Microsoft CEO Nadella Proposes 'Human Capital' and 'Token Capital': Enterprises Must Build AI Learning Loops to Avoid Value Capture by a Few Models

Microsoft CEO Satya Nadella published a long post on X in June 2026 titled 'A frontier without an ecosystem is not stable,' proposing that in an AI-driven economy, enterprises must simultaneously build 'human capital' (employee knowledge, judgment, etc.) and 'token capital' (AI capabilities owned by the enterprise) and establish a learning loop between them to achieve compound growth. Nadella warned that if all industry value is captured by a few AI models, it would repeat the mistakes of the first phase of globalization's industrial outsourcing, leading to economic imbalance. He emphasized that enterprises should build private evaluations, reinforcement learning environments, and knowledge bases to ensure models are replaceable while their own intellectual property is not lost. The post garnered over 28 million views, with Elon Musk commenting 'Interesting.'

Core Concepts: Human Capital and Token Capital

• Human Capital: Includes employee knowledge, judgment, relationships, creativity, and pattern recognition. Nadella believes human capital will not depreciate with the growth of token capital but will become more valuable.
• Token Capital: AI capabilities built and owned by the enterprise, referring to autonomously controlled AI compute and capability reserves. The two must form a 'cognitive loop' where AI models continuously absorb human expertise and provide feedback, rather than being used merely as tools.

Enterprise Strategy: Building a Learning Loop

• Nadella proposed that enterprises should establish an architecture where agent systems evolve over time while maintaining control over intellectual property. Key components include:
  • Private Evaluation: Measuring model improvements on actual business outcomes, not relying on external benchmarks.
  • Private Reinforcement Learning Environment: Optimizing models based on real execution trajectories within the organization.
  • Knowledge Base: Making institutional memory queryable to improve token usage efficiency.
• This loop is considered a new form of intellectual property with compounding effects: improved workflows generate better training signals, accelerating tacit knowledge accumulation.

Industry Background and Data

• Global AI investment surged: $202.3 billion in 2025, accounting for 50% of global venture capital; OpenAI and Anthropic alone accounted for 14% of global AI venture capital.
• High enterprise AI deployment but low maturity: McKinsey data shows 72% of enterprises have deployed at least one AI workload, but only 1% consider themselves mature.
• Employment impact: The World Economic Forum predicts 850,000 jobs will be replaced by AI by 2028; employment rates for workers aged 22-25 in high AI-exposure roles have dropped by 13%.
• Model competition shifting to agent competition: Gartner predicts that by the end of 2026, 40% of enterprise applications will embed task-oriented AI agents.

Political Economy Perspective: Avoiding Value Concentration

• Nadella warned that if value is captured by a few models, the political-economic system will not tolerate it, and society will not allow an AI future that 'hollows out entire industries.' He drew an analogy to the consequences of the first phase of globalization's industrial outsourcing.
• Currently, China and the US control 90% of frontier AI compute, and the San Francisco Bay Area accounts for 76% of global AI investment. The sovereign AI movement is rising, but national investments are far below Microsoft's single-quarter capital expenditure ($37.5 billion in Q2 FY2026).
• Nadella called for building a 'frontier ecosystem' rather than a single frontier model, allowing value to flow broadly to every company, industry, and country.

Microsoft's Own Practice

• Microsoft has shifted from relying on OpenAI to a multi-model strategy: investing over $13 billion in OpenAI since 2019, but after a 2025 restructuring, the IP license became non-exclusive.
• At the 2026 Build conference, Microsoft released seven self-developed MAI models covering reasoning, coding, and other areas, claiming cost efficiency up to 10 times higher than GPT-5 in specific scenarios.
• Microsoft Foundry (formerly Azure AI Foundry) supports multi-model deployment, positioning itself as an AI application and agent factory, reflecting the philosophy that 'platforms unlock more value.'