AI Has Started Behaving Like a Utility

I ran out of AI.

Not electricity. Not water. Not internet. AI.

I had exhausted the allowance on one of the highest plans available from a major lab, and the reset was still hours away. I pay for top plans across the major providers, so another endpoint was available. I also spend money every month on end-user APIs for small projects that do nothing glamorous. They remove friction from my life, one repetitive task at a time.

There was an even more obvious fallback in the room. I own seven GPUs, including two Blackwell 6000 Pros, with roughly 250 GB of VRAM across the machines. I can run capable local models. I know how to do it.

I waited.

That decision is more important than the hardware. For a short interruption, moving the work local was not worth the switching cost. Context would have to move. Tools would behave differently. The workflow would change just long enough to be annoying, then I would move everything back when the allowance reset.

The model did not fail me. The subscription did exactly what the subscription said it would do. What surprised me was my own response to losing it. I did not feel as if a piece of software had closed. I felt as if a service had gone out.

AI has started behaving like a utility.

The generator in the garage

I have a generator at home. It can carry the important loads when the grid disappears, but whether I start it depends on the expected duration of the outage. If the lights flicker and the utility is likely to recover in a few minutes, I do nothing. Starting the generator means fuel, noise, breakers, attention, and then another transition back to the grid. The backup is capable. The failover is not free.

If a hurricane has taken down poles across the neighborhood, the calculation changes. Then the switching cost is trivial compared with the cost of waiting.

Local AI now occupies that same place in my behavior. The machines are mine. The models are available. For the right workload, especially one that needs control or sustained use, local is absolutely the right answer. But when a frontier service is interrupted briefly, I treat my local capacity like a generator during a short grid outage: possible, ready, and not worth starting yet.

This is not an argument against local AI. It is evidence that the default has moved.

A technology stack is something you choose to operate. A utility is something you organize behavior around, often without noticing until it disappears. Utility status does not begin when regulators declare it or when service quality becomes perfect. It begins when the cost of absence becomes visible in ordinary life.

Mine became visible that afternoon.

I have seen enough waves to distrust the slogans

My first chapter in college was Production Engineering, with an emphasis on software factories. Ten years later, I returned to the same university for Economics. After that came years of reading evolutionary psychology, well beyond casual interest, because I wanted to understand why humans repeat behaviors that make little sense when viewed only through technology or finance. Sociology has always pulled at the same thread for me: how groups form, how institutions harden, how yesterday’s novelty becomes today’s expectation.

That background does not make me an economist or a sociologist. It gives me several imperfect lenses and a strong suspicion of explanations that fit on one slide.

The technology career helps too. It has been long enough to watch waves arrive carrying the same promises, attract the same camps, punish some early adopters, reward others, and then settle into something much less dramatic and much more consequential. Early adopter scars teach a different lesson than conference decks. You learn that being right about the direction does not make you right about the timing, the vendor, the price, or the implementation.

In a previous article on the three phases of technological adoption, I worked from a theory I first heard Michio Kaku describe and connected it to the economics of collapsing constraints. People begin with it will not work. Then comes it may work, but not for me. Eventually the technology becomes so normal that former skeptics remember themselves as participants.

AI is moving through those phases, but the corporate conversations are wonderfully contradictory. In one room, AI is the future and any company moving slowly is already dead. In another, smart operators say the people who approved the latest AI investments will be fired when the returns fail to clear. I have heard both positions recently, stated with complete confidence.

Both may be directionally right.

Capital can be misallocated around a capability that still becomes essential. Companies can overpay, implement badly, confuse demonstrations with operating models, and burn extraordinary amounts of money while society keeps adopting the underlying function. Market enthusiasm and social dependence are not the same variable. The first can collapse while the second keeps compounding.

The internet already taught us this, although memory has cleaned up the story.

Banks did not need the internet to be magical

Banks did not adopt the internet because websites were philosophically interesting. They adopted it because a customer could check a balance, transfer money, pay a bill, or answer a routine question without occupying a branch employee’s time.

That convenience was real for the customer. The economics were even more real for the bank.

Every transaction that moved from the branch to the browser changed the operating model. It reduced repetitive service work, altered the value of physical locations, extended service hours without extending every staffing schedule, and protected margin. The internet became a customer-service utility because both sides developed behavior around it. Customers expected access. Banks expected the channel to absorb toil.

Nobody needed the internet to replace banking. It only had to become the cheapest dependable path for a large class of ordinary interactions.

This is where much of the AI replacement argument misses the useful layer. I have written before that applied AI is human augmentation, not replacement. The utility thesis sits directly on top of that idea. A capability does not need to replace the physician, accountant, engineer, teacher, or service representative to become structurally important. It only needs to absorb enough friction that removing it forces the organization to add time, labor, or delay back into the system.

That is already happening.

Four hundred calls a day

A physician I know owns a clinic that receives roughly four hundred calls a day. He also has a computing background, which means he can look past the sales language and see a queue.

The calls do not disappear because hiring is expensive. Patients still need appointments, instructions, confirmations, answers, and a human when the situation actually requires one. Without another layer in front of that demand, the arithmetic points toward doubling the service-desk staff. Not because the existing staff failed. The system is receiving more repetitive work than the current operating model can absorb.

AI gives the clinic operational leverage. It can help handle the routine layer, reduce the queue, and leave people with more capacity for the cases where judgment and empathy matter. That is not replacing the clinic. It is not replacing medicine. It is changing the cost and shape of access.

Now remove that AI layer.

The business probably does not die. It degrades. Hold times rise. People return to repetitive explanations. The queue spills into work that should have received attention sooner. Payroll pressure returns, or service quality falls, and perhaps both happen for a while because organizations do not double staffing between lunch and dinner.

This is the early form of utility dependence. The dramatic test is not whether an organization ceases to exist when the service disappears. Most businesses do not instantly die when the internet goes down either. The useful test is whether the operation becomes materially worse, quickly enough that leaders have to plan around the absence.

Once degraded mode enters the management conversation, the capability is no longer a toy.

Participation changes the category

Where I live, families that cannot afford internet access at home can receive help through the public school system. The exact machinery matters less than the reason it exists. A child without connectivity is not merely missing a consumer convenience. That child is losing access to normal participation in education.

Society crossed that line gradually. Home internet began as an enthusiast’s connection, became a household product, then turned into the path through which schools distribute assignments, parents receive notices, students submit work, and families reach the institution. Once participation depended on it, access became a civic problem.

AI is not at that point yet. I would not pretend otherwise. The quality is uneven, the economics are unsettled, the rules are moving, and access is still distributed in ways that reflect income, language, geography, institutional policy, and simple awareness. We are early.

But the direction is visible in behavior. People now draft correspondence, navigate forms, interpret unfamiliar language, learn new subjects, organize family logistics, prepare for difficult conversations, and clear small administrative obstacles with AI beside them. None of those acts looks revolutionary alone. Put them together across millions of lives and the social meaning changes, because people begin to expect the extra capacity in themselves and in everyone around them, employers included, whether anyone has formally agreed that expectation is fair.

That last part will become uncomfortable.

Every utility creates a participation boundary. If one group can afford continuous access to high-quality assistance and another cannot, the difference is not only convenience. It becomes speed, confidence, comprehension, and the amount of bureaucracy a person can push through before exhaustion wins. The same capability that lowers toil can quietly raise the baseline imposed on everyone.

This is why the utility frame matters more than another argument about which model leads a benchmark this month. The social question is no longer only what AI can do. It is what institutions will assume people can do once AI is present.

You cannot cage a river

The current debate often assumes society can make a clean binary choice: permit AI or stop it, invest or retreat, accelerate or wait. That is not how technologies behave once they become entangled with incentives and daily routines.

You cannot cage a river. You can shape the riverbed.

The riverbed is access, pricing, education, labor expectations, institutional policy, and the boundaries we place around consequential decisions. It is also the honest economics inside a company: which work becomes cheaper, which service gets better, which cost merely moves somewhere less visible, and who absorbs the failure when the assistance is unavailable.

Shaping the riverbed is not surrender. It is the practical work. Banks did it when the internet became a service channel. Schools did it when connectivity became necessary for participation. The clinic is doing it now, not because AI is destiny, but because four hundred calls still arrive tomorrow morning.

My own evidence is smaller. I reached a limit, looked at seven GPUs, and chose to wait for the service to return. No benchmark could have told me I would do that. Behavior did.

This is the first piece in a short series about AI’s new role in society. The next question is larger, and I am less certain of the answer: what kind of phenomenon are we actually watching?

I do not think AI is the fifth industrial revolution.

It may be the first phenomenon of its kind.