion RMB | CAICT |
| 2026 compute rental market | ~260 billion RMB, +43% YoY | Industry research |
MaaS: on the edge of explosion
Model-as-a-Service packages large-model capability as a subscription API, so SMEs and developers can use it without training their own hundred-billion-parameter models.
- China's MaaS market is projected to exceed a trillion RMB between 2025 and 2030.
- Daily enterprise LLM calls hit 37 trillion tokens in H2 2025, up 263% from 10.2T in H1.
- Volcano Engine's Doubao model alone has crossed 120 trillion tokens per day.
- OpenRouter data: in February 2026, Chinese AI model calls jumped 127% over three weeks and Chinese models took four of the global top five, a combined 85.7% share.
Compute oversupply meets undersupply
Chinese compute shows a sharp supply/demand mismatch:
- Heavy idleness. CAICT reports average utilization of ~32% across online smart data centers. General-purpose enterprise compute is worse — only 10–15%. Huge swaths of capacity are dormant.
- Strong demand for sharing. The national compute platform has onboarded provincial nodes in 10+ regions with 1,000+ registered enterprises, yet efficiency still has large headroom.
- Idleness and shortage co-exist. As of March 2025, China had 10.43 million standard racks in use and 748 EFLOPS of smart compute, a meaningful fraction of it idle for operational reasons.
That gap is LinkCompute's arbitrage space: plug idle compute into the platform → deploy open-source models → sell as APIs to developers.
AI Agents: the next breakout
AI is moving from "model capability" to "agentic execution":
- Chinese enterprise Agent market: ~19 billion RMB in 2025
- 2025–2028 CAGR: >110%
- Full Chinese AI Agent market has already passed 58 billion RMB in 2025, with enterprise >70%
- 300+ Chinese Agent service providers, forming three layers: big-tech compute base → agent-dev platforms → vertical agent apps
Where the market has no answer
No one has stitched together the full chain of compute → models → evaluation → tools → AI orchestration.
- Cloud providers lock you into their ecosystem and can't offer neutral cross-platform pricing
- Model aggregators only forward APIs — no compute, no evaluation
- Leaderboards are mostly static benchmarks, disconnected from real usage
- Agent marketplaces without underlying call volume never reach escape velocity
That whitespace is LinkCompute's opening.
Five core pain points — one per layer
Each layer of our architecture maps directly to a pain point:
- Compute mismatch. Hyperscalers serve big accounts; smaller developers are underserved while compute centers sit idle.
- Model selection is hard. Too many models, opaque pricing, single channels, no authoritative comparison.
- No trustworthy benchmark. Existing leaderboards are static; the market lacks dynamic references built from real usage data (e.g. same model, different data centers, real price/latency).
- Tool distribution is broken. Independent Agent builders have no promotion channel and no monetization path.
- Going global is expensive. Chinese compute providers and AI developers lack a compliant, efficient, one-stop pipeline for cross-border distribution and settlement.
Last updated on