The Corpus Royalty

A recent essay titled The Private Capture of Public Genius has been making the rounds, staging the current artificial intelligence boom as a battle between civic virtue and corporate extraction. The piece articulates a widespread anxiety: that the foundational models powering today's tech revolution were trained on the collective intellectual output of humanity, and that tech giants are effectively strip-mining the open internet to sell it back to us as a proprietary service.

The sentiment resonated instantly. In a massive Hacker News discussion debating the essay, readers were quick to demand structural remedies to this perceived enclosure. One highly upvoted commenter laid out the prevailing argument for financial redistribution, arguing that the true offense extends beyond copyright to the enclosure of the open web itself. Because foundational models rely heavily on decades of unpaid forum posts and articles to function, the sentiment suggests that any billions extracted from historical data inherently demand a systemic dividend to compensate for fencing off the digital commons.

This intuitive grievance is incredibly strong. When massive tech giants vacuum up the sum total of the web's historical output — from prestigious publisher archives to the collective troubleshooting of open-source forums — to build substitutive products, it feels deeply unfair to the creators who actually generated the initial ideas. The fundamental issue, though, is this: economic value in technology markets fundamentally derives from current structural scarcity rather than any historical moral desert.

The Price of Abundance

To understand why a massive, systemic data dividend is a mirage, we first have to ground the discussion in economic physics rather than moral philosophy or copyright law. The internet was built on a very specific financial primitive, which is that the marginal cost of duplicating digital text is effectively zero. Once a server is running and connected to the network, the literal financial expense required to transmit an additional blog post, a Reddit comment, or an entire digital newspaper archive is a fraction of a cent.

Consequently, the natural economic price of public text is exactly zero. You cannot extract massive, enduring rents from total abundance. If we were to explicitly map this dynamic, you could picture a Scarcity versus Abundance matrix. On the horizontal axis, we plot the supply of the raw resource; on the vertical axis, we plot the capital required to process and serve it. Raw text sits firmly in the bottom-right quadrant — infinite supply and near-zero marginal cost to reproduce. Every single day, the internet generates exabytes of new text automatically, constantly replenishing the reservoir.

In the top-left quadrant, however, sits the true bottleneck: immense clusters of specialized silicon and the highly scarce machine learning talent required to make sense of the noise. The demand for a mandatory licensing toll — a "data dividend" — attempts to pull the bottom-right quadrant into the top-left by sheer legal force, pretending that a ubiquitous commodity is actually a scarce asset.

The Integration Layer

This matrix directly dictates the market structure of artificial intelligence. If we render the AI value chain node by node as a pipeline, the flow of capital becomes immediately clear. Abundant scraped data flows into massive compute clusters, which are then refined by Reinforcement Learning from Human Feedback (RLHF) alongside specialized engineering talent, before finally being served as inference to the end user.

Crucially, we must follow the actual dollars to see where the margin pools. At the very beginning of the pipeline, the raw text is acquired at almost no cost. The data scraping layer is effectively a rounding error in the operational budget of a modern AI lab. But moving to the next node — the compute layer — the capital requirements become staggering. OpenAI famously spent more than $100 million training GPT-4. That figure only accounts for the raw computational power required for one initial training run. It does not include the ongoing cost of running the models, the billions spent on securing future energy contracts, or the specialized engineering talent required to build the model architectures.

Furthermore, the RLHF node requires paying domain experts actual hourly wages to evaluate outputs, creating a high-friction, capital-intensive loop that aligns the model to human preferences. Finally, the inference node demands continuous, massive compute resources just to generate a single response for the end user. The cash-conversion cycle for these AI labs is brutal: they must sink billions of dollars into unforgeable physical infrastructure years in advance, depreciate those assets over time, and aggressively monetize inference just to keep the servers powered.

This is the reality of the economics: the margin simply does not pool in the commoditized layers of raw data. The value is captured entirely by massive capital expenditure at the integration layer. Rather than deriving their value from having read a forum's archives, these models generate worth through the nine figures of hard capex required to metabolize the entire internet into a probabilistic weights matrix. The data is a commodity — the physical integration of silicon, power, and talent is the moat.

The False Monopoly Analogy

The most glaring error in the demand for a massive data dividend is its reliance on historical precedent — specifically, comparing the current AI giants to the telecom monopolies of the twentieth century. The argument assumes that because AT&T extracted vast rents from its infrastructure, today's model builders are similarly hoarding public resources and should be forced to pay a systemic toll to the creators who laid the foundation.

This history is exactly backward. The AT&T monopoly was entirely a function of government-enforced scarcity. Its dominance was underpinned by legally protected patents that quite literally prevented competitors from building rival networks. The government granted the monopoly, and in exchange, it eventually demanded concessions. We know this because the dam only broke in 1956, when a landmark consent decree forced AT&T to license all of its 7,820 existing patents royalty-free. That forced distribution of a legally hoarded asset birthed the modern semiconductor industry.

The structural dynamic of modern AI is the mirror image of that era. Instead of withholding legally protected secrets from the public, foundational models achieve dominance by applying unprecedented capital expenditure to a ubiquitous, freely available commodity. There is no government-enforced scarcity preventing a startup from scraping the exact same public websites as Google, Meta, or OpenAI. The barrier to entry firmly remains the sheer physical reality of needing 100,000 GPUs in a highly optimized data center, pulling gigawatts of power, rather than any reliance on a patent portfolio. AT&T captured value through legal exclusion; AI labs capture value through physical execution.

Structural Leverage

The desire to tax the model builders stems from a profound misunderstanding of where the bottleneck actually lives. In other words, leverage strictly resides at the integration of silicon and highly specialized human talent, completely separate from the commoditized scraping of public blogs or news archives. As I wrote when analyzing the broader AI market structure, the real constraint on AI is compute, not data.

To put it another way, publishers and forum posters are demanding a cut of the profits generated by the printing press, simply because they provided the ink. The ink is necessary, but the ink is everywhere. The machine that stamps the pages is the only thing that actually costs money to build, maintain, and operate. When you have a market where one input is infinitely abundant and the other input requires billions of dollars in specialized hardware, all economic returns will naturally accrue to the hardware.

This is why the piecemeal data licensing deals we have seen struck by publishers in recent months are largely transitional artifacts. They are public relations expenses and legal risk-mitigation strategies masquerading as sustainable business models. They do not reflect the underlying economic physics of the space, and they will likely evaporate once the legal precedents are firmly established.

This represents a fundamental misreading of market reality.

There is an understandable desire to bend the arc of technological progress toward a more equitable distribution of wealth, especially for the creative class and civic institutions that historically fueled the open web. The market, however, does not care about what is fair; it cares about what is scarce. If you build a regulatory framework based on the assumption that abundant public text holds inherent monetary leverage, you will be consistently surprised when the market prices it at exactly zero.

If I had to bet on the outcome, the market will eventually clear the illusion of a data royalty, simply because you cannot build a long-term business model on taxing a zero-marginal cost good. When exactly regulators and creators internalize that reality, though, is a different question: epochs like this often take a decade or longer to settle the underlying economics.