Syntic
All posts

Published October 4, 2026 in Perspectives

Intelligence Is Becoming Infrastructure

By Syntic AI

In a 1961 lecture series marking MIT's centennial, John McCarthy suggested that "computation may someday be organized as a public utility, just as the telephone system is a public utility," with each subscriber paying "only for the capacity that he actually uses." It took about 45 years. In 2006 Amazon launched S3 and then EC2, which promised developers they would "pay only for capacity that you actually use." Five years later, the US National Institute of Standards and Technology wrote "measured service" into its official definition of cloud computing.

Machine intelligence is moving along the same path, and much faster. Within a few years it has gone from a product people visited to a metered input that other software calls. The economics and the policy documents now treat it the same way. America's AI Action Plan, published in July 2025, puts it plainly: "Like most general-purpose technologies of the past, AI will require new infrastructure."

The argument of this piece is that the utility analogy is right about supply and wrong about trust. Intelligence is acquiring the features of infrastructure: metering, capacity planning, physical build-out, public-interest regulation. But a kilowatt-hour is inert and interchangeable, and a unit of intelligence is neither. What a utility rests on is a guarantee that every unit is the same. Intelligence will need its own version of that guarantee, and nobody has built it yet.

What turns a technology into infrastructure

Economists have a precise term for technologies that end up underneath everything else. In their 1995 paper, Timothy Bresnahan and Manuel Trajtenberg defined general-purpose technologies by three traits: pervasiveness (they are inputs to many downstream sectors), "inherent potential for technical improvements", and innovational complementarities, meaning that progress in the core technology raises the productivity of innovation built on it. Steam, the electric motor and semiconductors fit the definition.

Large language models seem to as well. Tyna Eloundou and colleagues estimated that about 80% of the US workforce could have at least 10% of their tasks affected by these models, and about 19% could see at least half their tasks affected. They concluded that the models "exhibit traits of general-purpose technologies."

Being a general-purpose technology is not quite the same as being infrastructure. Infrastructure is what a general-purpose technology becomes once its supply is organised as a shared service that others build on, and once they stop owning it themselves. Electricity reached that stage when factories stopped running their own generators. Computing reached it when companies stopped buying servers for peak load. The signs are recognisable: the supply is metered, users and suppliers negotiate over reliability, interfaces are standardised, the physical build-out turns into a planning problem, and governments start asking who gets access and on what terms.

Intelligence now shows all five, and each changes how organisations should behave.

Metering came first

The token is intelligence's kilowatt-hour. It is how the input is counted, priced and budgeted, and it made intelligence something software calls rather than something people visit. As our analysis of why the cost of intelligence is falling while the bill keeps rising showed, it is also why software pricing is moving from seats to credits and usage.

The more interesting parallel is in what metering turns into once a utility matures. Samuel Insull, who built Chicago's electricity system, described his company as the pioneer of the Wright demand meter, which let a utility charge not only for the energy a customer used but for the peak capacity it required, because idle capacity is the utility's real cost. Intelligence pricing is going the same way. Model providers now sell asynchronous batch processing at a discount for work that can wait, a default tier for everything else, and in some cases reserved or prioritised capacity under term commitments. One major provider describes its default tier as serving requests "with best-effort availability."

"Best effort" is an honest description, and it is not what a utility sells.

Reliability is the unfinished part

A utility's real product is not the electron. It is the promise that the electron will be there. Grids are planned against reserve margins, and customers who need more certainty pay for it. For intelligence, that apparatus is still being assembled. Capacity is rationed through rate limits, overload errors appear at peak times, and the models themselves are retired on published schedules, a dependency risk we examined in why one AI model is not enough.

That matters more as AI moves from answering questions to doing work. A company whose support, scheduling and back office run on AI agents now has an outage risk that looks like a power cut, not like a slow website. The response is the one large electricity users settled on long ago: know your peak demand, keep a second source, decide in advance which loads are shed first, and test the failover before you need it. Few organisations yet treat their intelligence supply with that discipline. The ones that do will look less like software buyers and more like utility customers.

The build-out has become an energy problem

The clearest evidence that intelligence is now infrastructure is physical. The International Energy Agency's April 2026 report on energy and AI reads less like a technology survey than a grid-planning document.

In Texas, the queue of large loads seeking a grid connection, about three-quarters of them data centres, was around 63 GW in December 2024. It more than doubled to 130 GW within four months and passed 230 GW by January 2026. The state's all-time peak demand is 85 GW. Not all of those projects will be built, but the queue shows where the pressure is. The IEA reports that grid connections can take five to ten years in many jurisdictions and that transformer lead times average two to three years. It also cites estimates that in the second quarter of 2025, 20 data centre projects in the United States, worth about $98 billion, were blocked or delayed, partly because of local concerns about environmental impact and electricity affordability.

Governments have responded in the vocabulary of infrastructure. US Executive Order 14318 of July 2025 speeds up federal permitting for qualifying data centre projects, including those that add more than 100 MW of new load. The UK's AI Opportunities Action Plan of January 2025 created "AI Growth Zones" for data centre build-out and committed to expanding the national AI Research Resource "by at least 20x by 2030." The European Union has selected 19 AI Factories built on its public supercomputing network and opened a process for up to seven AI Gigafactories, each designed around more than 100,000 advanced AI processors.

The IEA's recommendations are those of a utility regulator: reform connection queues, improve demand forecasts with better disclosure from technology companies, and use tariffs to allocate the cost of grid upgrades fairly. Who pays for the substation is now an AI policy question.

Regulation is starting to look utility-shaped

Utilities were never regulated purely as products. In his 1898 presidential address to the National Electric Light Association, Insull argued for exclusive franchises "coupled with the conditions of public control," with charges based on "cost plus a reasonable profit." By 1909 he was calling electricity supply "a business which is a natural monopoly." Telephony went through a version of the same bargain. In the Kingsbury Commitment of December 1913, made under antitrust pressure from the US Attorney General, AT&T agreed to sell its Western Union stake, stop acquiring competing telephone companies and let independent carriers connect to its toll lines. Two decades later the Communications Act of 1934 wrote a public-interest goal into law: making service available "to all the people of the United States" at "reasonable charges."

AI regulation is starting to show the same features. The EU AI Act's obligations for general-purpose AI models have applied since 2 August 2025. Under Article 51, a model trained with more than 10^25 floating-point operations is presumed to have "high impact capabilities" and faces additional duties to assess and mitigate systemic risk. That presumption regulates a model by the scale of its production, which is how utilities and power plants have long been treated: the bigger the plant, the heavier the obligations. Meanwhile, under the Digital Omnibus changes politically agreed in May 2026, most rules for high-risk systems will apply from December 2027 or August 2028. The rules for the infrastructure are arriving before the rules for many of its applications.

The public-interest questions are also familiar. On access, the US National AI Research Resource pilot, the UK's research compute and the EU's AI Factories are all attempts to keep frontier-scale computing available to researchers and smaller firms. On concentration, the US Federal Trade Commission's January 2025 study of partnerships between the largest cloud providers and AI developers raised three concerns: influence over access to compute and talent, higher switching costs, and cloud partners' access to sensitive information. In telephony, antitrust pressure of a similar kind is what produced AT&T's interconnection commitment.

The question no jurisdiction has answered is what the bargain looks like for intelligence. Utilities accepted obligations on universal service, non-discrimination and reliability in exchange for permission to operate at scale. Permits, grid connections and public compute are now on offer to AI builders. What is being asked in return is still unclear.

Where the analogy breaks

The breaks in the analogy matter more than the parallels.

Intelligence is not fungible

A kilowatt-hour from one generator is identical to a kilowatt-hour from another. That interchangeability is what makes grids, markets and meters possible. A unit of intelligence is not like that. The same request can produce different outputs from different models, from different versions of the same model, and sometimes from the same model twice. Metering counts tokens. It says nothing about whether the tokens were any good.

The International AI Safety Report 2026, chaired by Yoshua Bengio, describes leading systems that "may excel at some difficult tasks while failing at other, simpler ones," and warns that "performance on pre-deployment tests does not reliably predict real-world utility or risk." No grid operator has to manage a supply whose quality varies with the question.

Carriers carry; models act

Telephone networks were regulated as common carriers in part because they carried messages without writing them. An AI model writes the message. Increasingly it also schedules, refunds and updates records itself. As we argued in AI agents are becoming the execution layer for business, it sits where business logic runs, not in the wires underneath. That is why carrier-style liability does not transfer. In the Air Canada chatbot case discussed there, a British Columbia tribunal held the airline responsible for what its chatbot told a customer. A utility does not answer for what customers do with its power. A company is likely to answer for what its intelligence does on its behalf.

Infrastructure without intent is an assumption

A grid has no objectives. The more autonomous AI systems become, the less that holds for them. Agents pursue goals over many steps, and the same safety report notes that "the inner workings of models remain poorly understood" and that some models can tell test settings from real deployment. Utility regulation assumes a passive supply whose failures are physical and measurable. Alignment, which means making sure a system pursues what its operators intended, has no equivalent in electricity.

What the gap implies

Electricity became safe to build on only when a second layer grew up around it. Electrical codes, independent testing laboratories and inspection made each socket trustworthy, not just each power station. Intelligence needs the same layer, adapted to a supply that varies in quality and takes actions. That means evaluation that keeps running in production and is not a one-off benchmark, provenance that travels with outputs (as with the records that follow generated media), identity and permissions for agents, and audit trails that explain decisions. Today most of that layer is built separately by each organisation.

What changes for companies, governments and individuals

For companies, the practical rule is to buy intelligence like a utility and govern it like a workforce. Plan capacity, budget compute alongside headcount, keep alternative suppliers and design failover, as the AI-native operating model requires. But do not outsource the quality layer. The supply is becoming a commodity; the evaluations, policies and accountability built on top of it are not, and they are where a company's advantage and its exposure both sit.

For governments, the task is to make the bargain explicit. If public permits, grid capacity and research compute are going to support private intelligence supply, the conditions should be clear: cost allocation that does not pass grid upgrades on to households, access commitments for researchers and smaller firms, interoperability that keeps switching costs reasonable, and oversight that looks at what systems do as well as how much compute trained them.

For individuals, intelligence will increasingly be ambient, built into the phone, the help line, the clinic's front desk and the code editor, much as electricity disappeared into the walls. The useful skill then shifts from operating AI tools to judging their outputs. The open civic question is the one telephony faced a century ago: whether access is universal, and on what terms.

Where Syntic stands

Syntic's view is that intelligence becomes more valuable as more people can use it, and that trust has to be earned through security, reliability, transparency and meaningful human control. Read through the infrastructure argument, the work is not only to make intelligence abundant but to make it dependable enough to build on and controllable by the people who rely on it.

The socket and the system

McCarthy's prediction came true for computing because a unit of compute is close enough to interchangeable that it could be pooled, metered and sold on demand. Intelligence is following the same route, through the same institutions: meters, capacity contracts, grid queues and regulators.

The difference is in what comes out of the socket. Electricity's hard problem was getting power to everyone. For intelligence, distribution is the easier half, and the harder half is knowing what the delivered unit will do. That is the part of the infrastructure still to be built, and it will decide how safely everyone else can build on top of it.

Sources