This week, OpenAI CEO Sam Altman offered an analogy at the BlackRock U.S. Infrastructure Summit: "The future we see is where intelligence will become a utility like electricity or water, and people will buy it from us based on volume."

This statement itself is not a new concept. The idea of ​​"AI as a utility" can be traced back at least a decade. But this time, Altman's statement has a clear subject and direction: "Buy from us based on volume." Specifically, it means paying by token to purchase intelligence from OpenAI.

No sooner had the words left his mouth than a torrent of criticism erupted on platforms like Reddit and X. One widely shared comment read: “They fed our lives and creativity to these models, trampled on copyright laws, and now they want to sell them back to us as utilities.”

This article presents a grand narrative geared towards the capital market, igniting an ethical debate within the creator community. It does not judge the speaker's motives, nor predict the course of legal proceedings. The core interest lies in whether this "utility" metaphor is logically, ethically, and commercially sound. Deconstructing this metaphor can help us understand the deep-seated contradictions unfolding within the AI ​​industry.

Narrative Deconstruction: Why "Public Utility"?

To understand the intention behind this metaphor, we need to return to the context of Altman's statement.

According to Business Insider and Rev's transcript of the meeting, Altman's statement was not about product launches or technological roadmaps, but rather a warning about a "computing bottleneck." He explicitly stated at the summit that if sufficient computing power infrastructure is not built now, three outcomes are possible in the future: AI services will be in short supply leading to soaring prices, only the wealthy will be able to afford them, or governments will have to intervene in allocation.

In other words, the "utility" metaphor is primarily a narrative geared towards infrastructure investors, not towards user-oriented pricing strategies.

There's a clear business logic to packaging AI as a utility. Utilities are capital-intensive, long-cycle, and have stable cash flow, making them naturally suitable for the capital structures of pension funds and infrastructure funds. When OpenAI needed to convince asset management giants like BlackRock to fund multi-billion dollar data center projects, "AI as a utility" was more likely to pass investment committee approval than "AI as a technology product."

This assessment is not speculation. OpenAI President Greg Brockman has mentioned that the company will need approximately $1.4 trillion in data center investment commitments over the next eight years. While the specific structure and implementation progress of this figure remain to be verified, it is sufficient to illustrate that the "utilities" Altman refers to primarily target the capital markets, rather than end users.

"Incremental construction" or "restructuring of existing assets"?

Critics' anger focuses on a fundamental difference that the "utility" metaphor obscures.

Hydropower is "incremental construction." When humans build dams, lay pipelines, and erect power grids, they are creating supply capacity that did not originally exist in nature. The investment is used to build new physical assets that are not dependent on the existing labor of anyone else.

AI model training is a process of "reorganizing existing resources." The training data for the GPT series models comes from large-scale crawling of publicly available content across the entire internet, covering books, articles, artwork, forum posts, code repositories, and even users' private conversations on social media. This represents decades of accumulated human creation, the vast majority of which is unauthorized by the creators and without any copyright fees paid.

One Medium author wrote: "They are trying to compress decades of collective human creation into a single commodity, then reprice it as a utility and sell it back to the people who provided the raw materials for free in the first place."

This is not an emotional outburst, but a precise identification of the logic of property rights. Utility companies like hydroelectricity either build their own "raw materials" (dams to store water) or purchase them at market prices (coal and gas). In contrast, the "raw materials" acquired by AI companies during the training phase fall into a legal gray area of ​​"fair use," and commercially do not result in any cost transfer.

This "free access, paid sale" model makes what critics see as "public utilities" sound more like a "land grab": first appropriating public resources, building walls, and then charging original users an entry fee.

The Distance Between Token Billing and Universal Service

Even setting aside the controversy over the source of the data, the notion that "AI is a utility" is difficult to justify in terms of pricing mechanisms.

True public utilities, such as water, electricity, and gas, are obligated to provide universal service in most economies. Government regulators require them to guarantee the supply of basic necessities, and pricing mechanisms are typically based on cost-plus pricing with strictly regulated profit margins. Residential electricity prices are not differentiated based on whether you use it to turn on a light bulb or run a server.

The pricing of AI tokens is entirely different. According to KongHQ's monitoring data on enterprise AI costs and Artefact's analysis, the absolute price of per-tokens has decreased by about 75% in the past year, but enterprises' actual AI spending has increased rather than decreased because the growth rate of usage far exceeds the price decline. This "lower unit price, higher total price" model is known as the "token cost illusion".

Even more telling is the structural difference in token fees. Output tokens are typically 3 to 10 times more expensive than input tokens. For the same amount of information, the cost for AI to "read" it is far lower than the cost to "write" it out. If you submit a document to AI for summarization, the input stage is almost free, but every word generated in the summary becomes a high-cost area.

The pricing logic of a public power grid is that electricity itself is homogeneous; one kilowatt-hour of electricity costs the same whether it's used to power a refrigerator or a server. The pricing logic of AI tokens is that the service itself is broken down into huge price differences, and these price differences are entirely defined unilaterally by the supplier.

In other words, this isn't utility pricing; it's discriminatory pricing based on usage. It's not about making smart technology accessible to everyone, but about maximizing revenue from the amount of smart technology consumed.

The moat of "reasonable use" is loosening.

Despite the loud criticism, from a legal perspective, AI companies are not as vulnerable as they appear when it comes to training data.

According to Morrison & Foerster's "AI Trends 2026" report and Norton Rose Fulbright's tracking of AI copyright litigation, US courts are currently inclined to recognize that training general-purpose AI models is "highly transformative" and therefore more likely to meet the statutory standard of "fair use." Anthropic's successful persuasion of a court to dismiss a copyright lawsuit in mid-2025, while details still pending verification, has become an important source of confidence in the AI ​​industry.

However, the legal moat is being eroded by the AI ​​industry's own behavior in terms of business logic.

An analysis by TechPolicy.press points out that as AI companies begin to purchase licensed training data on a large scale—for example, the agreements OpenAI reached with Reddit, News Corp, and others—the defense of "free scraping equals fair use" is being contradictorily weakened. If training data can indeed be used "fairly" without discrimination, then why spend a lot of money to purchase licenses from specific sources? If data owners truly have no rights to claim, then what is the legal basis for these licensing agreements?

The act of purchasing itself constitutes a commercial negation of the presupposition of "free raw materials".

Returning to Altman's "hydropower theory," this contradiction becomes even more acute. Hydropower companies, when building infrastructure, don't face the collective question of "whether your water source is legal." Yet, when AI companies claim to be the next generation of utilities, the question of "where do the raw materials come from?" remains unconvincing.

The infrastructure development process needs to address the issue of allocation.

Altman's "hydroelectricity theory" captures a real trend in AI development. Large-scale models are transforming from laboratory products into underlying capabilities, being embedded in search engines, office software, design tools, and even industrial processes. When AI is ubiquitous, it truly approaches "infrastructure" in function.

However, the three cracks in this metaphor's current stage of evolution cannot be ignored.

First, there's the issue of property rights. Hydropower creates new value, while AI restructures existing assets. Restructuring itself has value, but the premise is that "existing assets can be used free of charge." This premise has neither gained moral consensus nor received final legal confirmation.

Second, the pricing gap. "Universal service" in utilities implies low profit margins and non-discriminatory pricing, while token pricing is market-driven, tiered, and unilaterally defined by the supplier. The two have virtually no overlap in their business logic.

Third, addressing the shortcomings. The hydropower industry has independent regulatory bodies, transparent cost accounting, and public participation mechanisms for price hearings. The AI ​​industry currently lacks any form of public governance framework, with "pay-as-you-go" rules being set by a few companies themselves.

For ordinary users, the trend of charging for AI based on usage is unlikely to change in the short term. While the positive impact of declining token prices continues, the increasing volume of usage will offset this benefit. It is recommended that when choosing AI tools, users should not only focus on the unit price but also assess the actual trend of their usage.

For developers and enterprise customers, cost controllability in high-token-consumption scenarios such as code generation and long-text analysis is more important than unit price. A token pricing system that relies on a single vendor means that the cost structure is completely controlled by that vendor.

For creators, the proliferation of the "AI utility" narrative is itself a signal: the probability of your work being used for training is increasing, but a mechanism for rewarding it has not yet emerged. The infrastructuralization of the industry should not just mean turning model companies into the next power companies, but should also include establishing a reasonable and traceable data revenue distribution mechanism.

The current reality is that AI is becoming infrastructure, but not yet a utility. This latter title requires much more to justify, beyond just computing power and token-based billing.