2026向代币经济学战略性转向_驾驭AI支出新格局研究报告_28页_14mb

Deloitte AI Tokenomics Summary

Core Content

This document explores the evolving economics of AI, focusing on token-based pricing and its implications for enterprise cost management. It highlights how traditional cost models are inadequate for capturing the complexities of AI spend, which is driven by token consumption. The paper emphasizes the need for precision economics, where token costs are tracked, predicted, and optimized to align with business value.

Main Points

1. AI Spend Dynamics

• Rapid Growth: AI has become the fastest-growing line item in corporate technology budgets, consuming up to half of IT spend in some firms.
• Cloud Cost Increases: Cloud bills are rising by nearly 20% annually due to AI workloads.
• Geopolitical Pressures: Organizations are increasingly concerned with data sovereignty and infrastructure independence.
• Token as Cost Unit: Tokens—small data chunks processed by AI systems—serve as the practical unit of cost, and their consumption drives AI spend.

2. Token Economics Overview

• Nonlinear and Unpredictable Costs: AI costs scale in complex and unpredictable ways based on workload design, algorithmic complexity, and infrastructure intensity.
• Token-Based Pricing: Token pricing is tied to the actual work AI performs, and it is influenced by the tech stack, hosting model, and AI model customization.
• Three Pricing Models:
  • Packaged Software: Tokens are abstracted, with costs hidden in vendor contracts.
  • API Consumers: Token use is explicit, but costs are volatile due to workload design and infrastructure provider choices.
  • Self-Hosted AI Factory: Tokens are a direct outcome of infrastructure decisions, offering the most control but requiring significant capital and technical investment.

3. AI Factory Concept

• Definition: An AI factory is a self-hosted infrastructure solution that includes compute, storage, and networking, along with optimized software and services.
• Key Advantages:
  • Greater control over cost, latency, and data sovereignty.
  • Better long-term cost predictability and management.
• Best Use Cases:
  • High-volume, predictable, and latency-sensitive workloads.
  • Scaled, high-value applications where infrastructure investment pays off.

4. Hosting Models and Their Economics

5. AI Model Selection

• Open-Source Models:
  • Generally free.
  • Run in self-hosted environments.
  • Offer greater control, customization, and data sovereignty.
  • Examples: Meta Llama, Mistral, NVIDIA NIM Microservices.
• Proprietary Models:
  • Typically billed per token.
  • Enable quick deployment with no upfront investment.
  • Pretrained and have strong out-of-the-box functionality.
  • Examples: Anthropic Claude, Google Gemini, OpenAI GPTs.
  • Higher per-token costs and less flexibility.

6. Token Cost Curve and Jevons' Paradox

• Efficiency vs. Consumption: As AI efficiency improves, token consumption increases, leading to higher total costs.
• Token Price Trends: Token prices are dropping rapidly, from dollars per thousand to pennies per million.
• TCO Projections: Deloitte projects the average inference cost will drop from $0.04 per million tokens in 2025 to about $0.01 by 2030.

Key Takeaways

• Token Awareness is Critical: Understanding token dynamics is essential for managing AI costs effectively.
• Hybrid Strategies are Common: Most enterprises adopt hybrid models, balancing API access for smaller workloads with self-hosted AI factories for larger, more complex needs.
• Cost Inflection Point: At around 7 billion tokens per month (or 84 billion per year), AI factories become more cost-effective than API or neo-cloud providers.
• Compute Dominates Costs: At 10 billion tokens, compute costs make up the majority of TCO, with facilities, software, and networking costs following closely.
• Strategic Flexibility is Key: Enterprises should design flexible, modular, and hybrid architectures to adapt to evolving AI economics.

Strategic Recommendations

• Know Your Workloads: Understand and prioritize AI workloads to make informed infrastructure decisions.
• Assess Consumption Scale: Both current and future token demand will shape hosting strategy and AI model choice.
• Avoid Vendor Lock-In: Opt for open-source models or modular architectures to maintain flexibility and control.
• Plan for Future Costs: Develop forward-looking infrastructure strategies that account for rapid technological advancements and evolving cost structures.

Conclusion

The economics of AI are shifting from traditional TCO models to token-based precision economics. As AI adoption grows, so does the complexity of managing its costs. Organizations must navigate this new landscape by understanding token dynamics, evaluating hosting options, and strategically selecting AI models. The decision to build an AI factory or continue with cloud or API-based solutions depends on workload scale, predictability, and business priorities. With careful planning and infrastructure optimization, enterprises can turn token spend into a strategic advantage.