AI Capacity Needs A Civic Interface

The most revealing AI interface right now may not be the chat box, the code editor, or the little sparkle button that keeps appearing in places where a perfectly serviceable menu used to live. It may be the substation.

There is a strange split in how AI capacity is being presented. On the product side, intelligence is becoming smoother, cheaper, faster, and more ambient. OpenAI's announcement with Broadcom for its Jalapeno inference chip describes a full-stack push toward faster, more reliable, more affordable AI, with deployment planned at gigawatt scale. The surface story is access. More people, more workflows, more useful intelligence, less waiting.

Then the other story arrives with a hard hat on. Business Insider reported today that Anthropic has signed a 20-year lease for a Kentucky AI data center campus expected to supply roughly 400 megawatts of capacity. The site is not an abstract cloud. It is a former industrial property in Hawesville, a place with roads and neighbors and power lines. The International Energy Agency's Energy and AI report puts the larger issue plainly: there is no AI without electricity for data centers.

That sentence ought to sit closer to the product demo.

The trouble is not that AI uses electricity. Everything useful uses something. Hospitals use electricity. Elevators use electricity. Dishwashers use electricity, which is good because no civilization has been improved by requiring every dinner guest to develop a personal relationship with a sponge. The problem is that AI tools are being designed to feel weightless while their physical footprint becomes harder to ignore. The user gets a glowing prompt field. The town gets a load forecast.

This makes capacity a design problem, not only an infrastructure problem. Interfaces teach people what kind of thing they are using. A thermostat teaches that heat has a setting and a lag. A battery icon teaches that mobility has limits. A calendar teaches that attention occupies space before and after the meeting itself, though many organizations continue to treat this as a rumor. But most AI products still teach abundance. They make intelligence feel like a faucet without a meter, a colleague without a commute, a machine room without a door.

The new AI stack is no longer just models and apps. It is chips designed around inference workloads, racks tuned for model serving, cloud contracts, transmission planning, cooling choices, backup generators, tax deals, and utility negotiations that can outlive several versions of the product that justified them. A manager asking an assistant to summarize another dozen decks may not need a live carbon odometer in the corner. I am not eager to turn every sentence into a moral invoice.

But a humane interface should not make major consequences disappear simply because they are inconvenient to narrate.

We have a habit of treating infrastructure as visible only when it breaks. The road matters when the pothole eats a tire. The grid matters when the lights flicker. AI capacity is sliding toward the same pattern. It is invisible when the answer arrives quickly and suddenly everyone else's problem when the data center wants water, power, land, tax treatment, or an emergency exception to a planning process built for smaller machines.

The design language of AI starts to feel too clean for the system it represents. "More accessible" is a lovely phrase when it means a student can get help, a small business can automate a tedious task, or a developer can run a test that used to be too expensive. It is less complete when the access is made possible by regional loads that communities discover through hearings, utility filings, or a construction fence. The benefit may be real. The missing interface is also real.

Some of this belongs in public policy. Utilities and regulators need serious planning, not vibes wearing a reflective vest. Researchers are already warning that spatially concentrated data center growth can stress regional power systems even when global averages look manageable. The honest system lives in boring distinctions: direct water use, indirect water use, carbon accounting, local grid stress, and the timing of clean-energy additions.

But product teams and AI companies do not get to hand the whole problem to city councils after designing their tools as if compute were a private magic trick. The interface to AI capacity should include cues for scarcity, timing, and shared cost. Enterprise tools could distinguish urgent agent work from batchable work and make slower, cheaper, lower-impact modes feel normal. Organizations could build their own capacity etiquette: what deserves an agent sprint, what can wait, what should be done locally, what should be reviewed before another background run starts chewing through invisible resources.

This is not a call to make AI miserable to use. Friction is not a virtue by itself, and there is already enough software in the world designed by someone personally offended by lunchtime. The point is to make consequence legible before it hardens into resentment. People are more capable of tradeoffs when the tradeoffs are treated as part of the work rather than a scandal discovered afterward.

The civic version matters even more. If a town is asked to host the physical body of an AI system, it should not be offered only jobs, revenue projections, and the usual glossy promise that this time the future will be clean and quiet. It should get a usable account of power, water, noise, emergency planning, tax exposure, upgrade cycles, and who bears the cost if the business logic changes before the infrastructure does. The fact that a model can write a good email does not mean a county should have to infer the rest from a bond document.

AI has spent the last few years trying to become intimate: inside our documents, calendars, browsers, codebases, meetings, classrooms, and health decisions. Now it is becoming municipal. It will sit in places, draw from places, shape places, and compete with other needs that also have human faces attached. That does not make AI illegitimate. It makes the design problem larger than the screen.

The next interface for AI capacity should help people understand not only what the system can do for them, but what it is asking from the world around them. The answer field is only one side of the machine. The other side has a fence, a transformer, a water plan, a lease term, and neighbors who deserve more than a progress spinner.