AI’s hidden economy: The emergence of cosine capital in the age of LLMs

The research details how embeddings operate as “general intellect” - a collective, machine-readable form of human thought. Every text, image, or action that enters an LLM becomes part of a recursive process: embeddings create content, content creates more embeddings, and the cycle continues. represents this as an E–C–E′ loop, where each round of embedding and generation compresses the boundaries between human and machine knowledge.

A new study examines how artificial intelligence systems, particularly large language models (LLMs), are reshaping global capitalism by embedding all aspects of human communication into mathematical relationships.

The research, titled “Cosine Capital: Large Language Models and the Embedding of All Things,” published in Big Data & Society, introduces the concept of “cosine capital” to describe how meaning, value, and labor are redefined through the computational logic of AI embeddings.

The emergence of cosine capital

argues that LLMs such as ChatGPT and similar systems mark a major turning point in the organization of digital economies. At their core, these models use cosine similarity to measure the closeness of vectors, mathematical representations of words, objects, or ideas, in a high-dimensional space. This geometry of proximity, according to the study, is not just a technical mechanism but an economic one. It turns meaning and language into measurable and tradable entities.

In this new paradigm, “cosine capital” refers to a system in which social relations, cultural production, and intellectual labor are embedded as coordinates in machine-learning models. These embeddings become forms of capital when they are owned, circulated, and monetized by the institutions that control the models. Unlike traditional data-driven economies that rely on discrete datasets, this emerging model depends on relational similarity, how closely one concept aligns with another in vector space.

The study identifies this as a historical shift from the “bit economy”, based on digital records and classification, to an “embedding economy”, where everything from language to emotion becomes a latent representation within AI systems. This marks the rise of a new infrastructure of power that operates invisibly but continuously, shaping how meaning itself is produced and distributed.

From labor to embedding: A new logic of value

Traditional economic systems extract value from labor and production, but in the age of LLMs, value emerges from contextual coherence, how well something fits within a model’s embedded structure. The author connects this transformation to Marxian economic theory, describing LLMs as machines that perform abstraction at the scale of human language, converting intellectual and emotional work into predictable and reproducible relationships.

The research details how embeddings operate as “general intellect” - a collective, machine-readable form of human thought. Every text, image, or action that enters an LLM becomes part of a recursive process: embeddings create content, content creates more embeddings, and the cycle continues. represents this as an E–C–E′ loop, where each round of embedding and generation compresses the boundaries between human and machine knowledge.

As these embeddings multiply, they begin to structure the social world. The author warns that this process leads to what he calls epistemic enclosure - a situation where knowledge, culture, and even identity are captured within privately owned AI architectures. Meaning becomes measurable and controllable, with algorithmic systems determining what counts as relevant, coherent, or valuable.

The author compares this with earlier forms of algorithmic capitalism based on eigencapital - a concept describing systems that classify and rank people through named attributes and tables, such as credit scores or customer databases. Cosine capital, in contrast, deals not in fixed attributes but in fluid relations. It operates through statistical proximity rather than explicit categorization, creating a subtler but more pervasive form of control. It no longer matters what a person is labeled; what matters is how closely they align with other nodes in the model’s latent space.

The politics of embedding and the future of AI economies

According to the paper, cosine capital is a new stage of computational capitalism, one that extends deep into the geometry of human thought. argues that control over embedding systems, those that define how concepts relate and which associations count, has become the new frontier of power. The companies that own and train large language models effectively govern the infrastructure of meaning, determining how information flows and whose knowledge becomes visible.

This raises urgent political and ethical questions. If embeddings determine what can be said, generated, or retrieved, then access to meaning becomes a proprietary service. AI models, optimized for alignment and similarity, may unintentionally flatten cultural diversity by privileging average or majority associations. In ’s view, this results in a narrowing of imagination - a recursive reinforcement of what the model already “knows.”

Moreover, the paper highlights the feedback mechanism inherent in AI training cycles. As models learn from their own outputs, the boundaries of embedded knowledge contract, leading to increasing homogeneity. To counteract this tendency, developers continually expand data collection, integrating more content from human communication, which further extends AI’s reach into personal and collective life.

The author warns that this expansionist logic of cosine capital blurs the line between representation and reality. It embeds human creativity, emotion, and cognition into machine-readable patterns that circulate as commodities. The study suggests that the social impact of this process is comparable to the historical rise of industrial capitalism, except that today’s raw material is not physical labor, but meaning itself.