AI’s Next Bottleneck Is Permission

Written By Stevan Perea

For the last two years, AI has mostly been discussed as if it lives on a screen. Everyone talks about models, chips, agents, benchmarks, and valuations. The conversation has been technical, financial, and philosophical all at once: how capable are the systems, how fast will they improve, who has the compute, which jobs change, which investors got in early.

My partner and I have spoken to a large sample of experts and scientists in the space, and they all agree: the next phase looks much less abstract. It looks like electricity bills, water use, substations, transmission lines, zoning hearings, interconnection queues, and county commissions. It looks like local residents asking why their town is being asked to carry the physical burden of someone else’s technological future.

That shift matters, because every major technology eventually stops being a story about possibility and becomes a story about permission. AI is reaching that point now because it has become more important and more real. Intelligence at scale is not solely contained to software. It is buildings, power, cooling, land, water, capital, and public tolerance.

For a while, the obvious constraint was model capability. Could the systems reason, write, code, predict, diagnose? Then it was chips. Who could secure GPUs and finance the compute? Then it was capital. Who could raise enough to stay in the race? Now another constraint is becoming visible.

Permission.

Partially in a narrow legal sense, but permission in the broad sense. Social, political, community, regulatory, environmental. The permission an abstract technology needs once it starts making physical claims on the world.

The industry talks about adoption as if it were a straight line. A product gets better, costs fall, customers see the value, capital flows, infrastructure gets built, the future arrives. But adoption passes through people who may not experience the technology as progress. A homeowner worried about electricity rates. A farmer worried about water. A local official trying to understand a utility agreement. A resident who hears “national AI competitiveness” but sees construction traffic, land conversion, and unclear local upside.

The people closest to the technology experience the buildout as momentum. The people closest to the consequences may experience it as imposition, or far worse. That gap is where the next set of AI problems will live, and where a lot of company-building will happen.

You can already see the gap in the numbers. More than 2,000 gigawatts of generation and storage, nearly twice the entire existing U.S. power plant fleet, sits waiting in interconnection queues. The typical wait from request to switch-on has stretched to roughly five years. Most of those projects will never get built at all.

The magic on the screen is real. The line to plug it in is years long.

And when the buildout reaches a specific town, the vote gets literal and consequential. Last fall, the board and planning commission of Saline Township, Michigan both rejected a giant OpenAI and Oracle data-center project after near-universal local opposition. The project’s expected power demand was compared to the output of a large power plant. The developer sued, the township settled, and construction began weeks later.

The project is getting built. But notice what it cost to get there: litigation, a fractured community, and a town that will remember. Permission moved the project forward and changed the cost of getting it built.

That is the part operators understand in their bones, and it is one reason operators sometimes see markets differently than investors. “Demand exists” does not mean “the system will let you serve it.” A customer wanting something is one part of the equation. You still need supply, approvals, working capital, vendors, distribution, and the thousand small pieces that turn a good idea into an operating reality.

AI is now running into its operating reality, and the physical world always gets a vote.

None of this means the backlash is simple or one-sided. There are real national-competitiveness arguments, real economic-development arguments, and real grid-modernization opportunities. There are also real concerns about utility costs, water, tax incentives, and who absorbs costs they never agreed to.

By one tally, more than $60 billion in data-center projects have already been blocked or delayed by local opposition. One side is not right. The point is that the fight itself shows where the bottleneck has moved.

When a technology is early, the question is whether it can work. When it scales, the question becomes whether the world can absorb it. That second question requires different instincts: the ability to translate between the technical and the civic, between national narratives and local consequences.

In infrastructure-adjacent markets, that translation is a big part of the product.

A company that reduces energy demand is not just selling efficiency. A company that makes data-center load flexible is not just selling grid software. A company that improves cooling, siting, permitting, interconnection, or community-benefit design is not operating around the edges of AI. It may be operating at the real constraint.

That is what makes this moment interesting from a venture perspective. The obvious AI companies are already obvious: the model labs, the chip companies, the enterprise wrappers, the capital flowing to the center. But markets also create opportunity at the edges of consensus.

When a category scales fast, it stresses the systems that were not built for that pace: power, siting, interconnection, public trust, and the mundane places where the future has to become installable.

Saying “AI will be big” is consensus now. Noticing which parts of the world AI will stress first is not.

The harder discernment is telling a temporary bottleneck from a structural one. Some bottlenecks get solved by money, some by engineering, some by regulation, some by time. Some never fully disappear, because they are tied to human behavior, institutional trust, and how costs and benefits get distributed.

Permission is one of the harder ones.

You cannot spend it away or benchmark it into submission. And the more important a technology becomes, the more people ask who gets to decide how it enters their lives. That is a very human reaction.

People evaluate technology by proximity: what changes near them, who benefits first, who pays, whether they were asked.

So the next phase may reward a different kind of founder. The one who can build a better model, but also the one who can move between engineers, regulators, customers, utilities, financiers, and communities. The one who understands that a technical breakthrough still has to become acceptable in the world.

Adoption is a negotiation. None of it is inevitable.

Technological change usually begins with a few people seeing capability before consensus. But the second act is different. Once the world sees the technology, the question becomes whether the technology can see the world back.

Can it understand the systems it is entering and the people it affects? Can it adapt to constraints that are not purely technical? Can it earn permission?

AI has spent a few years proving it can generate, reason, and accelerate. Now it has to prove that it can be built into the physical world without treating the physical world as an afterthought.

The next bottleneck is permission.

Sources / Further Reading

This piece was informed by recent reporting and regulatory developments around AI data centers, electricity demand, grid interconnection, and local opposition to data-center buildouts.

Public opinion

  • Reuters/Ipsos polling found broad public concern around AI data centers, including worries about electricity costs and discomfort with nearby buildout.

Electricity demand

Grid interconnection

Local opposition and the buildout on the ground

Stevan Perea
Previous

Attention Compounds
Next

Writing From the Middle