The Meter and the Mirage: A Review of the Utility Thesis for Intelligence
Imagine two bills arriving on the same day. One is from the water company: a number of liters, a fixed rate, a total. You don't need to know which water molecules you got. The other is from a law firm: a number of hours, a rate per hour, a total. But you absolutely need to know which lawyer worked those hours, because an hour from a first-year associate and an hour from a senior partner are not the same product wearing two prices. They are two different products that happen to share a billing format.
This is the distinction we think gets erased every time someone says intelligence will be sold "on a meter, like electricity or water." Sam Altman made exactly that comparison at BlackRock's Infrastructure Summit in Washington in March 2026, telling the room that OpenAI sees a future where intelligence is a utility and people buy it from the company by the unit, the way they buy kilowatt-hours or gallons. He went further and reached for a phrase from 1954, when Lewis Strauss, then chairman of the US Atomic Energy Commission, predicted that nuclear power would one day be too cheap to meter. Altman wants the same fate for intelligence: so abundant that billing it becomes almost beside the point.
We think the comparison is worth taking seriously precisely because it is doing real work, not decorative work. It is not a throwaway metaphor. It is an argument, compressed into an analogy, about what kind of thing intelligence is, how it should be regulated, who gets to price it, and what happens to the people who cannot afford the meter running. Our job in this piece is to pull that argument out of the analogy and test it, plank by plank, against the two things it borrows from: what electricity actually is, and what electricity regulation actually became.
Why the analogy is harder to defend than it sounds
A utility comparison smuggles in two claims at once, and only one of them is visible. The visible claim is about delivery: continuous availability, metered consumption, billing by usage rather than by ownership. The invisible claim is about the product itself: that it is homogeneous enough for a meter to mean anything.
A kilowatt-hour is, for almost every practical purpose, identical to every other kilowatt-hour. That homogeneity is not incidental to why electricity became a regulated utility. It is the precondition for it. Regulators can set a fair rate per unit because the unit does not vary in quality from one supplier to the next. A liter of clean tap water behaves the same way. The entire logic of utility regulation, going back to the natural monopoly theory that economists like Alfred Kahn formalized in the mid-twentieth century, rests on treating the good as fungible and the infrastructure that delivers it as the only thing worth regulating.
Intelligence, as OpenAI and its competitors currently sell it, fails that test almost immediately. A token from a smaller, faster model and a token from a frontier reasoning model are not the same unit dressed in different prices. They differ in accuracy, in reasoning depth, in reliability, in the kind of task they can be trusted with at all. Nobody buying electricity asks whether this particular kilowatt-hour will hallucinate a wrong answer under time pressure. Buyers of intelligence ask exactly that question, constantly, and route their spending accordingly. The moment you have to ask which supplier's unit you are getting, you have left utility economics and entered something closer to a professional services market, where an hour of expertise varies by who is providing it and what they know.
This is the part we think the "too cheap to meter" framing quietly papers over. It borrows the emotional promise of abundance from electricity while skipping the structural feature, homogeneity, that made electricity meterable and regulable in the first place. An analogy that keeps the abundance and drops the homogeneity is not really the same analogy anymore.
What the history of electricity actually says
There is a second problem, and it comes from history rather than economics. The 1954 phrase Altman is borrowing did not describe a policy that succeeded. It described a promise that failed, spectacularly, and the industry it was made about went in the opposite direction from where the promise pointed. Nuclear power did not become too cheap to meter. Electricity generally did not either. Instead it became one of the most heavily regulated industries in the developed world, with rate commissions, price caps, mandated service obligations, and decades of litigation over what counts as a fair return for a utility holding a state-granted monopoly.
We find this detail more instructive than the quote itself. If Altman's own reference point is a prediction that history overturned, that should temper, not reinforce, confidence in the modern version of the same prediction. The lesson electricity actually teaches is not "abundance eventually removes the need for pricing." It is "an infrastructure this essential eventually gets pulled into public oversight, whether or not the original abundance promise ever arrives." If intelligence follows the electricity precedent all the way through, the destination is not a market quietly billing by the token forever. It is regulatory intervention, rate scrutiny, and public utility commissions asking AI companies to justify their pricing the way they once asked power companies to justify theirs. That is a very different future than the one the metaphor is being used to sell, and we think it is worth saying so plainly: if you invoke the electricity precedent, you should be prepared to inherit the electricity outcome, not just the electricity marketing.
The layer the metaphor collapses
Here is where we think a review of this idea needs to do more than point out that the metaphor is imperfect. The more useful move is to separate two layers that the utility framing flattens into one.
The first layer is infrastructure: the data centers, the chips, the power draw, the cooling systems, the fiber connecting all of it. This layer genuinely resembles a utility in structural terms. It has enormous fixed costs, it benefits from scale, and a small number of firms are building nearly all of it. If any part of the AI stack deserves language borrowed from electricity regulation, it is this layer, not the one above it.
The second layer is the model itself, the actual intelligence being sold. This layer behaves nothing like a commodity utility. It is closer to a professional service, priced by capability and trust rather than by a uniform physical unit. Providers compete on the quality of what they deliver, not merely on their ability to deliver more of an identical thing. Altman's sentence merges these two layers into a single "us" providing a single metered good. We think that merger is the actual sleight of hand in the analogy, more than the electricity comparison itself. It lets a claim that is true of the infrastructure (this is capital-intensive, this behaves like a utility, this may eventually need utility-style oversight) lend credibility to a claim about the product (this is homogeneous, this is fungible, this should be priced like a commodity) that is not true at all. A reviewer's job is to notice when a true claim about one layer is being used to smuggle in a false claim about the layer sitting on top of it.
Sustainability, and a case study that just happened
Any honest review of the utility thesis has to ask what happens to economies that come to depend on it, and we think sustainability here splits into three separate questions that get run together far too often.
The environmental question is the most visible one: data centers already strain local power grids and water tables, and a genuine utility-scale rollout of AI multiplies that strain rather than shrinking it. The economic question is quieter but just as real: current pricing on most frontier AI products does not yet cover the true cost of serving it, so what looks like abundance today may be subsidized scarcity being deferred, not solved.
The third question, the geopolitical one, is the one we think gets the least attention relative to how fast it can move, and 2026 already gave us a live demonstration. In June, Anthropic released two new frontier models under its Mythos tier. Within days, the US Department of Commerce's export controls forced Anthropic to suspend access to both of them entirely. Access was not restored until the controls were lifted at the end of the month. No electric utility in memory has been switched off for an entire customer base by an export-control decision inside a matter of days, then switched back on weeks later by the same mechanism in reverse. That is not a hypothetical vulnerability we are speculating about for the sake of argument. It is a documented event, and it shows that the "grid" Altman describes can be a national policy lever in a way a power grid essentially never is for ordinary domestic users. Anyone building the utility metaphor into national infrastructure planning needs to reckon with the fact that the supply, in this case, is far more politically contingent than the metaphor implies.
Metered pricing: which gap does it close, and which one does it leave open
We think the honest answer to whether pay-per-use pricing helps or harms equity is that it does both, at two different points in the pipeline, and conflating them is where a lot of the online debate about Altman's comment goes wrong.
At the point of entry, metering genuinely lowers a real barrier. A student, a small business, or a researcher in a country without access to serious compute no longer needs capital to buy hardware before they can use a frontier model. They need only a card on file and a willingness to pay for what they consume. That is not nothing, and dismissing it as pure marketing undersells a real shift from capital expenditure to operating expenditure that has opened doors that were previously closed by cost of entry alone.
At the point of sustained, heavy use, the same mechanism reproduces an older and more familiar inequality. Large customers negotiate volume pricing that small customers never see. Retail users pay list price for the same units enterprise users get at a discount, which is precisely the dynamic that makes electricity and water bills regressive for low-income households in places without strong public subsidy. Reporting on Altman's comments from earlier this year noted, correctly in our view, that millions of people who already struggle to pay ordinary utility bills are unlikely to be reassured by hearing that a second essential resource will now be billed the same way. There is a deeper barrier that metered pricing does not touch at all, and we think this is the point most commentary skips past. Removing the capital barrier to entry does nothing about the access barrier: reliable connectivity, a capable device, and the literacy to use the tool effectively once you have it. A meter assumes you are already standing at the tap. For a large share of the population the utility thesis is implicitly addressed to, that assumption does not hold, and no amount of clever token pricing changes it.
A better fit than electricity, and a better fit than the internet
Our own reading, laid out plainly rather than hedged, is that neither of the two metaphors usually offered here, electricity and the internet, is the right one, and that the professional service frame explains the current market better than either.
Electricity fails on homogeneity, as we argued above: the product varies too much between providers for a uniform meter to price it honestly. The internet comparison, which several commentators reach for as a corrective to the electricity one, does better on delivery (continuous, metered or subscription-based, infrastructure-heavy) but still misses something electricity and bandwidth share and intelligence does not: neither electricity nor bandwidth makes decisions. A kilowatt-hour does not choose how to route itself. A gigabit of bandwidth does not interpret the data passing through it. A model's output reflects training choices, alignment decisions, and embedded assumptions that shape what you get back, not merely how much of it you get. That is a property utilities and pipes structurally lack and professional services structurally have. An accountant's advice reflects their training and judgment, not just their availability. We think that is the closer analogy: intelligence is being sold with utility-style billing wrapped around a professional-service-style product, and the mismatch between the billing metaphor and the product itself is exactly what produces the discomfort so much of the public reaction has been circling without quite naming.
There is also a third framing worth putting on the table alongside utility and professional service, because it is the one the industry itself tends to avoid: the commons. Elinor Ostrom's work on governing shared resources describes conditions under which a resource is managed collectively rather than sold by a single controlling firm, with rules set by the community of users rather than dictated from outside. The rise of serious open-weight models is, functionally, an attempt to build exactly that alternative to the "buy it from us on a meter" model Altman is describing. That two incompatible visions, centralized metered access on one side and distributed collective ownership on the other, are being pursued simultaneously and at real scale is itself worth noting as a live and unresolved tension, not a settled matter.
What would actually have to be true
To treat the utility framing as more than a persuasive analogy, we think four conditions would need to hold, and it is worth checking each one against where things actually stand rather than where the metaphor implies they stand.
The product would need to be roughly homogeneous across providers, the way a kilowatt-hour is. It currently is not, and model quality is arguably the single most differentiated axis in the entire industry right now. The market structure would need to resemble a natural monopoly, where duplicating the infrastructure is wasteful rather than competitive. The infrastructure layer arguably does resemble this; the model layer, with several serious competing labs, does not. The resource would need to be treated as functionally essential and non-substitutable, the way water and electricity are. Intelligence is heading in that direction for some tasks but is nowhere near it for most of the economy today. And the good would need political recognition as a utility, meaning legislatures and regulators formally decide it deserves that status and the obligations that come with it, such as universal service requirements and rate oversight. That recognition does not yet exist anywhere, and we suspect the industry saying "treat us like a utility" while resisting the obligations utilities actually carry is not a coincidence.
None of the four conditions is fully met. Two are partially met at the infrastructure layer and not at the product layer, which brings us back to the layer-collapse problem we raised earlier: it is the same gap, showing up again from a different angle.
Closing
We do not think Altman's comparison is dishonest so much as it is convenient, in the specific sense that any founder with a large stake in usage-based billing has every reason to describe that billing model using the most reassuring vocabulary available, and "electricity" reads as far more reassuring than "professional fees you cannot fully predict in advance." That does not make the underlying business model wrong. It does mean the metaphor deserves exactly the scrutiny we have tried to give it here rather than the applause it got in the room where it was first said.
What we would actually want, going forward, is less debate over whether intelligence resembles electricity and more precision about which layer of the stack any given claim applies to. The infrastructure genuinely rhymes with a utility and may reasonably end up regulated like one. The model sitting on top of that infrastructure behaves like a differentiated professional service and should be evaluated, priced, and eventually regulated on those terms instead. Collapsing the two into a single metered "it," the way the original comparison does, produces a clean sentence and a bad map. The clean sentence is what gets quoted. The bad map is what everyone else has to live with once the pricing, the regulation, and the access questions actually have to be worked out in practice.
Editorial Note...
___________________________________________________________________________________
Articles published in the Reviews section provide analytical, interpretive, and occasionally forward-looking perspectives on scientific, technological, and policy developments. While grounded in available evidence and referenced sources where appropriate, they may include reasoned critique, synthesis, or informed judgment. They should not be interpreted as representing scientific consensus or definitive conclusions.