The last major AI enterprise in China is also struggling. Baidu’s ERNIE Bot has announced that it will be free. How will future AI large models make money?

I. Direct Monetization: Technology as a Service (TaaS)

1. API Call Fees
   • Model: Charge per request, output length, or computational load (e.g., OpenAI’s GPT-4 at ~$0.06 per 1k tokens).
   • Advantage: Lightweight, scalable, and ideal for developers/enterprises.
   • Examples: Anthropic’s Claude API, Google’s PaLM 2 API.
2. Vertical-Specific Model Subscriptions
   • Model: Offer high-precision models for industries like healthcare, law, or finance via subscription plans.
   • Logic: Generic models cannot replace domain expertise; proprietary data creates pricing power.
   • Example: Morgan Stanley’s customized GPT-4 for wealth management analysis.

II. Industry Transformation: Embedding into Workflows

1. Industrialized Content Generation
   • Applications: Movie script/storyboard generation, ad copywriting, game NPC dialogue systems.
   • Revenue: Revenue-sharing based on content value (e.g.,短视频 scripts tied to viewership).
   • Example: Runway ML’s role in generating effects for Everything Everywhere All at Once.
2. Enterprise Process Automation
   • Use Cases: Contract review, customer service ticket handling, supply chain forecasting.
   • Model: SaaS deployment with modular pricing (e.g., Salesforce integrating GPT).
   • Data Edge: Fine-tuning models with private enterprise data to build competitive moats.

III. Indirect Revenue: Data & Ecosystem Building

1. Data Flywheel Effect
   • Logic: User-generated feedback data continuously improves models, creating a closed loop.
   • Example: MidJourney refining its image models via user-generated content.
2. Developer Ecosystem Revenue Sharing
   • Model: Open model capabilities to developers, taking a cut of app revenue (similar to App Store).
   • Example: Hugging Face’s 10-15% commission on its model marketplace.

IV. Hardware & Computing Power Competition

1. Dedicated Chips & Compute Leasing
   • Demand: Training LLMs costs millions per run; inference requires cost-efficient hardware.
   • Key Players: NVIDIA (H100 GPUs), AWS (Trainium chips), Cerebras (wafer-scale engines).
   • Trend: Computing power as the “new oil”—hardware dominance drives model evolution.
2. Edge Computing Deployment
   • Use Cases: Lightweight local models (e.g., on-device Stable Diffusion) to reduce cloud costs.
   • Example: Apple integrating locally run LLMs in iOS 18 to push hardware upgrades.

V. Policy & Compliance Opportunities

1. Government Custom Solutions
   • Focus: Smart city management (traffic prediction), public sentiment analysis, automated governance.
   • Logic: Policy-driven demand with stable budgets (e.g., China’s “East Data West Computing” project).
2. Compliance & Auditing Services
   • Demand: EU AI Act mandates transparency for high-risk systems, creating compliance markets.
   • Example: IBM’s AI Ethics Toolkit for regulatory alignment.

VI. Future Battlegrounds & Risks

1. Key Barriers:
   • Data Quality: Synthetic vs. real-world data collection capabilities.
   • Energy Efficiency: Reducing training costs (e.g., Google’s Pathways cutting power use by 50%).
   • Multimodal Integration: Unified modeling of text, image, and video (e.g., GPT-5’s direction).
2. Risks:
   • Regulatory Scrutiny: Monopolistic control of foundational models may trigger antitrust actions.
   • Open-Source Disruption: Models like Llama 3 eroding profits of closed-source alternatives.
   • Efficiency Revolution: Smaller models (e.g., Mistral 7B) challenging the “bigger is better” paradigm.

Summary: Three-Tier Revenue Pyramid

The most sustainable profits will focus on the top application layer—transforming LLM capabilities into indispensable tools by solving industry-specific pain points, while building data moats and user loyalty.