📱Ifanr (爱范儿)•Stalecollected in 57m
The Battle for AI Metrics: Token vs. DAA
#ai-metrics#business-strategy#kpi-definitionai-metrics-&-evaluation-standardsnvidiabaidujensen huangrobin li
💡Understand the shift from Token-based metrics to Agent-based KPIs to better align your AI strategy with industry trends.
⚡ 30-Second TL;DR
What Changed
Jensen Huang advocates for 'Token economy' as the core metric for AI productivity.
Why It Matters
This debate signals a shift in how investors and companies evaluate AI ROI, moving from infrastructure capacity to actual user engagement and agentic utility.
What To Do Next
Evaluate your AI product's success by tracking agent-based task completion rates rather than just raw token consumption.
Who should care:Founders & Product Leaders
Key Points
- •Jensen Huang advocates for 'Token economy' as the core metric for AI productivity.
- •Robin Li proposes 'DAA' (Daily Active Agents) as a more practical measure of AI utility.
- •The industry is moving away from purely technical benchmarks toward value-based metrics.
🧠 Deep Insight
Web-grounded analysis with 18 cited sources.
🔑 Enhanced Key Takeaways
- •Jensen Huang's 'Token economy' frames tokens as a commodity and the fundamental unit of value, pricing, and competition in the AI industry, envisioning data centers as 'AI factories' that produce tokens at scale.
- •Robin Li's 'Daily Active Agents' (DAA) is introduced as the AI era's equivalent of Daily Active Users (DAU), focusing on measuring the number of AI agents actively working for humans and delivering tangible results, thereby shifting the emphasis from token consumption as a cost to agent-driven task completion and outcome delivery.
- •The industry is undergoing a significant paradigm shift, moving away from traditional technical benchmarks like accuracy and precision, and even input-based metrics such as token consumption, towards value-based metrics that prioritize real-world applicability, business impact, and the actual delivery of results.
- •NVIDIA's 'Token economy' concept includes a proposed tiered pricing structure for token delivery and highlights 'token efficiency' (e.g., tokens per watt, per dollar) as a crucial operating metric for enterprises and AI service providers.
- •Baidu's DAA proposal is intrinsically linked to the emergence of 'super individuals' augmented by AI agents and a transformation in organizational structures towards human-agent collaborations, where AI agents evolve from passive responders to autonomous executors capable of continuous learning.
📊 Competitor Analysis▸ Show
| Feature/Metric | Jensen Huang's 'Token Economy' | Robin Li's 'Daily Active Agents' (DAA) |
|---|---|---|
| Advocate | Jensen Huang (NVIDIA CEO) | Robin Li (Baidu CEO) |
| Primary Focus | Monetizing inference at scale; AI compute as a commodity. | Measuring AI utility and value creation through task completion. |
| What it Measures | Cost per token, throughput (tokens per second/megawatt), token efficiency, and revenue generated from token production. | Number of active AI agents completing 'task loops' and delivering results for humans. |
| Underlying Philosophy | AI data centers are 'factories that manufacture intelligence,' with tokens as the unit of production and cost. | Tokens are a cost (input), not revenue (output); true value lies in agents actively working and delivering tangible outcomes. |
| Implication for AI Development/Business | Drives optimization of hardware and software for maximum token throughput and efficiency; encourages tiered pricing models for AI services. | Shifts focus from model intelligence to application effectiveness and agent capabilities; necessitates 'agent-native' infrastructure and new business models based on agent utility. |
🛠️ Technical Deep Dive
- Tokens: These are the most basic units of output from large language models, with approximately 1,300 tokens typically generating 1,000 words of text. They serve as the fundamental computational unit for AI models to process and generate data. NVIDIA's Blackwell architecture, for instance, demonstrates significant improvements in token generation, achieving up to 65 times more tokens per second per GPU and around 50 times more tokens per megawatt compared to previous architectures, leading to approximately 35 times lower cost per million tokens. The concept emphasizes 'token efficiency' (e.g., tokens per watt, per dollar) as a critical operating metric, viewing data centers as 'AI factories' dedicated to producing tokens.
- Daily Active Agents (DAA): This metric is tied to the performance and utility of 'AI agents,' which are described as general-purpose AI entities designed for task completion, offering a higher value ceiling than traditional chatbots. To support the widespread adoption and functionality of these agents, Baidu is actively developing an 'agent-native' infrastructure. This infrastructure spans various layers, including proprietary Kunlunxin AI chips, cloud computing services, and advanced large models like ERNIE 5.1, all optimized for large-scale agent applications. Examples of such task-completing agents include Baidu's DuMate (a general-purpose agent), Miaoda (a coding agent), and Famou Agent 2.0 (a self-evolving agent for optimization tasks). DAA specifically measures how many of these agents successfully complete a 'task loop' in real-world scenarios on a daily basis.
🔮 Future ImplicationsAI analysis grounded in cited sources
Outcome-based metrics will reshape AI investment priorities.
Companies will increasingly prioritize AI projects that demonstrate tangible business value and measurable results, such as revenue growth or cost reduction, over those focused solely on technical performance or input consumption, driving a re-evaluation of AI development strategies.
The proliferation of AI agents will necessitate specialized infrastructure and new business models.
As AI evolves from models to task-completing agents, platforms will need to support large-scale agent applications and measure their active utility, leading to a focus on 'agent-native' infrastructure and service-oriented business models.
The 'Token economy' will lead to a commoditization and tiered pricing of AI compute resources.
Nvidia's advocacy for tiered token pricing and emphasis on token efficiency suggests that AI compute will be increasingly treated as a utility, priced based on throughput and speed, making cost-per-token a critical competitive and operational factor.
⏳ Timeline
2023
Significant decrease in price-per-token and rapid scaling of token volume.
2025-02
Industry discussions on the evolution of AI metrics, emphasizing ROI, time to value, and continuous improvement.
2025-09
MIT report highlights that 95% of generative AI projects fail to deliver measurable ROI, sparking debate on appropriate AI metrics.
2026-01
Jensen Huang at CES 2026 discusses resetting the economics of AI factories, emphasizing system-level performance and token demand.
2026-03
Jensen Huang at GTC promotes 'Token economy' as the core unit of value for AI, positioning data centers as 'AI factories' and proposing tiered pricing for tokens.
2026-05-13
Robin Li at Baidu Create 2026 proposes 'Daily Active Agents' (DAA) as a new metric for the AI industry, focusing on agent-driven task completion over token consumption.
📎 Sources (18)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: Ifanr (爱范儿) ↗