This week, OpenAI CEO Sam Altman presented an analogy at the BlackRock U.S. Infrastructure Summit: "The future we see is intelligence becoming a utility, like electricity or water, where people buy it from us by quantity."
This concept itself is not new. The notion of "AI as a utility" dates back at least a decade. This time, however, Altman's statement included a clear subject and direction: "buy it from us by quantity." Specifically, it means purchasing intelligence from OpenAI based on token usage.
Shortly after his remarks, a wave of criticism gathered on platforms like Reddit and X. A widely shared comment read: "They used our lives and creativity to feed the model, trampling copyright laws, and now they want to sell these things back to us in the form of a utility."
Proposing a grand narrative for the capital markets while igniting an ethical inquiry from the creator community. This article does not judge the speaker's motives nor predict the outcome of legal battles. The core point of interest is: does this "utility" metaphor hold up logically, ethically, and commercially? Deconstructing this metaphor can help us see the deep-seated contradictions unfolding within the AI industry.
Deconstructing the Narrative: Why "Utility"?
To understand the intent behind this metaphor, one must return to the context of Altman's statement.
According to reports from Business Insider and transcripts from Rev, Altman's remarks did not stem from a product launch or a technical roadmap, but rather from a warning about a "compute bottleneck." At the summit, he clearly stated that without building sufficient computing infrastructure now, three potential outcomes could emerge in the future: AI services could become scarce and prices could skyrocket, only the wealthy could afford them, or governments might have to intervene in allocation.
In other words, the "utility" metaphor is primarily a narrative aimed at infrastructure investors, not a pricing plan for end-users.
Packaging AI as water or electricity has clear business logic. Utilities like water and electricity are capital-intensive, long-cycle, and cash-flow-stable industries, naturally suited to the capital structures of pension and infrastructure funds. When OpenAI needs to convince asset management giants like BlackRock to fund data center projects worth trillions of dollars, "AI as a utility" is more likely to pass investment committee approvals than "AI as a tech product."
This assessment is not speculation. OpenAI President Greg Brockman mentioned that the company needs approximately $1.4 trillion in data center investment commitments over the next eight years. While the specific structure and progress of this figure require verification, it is sufficient to indicate: the primary audience for Altman's "utility" is the capital market, not end-users.
"Incremental Construction" vs. "Reorganization of Existing Stock"
Critics' anger focuses on a fundamental difference obscured by the "utility" metaphor.
Water and electricity are forms of "incremental construction." Humans build dams, lay pipelines, and erect power grids to create supply capabilities that did not exist in nature. Investments are made to build new physical assets that themselves are not dependent on the prior labor of others.
AI model training is a "reorganization of existing stock." GPT series models are trained on data from large-scale web crawling of publicly available content, covering books, articles, artworks, forum posts, code repositories, and even users' private conversation records on social media. This represents decades of accumulated human creation, the vast majority obtained without creator authorization or payment of any copyright fees.
A Medium author wrote: "They are trying to compress decades of collective human creation into a commodity, then re-price it under the name of a utility, selling it back token by token to the very people who provided the raw materials for free."
This is not emotional venting but a precise identification of a property rights logic. The "raw materials" for utility companies like water and electricity are either self-built (dams storing water) or purchased at market prices (coal, gas). In contrast, the "raw materials" acquired by AI companies during the training phase reside in the legal gray area of "Fair Use," and commercially, have incurred no cost transfer.
This model of "free acquisition, paid sale" makes the "utility" in critics' eyes sound more like an "enclosure movement": first appropriating resources from the public domain, building walls around them, and then charging the original users for entry.
The Distance Between Token Pricing and Universal Service
Even setting aside the data source controversy, the "AI as a utility" metaphor struggles to hold up in terms of pricing mechanisms.
True utilities, such as water, electricity, and gas, in most economies, bear the obligation of "Universal Service." Regulatory agencies require them to guarantee basic supply for public welfare. Pricing mechanisms are typically based on "cost-plus," with profit margins strictly constrained. The price of residential electricity does not differ based on whether you use it to light a bulb or run a server.
AI's token pricing is entirely different. According to KongHQ's monitoring of enterprise AI costs and analysis by Artefact, while the absolute price per-token has decreased by about 75% over the past year, actual enterprise AI spending has increased rather than decreased, as usage growth has far outpaced price reductions. This pattern of "falling unit price, rising total cost" is referred to as the "Token cost illusion."
More telling is the structural difference in token fees. The price of an output token is typically 3 to 10 times that of an input token. For the same amount of information, the cost for AI to "read in" is much lower than to "write out." If you submit a document to an AI for summarization, the input stage is almost free, but every word generated in the summary falls into a high-charge zone.
The logic of the public power grid's pricing is: electricity itself is homogeneous; 1 kilowatt-hour costs the same for a refrigerator as for a server. The logic of AI's token pricing is: the service itself is split into vastly different price tiers, and this price differential is defined unilaterally by the supplier.
In other words, this is not utility pricing; this is discriminatory pricing based on usage volume. Its goal is not to ensure everyone can access intelligence, but to extract maximum revenue from the consumption of intelligence.
The "Fair Use" Moat is Beginning to Crumble
Although critics are loud, legally, AI companies are not in as precarious a position regarding training data as it might seem on the surface.
According to the Morrison & Foerster law firm's "2026 AI Trends" report and Norton Rose Fulbright's tracking of AI copyright litigation, US courts currently tend to hold that training general-purpose AI models has a "highly transformative" nature, making it easier to meet the statutory standards for "Fair Use." The Anthropic case in mid-2025, where they successfully persuaded a court to dismiss a copyright lawsuit (though details need verification), has become a significant source of confidence for the AI industry.
However, this legal moat is being eroded by the AI industry's own actions in terms of business logic.
An analysis by TechPolicy.press points out: As AI companies begin to purchase licensed training data on a large scale—such as OpenAI's agreements with Reddit, News Corp, and others—the defense of "free crawling constitutes Fair Use" is being paradoxically weakened. If training data can indeed be "fairly used" indiscriminately, why spend large sums to purchase licenses for specific sources? If data owners indeed have no right to make claims, what is the legal basis for these licensing agreements?
The act of purchasing itself constitutes a commercial negation of the "raw materials are free" premise.
Returning to Altman's "water and electricity theory," this contradiction becomes more acute. Water and electricity companies, when building infrastructure, do not face collective questioning like "is your source of water legally obtained?" But when AI companies proclaim themselves as the next-generation utility, the question "where does the raw material come from?" still lacks a convincing answer.
Becoming Infrastructure Requires Solving Distribution Issues
Altman's "water and electricity theory" captures a real trend in AI development. Large models are transitioning from lab products into foundational capabilities, embedded into search engines, office software, design tools, and even industrial processes. When AI becomes ubiquitous, it does functionally approach "infrastructure."
But three fractures in this metaphor at the current stage of evolution cannot be ignored.
First, the property rights fracture. Water and electricity create increments; AI reorganizes existing stock. Reorganization itself has value, but its premise—"existing stock can be used for free"—has neither achieved moral consensus nor received final legal confirmation.
Second, the pricing fracture. The "universal service" of utilities implies low profit margins and non-discriminatory pricing, while token pricing is market-based, tiered, and unilaterally defined by the supplier. The two have almost no intersection in business logic.
Third, the governance fracture. The utility industry has independent regulatory bodies, transparent cost accounting, and public price hearing mechanisms. The AI industry currently lacks any form of public governance framework; the rules for "pay-per-use" are set by a handful of companies themselves.
For ordinary users, the trend of AI pay-per-use is unlikely to change in the short term. The benefit of falling token prices continues, but the "ever-increasing usage" will offset this advantage. When choosing AI tools, it is recommended to not only focus on unit price but also assess your own actual usage trend.
For developers and enterprise clients, the cost controllability of high-token-consumption scenarios like code generation and long-text analysis is more critical than unit price. Relying on a single supplier's token pricing system means your cost structure is completely at their mercy.
For creators, the spread of the "AI utility" narrative itself is a signal: the probability of your work being used for training is increasing, while a mechanism for receiving compensation has yet to emerge. The infrastructuralization of the industry should not merely turn model companies into the next power companies; it should also include establishing a reasonable, traceable mechanism for distributing the benefits derived from data.
The current reality is: AI is becoming infrastructure, but it has not yet become a utility. The latter title requires more to support it, not just compute scale and token-based billing.
