AI wealth concentration and the new battle for enterprise value
AI wealth concentration is the growing risk that a small number of dominant AI companies capture most of the economic gains from artificial intelligence, even though many other firms supply the data, expertise, and real-world work that make those gains possible. Satya Nadella warns that advanced models are becoming so good at absorbing corporate know-how that the firms generating this knowledge could lose control of it. Instead of AI acting as an equalizer, value could flow upwards to a few AI platforms, leaving traditional businesses as data suppliers with shrinking margins. This dynamic mirrors earlier technology and outsourcing waves that boosted aggregate productivity but hollowed out industries. If enterprises treat AI as a commodity tool rather than a strategic asset, they risk donating their institutional memory and competitive secrets to external models that “eat everything they see” and return only generic insights.
Token capital and human capital: a new form of corporate IP
Nadella argues that the core of a modern enterprise AI strategy is the loop between human capital and token capital business assets. Human capital is the knowledge, judgment, relationships, ingenuity, and pattern recognition held by people. Token capital is the proprietary AI capability a firm builds and owns, often encoded in models, prompts, evaluations, and internal datasets. According to Microsoft CEO Satya Nadella, “the real opportunity is not in picking the best model but instead in building a learning loop on top of models where human capital and token capital compound.” In this view, human expertise and AI capability are not rivals. Human agency sets goals, spots meaningful patterns, and connects domains, while token capital remembers, scales, and repeats successful decisions. Together they become a new type of intellectual property: a living, adaptive system that keeps learning from every customer interaction, project, and internal experiment.
Why AI learning loops, not models, are the real competitive edge
As foundational models get commoditized, Nadella says the durable value sits in AI learning loops built on top of them. These loops combine human decisions, enterprise data, and model behavior in a compounding cycle: people delegate tasks to AI agents, steer outcomes at key moments, then feed the results back into the system. Private evaluation pipelines check performance against real business metrics, not public benchmarks. Private reinforcement learning environments train models on the company’s own activity traces. A governed knowledge base makes institutional memory searchable and keeps token usage efficient. The result is what he calls a “hill climbing machine” that gets smarter with every use. Critically, companies should be able to swap a generalist model without losing this “company veteran” intelligence. If the learning loop is portable, the firm controls its token capital; if not, it has rented intelligence and ceded value to external platforms.
Monopoly risks: from globalization to AI wealth concentration
Nadella draws a stark parallel between AI wealth concentration and the first wave of globalization. Then, outsourcing raised GDP on paper while eroding industrial bases, jobs, and local expertise. Today, a similar pattern could unfold in digital form: industries push their best practices into external models, those models centralize expertise across sectors, and economic returns concentrate in a handful of AI leaders. He warns that there is “no societal permission for an AI future that hollows out entire industries.” The stakes are higher than traditional tech monopoly fears because AI can absorb and replicate the judgment of whole professions, not only their tools. If every sector feeds its knowledge into a few frontier models, competitive differentiation shrinks, bargaining power shifts, and inequality widens as value flows to platforms rather than producers of real-world goods and services.
How enterprises can build sovereign AI advantage now
To avoid becoming mere data suppliers, companies need an enterprise AI strategy centered on owned AI learning loops. First, they should identify critical workflows where human capital is strongest—customer relationships, domain judgment, operational know-how—and design agentic systems that learn from these interactions rather than replace them. Second, they must keep key components private: evaluation suites, reinforcement learning environments, and knowledge graphs built from internal data. Third, they should insist on model portability so they can change providers without losing accumulated intelligence. Nadella’s alternative vision is a “frontier ecosystem” where platforms enable more value than they capture, and each organization owns the loop that encodes its institutional memory. For leaders, the choice is immediate: treat AI as a procurement line item, or start building token capital that compounds with human capital into a lasting competitive moat.
