💰 1. 8,790 Billion by 2030: The Numbers That Matter
Goldman Sachs dropped a 50-page deep-dive report on July 10, 2026, and the headline figure is hard to ignore: China's AI large language model (LLM) API and subscription revenue is projected to grow from an estimated 350 billion yuan in 2026 to 8,790 billion yuan by 2030 — a 25-fold increase in four years.
Daily token consumption tracks the same trajectory: from 350 trillion today to 4,600 trillion by decade's end. By 2030, 55% of China's AI model tokens will be consumed overseas — making these models a genuinely global infrastructure layer.
For brand marketers deciding which AI platforms to build on, this report provides the first systematic competitive landscape from a top-tier global investment bank. The findings directly affect content strategy, AI tool selection, and marketing technology investment decisions.
| Metric | 2026 (Estimate) | 2030 (Projected) | Growth |
|---|---|---|---|
| API + subscription revenue | 350B yuan | 8,790B yuan | 25× |
| Daily token consumption | 350T | 4,600T | 13× |
| Industry ARR | ~$10B | ~$125B | 12.5× |
| Training cost | ~$4B | ~$20B | 5× |
| Overseas token share | — | 55% | Dominant |
🔄 2. Cost Revolution → Intelligence Revolution: The Framework
Goldman Sachs frames China's AI evolution in two phases:
| Phase | Timeline | Defining Event | Market Dynamic |
|---|---|---|---|
| Cost Revolution | 2025 | DeepSeek proves "Chinese models can be cheap" | Token consumption explodes, API prices collapse |
| Intelligence Revolution | 2026+ | Zhipu GLM proves "Chinese models can be world-class" | Quality differentiation emerges, two-tier market forms |
The report's core thesis: China has moved from competing on price to competing on intelligence. The question for the next five years is not "can China catch up?" but "who survives the consolidation?"
Three factors drive the transition:
- MoE (Mixture of Experts) architecture keeps parameter activation ratios at just 3-5% — dramatically reducing inference costs without sacrificing quality
- Reinforcement learning post-training (RLHF) pushes coding and agent-task capabilities toward commercial viability
- Open-source distribution creates data flywheel effects — wider deployment → more feedback → faster iteration
⚖️ 3. The Two-Tier Market: Who Wins Where
Goldman Sachs identifies a "double-layer structure" forming in China's AI market:
Premium tier ($1/million tokens):
| Leader | Model | Parameters | Key Advantage |
|---|---|---|---|
| Zhipu (GLM) | GLM 5.2 | 0.7T | Arena text ranking #1, 2026E ARR target $1B |
| Alibaba (Qwen) | Qwen 3.7 Max | — | Cloud ecosystem distribution |
| DeepSeek | V4 Pro | 1.6T | Peak/off-peak pricing from mid-July |
Premium models price at $1/M token — approximately 10-25% of US equivalents ($4-8/M token) — while maintaining 10-20% inference gross margins.
Volume tier ($0.06-0.2/million tokens):
Targeting price-sensitive global SME and consumer markets. For example, MiniMax generates 60-70% of its revenue from overseas users.
The competitive framework (Goldman Sachs 3D model):
| Dimension | Text Model Leaders | Multimodal Leader |
|---|---|---|
| Pricing power | Zhipu, DeepSeek | — |
| Cost advantage | Zhipu, DeepSeek | — |
| Financial strength | — | ByteDance (Seedance) |
| Multimodal moat | — | ByteDance: 70% gross margin, $2B+ ARR |
⚙️ 4. How China Models Achieve 10-25x Cost Efficiency
The math behind China's price advantage matters for brands because it directly affects the cost of AI-powered marketing tools:
| Efficiency Factor | Mechanism | Impact |
|---|---|---|
| MoE architecture | Only 3-5% of total parameters activated per query | 95-97% reduction in compute per token |
| Speculative decoding | DeepSeek DSpark framework boosts generation speed 60%+ | Same GPU serves more users |
| Parameter economy | 200B-1.6T parameters vs 10T+ for global closed models | Lower training cost per iteration |
| Open-source flywheel | Community improvements feed back into base models | Free performance upgrades via ecosystem |
The licensing model is also evolving. The report predicts a shift from pure MIT open-source (free for all) toward "open weight community license" — where commercial use requires revenue-sharing agreements with the model developer. MiniMax M-series has already adopted this approach.
🎯 5. What This Means for Brand AI Strategy
For brands building their AI marketing stack in China, five implications:
| Implication | Detail | Action |
|---|---|---|
| 1. Model costs will keep falling | Premium models at $1/M token today → likely 50-70% lower by 2028 | Avoid long-term exclusive contracts with any single model provider |
| 2. The two-tier market demands tier-aware sourcing | Premium for accuracy-critical tasks (GEO content, compliance); volume for high-frequency tasks (A/B testing, draft generation) | Map each marketing use case to the right tier |
| 3. ByteDance's multimodal lead matters | Seedance dominates video generation (70% margin, $2B+ ARR) | Prioritize ByteDance ecosystem for short-video/visual content tools |
| 4. Open-source creates optionality | Qwen, GLM, DeepSeek all open-source; can be deployed on any cloud | No vendor lock-in; test multiple models in parallel |
| 5. The globalization premium | 55% overseas token share by 2030 means models optimized for international use | Cross-border brands should test Chinese models for non-China markets too |
📋 6. Key Takeaways
- 8,790 billion yuan by 2030. China's AI LLM market is not speculative — it has Goldman Sachs' 25× growth projection with detailed unit economics.
- The transition from cost to intelligence is underway. The market is bifurcating into premium ($1/M token) and volume tiers ($0.06-0.2) — brands need tier-specific sourcing strategies.
- ByteDance leads multimodal; Zhipu + DeepSeek lead text. The competitive map is clear for the first time: choose your AI provider based on use case, not hype.
- MoE efficiency gives China models a structural cost advantage that compounds over time. 3-5% parameter activation means inference costs will continue to decline faster than global peers.
- 55% overseas token share by 2030. Chinese AI models are not just a domestic story — they are becoming global infrastructure, relevant for any brand with cross-border operations.
- Open-source is the default, but it is evolving into a compensated model. Plan for revenue-sharing agreements as the industry matures beyond the MIT-license era.