As AI expands, we’ve seen how building data centers taxes the water and energy resources of communities. I figured, why not ask AI to solve this. Here was the suggestion:

My solution would be “AI pays its full cost, and growth is allowed only where it makes the grid and community stronger.” Not a ban on AI, but a much tougher operating model.

Data-center electricity demand is real and growing quickly. The IEA projects global data-center electricity use will more than double to about 945 TWh by 2030, with AI as the major driver; in the U.S., data centers could account for nearly half of electricity-demand growth through 2030. Water is also a serious local issue: large data centers can use millions of gallons per day, and many new centers are being built in water-stressed regions.

Here is the package I would push:

1. Make data centers pay for the infrastructure they require

Residential and small-business customers should not subsidize AI buildout. Utilities should create special data-center tariffs that charge large users for new transmission, substations, backup capacity, grid congestion, and local environmental costs. The Phoenix/Arizona debate shows the core problem: massive new power needs can force utilities to propose rate increases for both data centers and residents unless costs are allocated carefully.

2. No new large data center without new clean power

A company should not be allowed to simply plug into an already strained grid and claim it bought renewable credits somewhere else. Approval should require additional, local or regional clean power: solar, wind, geothermal, nuclear, hydro where appropriate, plus storage. Reuters reports that the AI power crunch is already pushing faster solar-plus-storage development because gas plants and turbines face long delays.

3. Site data centers where energy and water make sense

We should stop treating location as just a tax-incentive decision. Data centers should be steered toward places with cooler climates, abundant non-potable water options, low-carbon electricity, and available grid capacity. Building water-intensive facilities in hot, dry, water-stressed regions is exactly the wrong incentive structure.

4. Require water-neutral or water-positive operation

For water-stressed regions, the default should be: no freshwater cooling unless the operator can show no meaningful harm. Require air cooling, closed-loop systems, reclaimed wastewater, or other low-water approaches. Companies should also have to disclose water use publicly, not just energy use.

5. Price AI compute according to its real environmental cost

Right now, many AI services feel artificially cheap because the environmental and infrastructure costs are hidden. Heavy AI users should face pricing that reflects energy, water, and carbon intensity. That would encourage more efficient models, smaller models, caching, and using AI only where it adds real value.

6. Set efficiency standards for AI models and data centers

We regulate appliance efficiency; we can regulate compute efficiency. Require reporting on energy per training run, energy per inference, water-use effectiveness, power-usage effectiveness, and carbon intensity. Companies should have to demonstrate continuous efficiency improvements.

7. Use AI less wastefully

Not every task needs a giant frontier model. Many uses can run on smaller models, local models, or retrieval systems that consume far less compute. A lot of “AI everywhere” is product hype, not necessity. We need a norm of right-sized AI: use the least energy-intensive tool that does the job well.

8. Give communities approval power and revenue

Communities hosting these facilities should receive enforceable benefits: ratepayer protections, local tax revenue, water protections, emergency power commitments, noise limits, transparency, and the right to say no when the tradeoff is bad.

So my answer is not “stop AI.” It is: stop letting AI companies externalize the costs. If AI is valuable enough to build, it is valuable enough to pay for clean power, water protection, grid upgrades, and community safeguards.

Ah, but how would any of these policies actually get implemented? There’s so much corporate money and politics seems paralyzed. Here was AI’s solution to that.

I would not start with Congress. That is where good ideas often go to die.

I’d implement it through the places where decisions are already being made: state public utility commissions, local zoning boards, state legislatures, water authorities, and utility rate cases. That is less glamorous than a national AI law, but it is much harder for special interests to block everywhere at once.

The strategy would be:

1. Reframe the issue as “ratepayer protection,” not anti-AI

The winning message is not “stop AI.” It is:

Families, schools, and small businesses should not pay higher electric or water bills so trillion-dollar technology companies can build private infrastructure.

That framing can attract conservatives, progressives, rural communities, environmental groups, consumer advocates, and small businesses. There is already movement in this direction: more than 65 large-load utility tariffs had been proposed or approved in over 30 states by late 2025, and Oregon recently approved a large-load tariff framework for Portland General Electric that assigns more infrastructure costs to fast-growing customer classes such as data centers.

2. Use public utility commissions as the main lever

Public utility commissions already decide who pays for new substations, transmission lines, generation capacity, and grid upgrades. That makes them one of the most powerful venues.

Consumer advocates should intervene in every major data-center rate case and push for rules like:

• data centers pay upfront for grid upgrades;
• long-term contracts so they cannot leave residents with stranded costs;
• higher charges during peak demand;
• proof of new clean power;
• no cost-shifting to residential customers.

This is less vulnerable to national political theater because it happens state by state, case by case.

3. Build strange-bedfellow coalitions

The coalition should not be branded as environmentalists versus tech. It should include:

• ratepayer advocates;
• local chambers of commerce;
• farmers and water users;
• school districts and municipalities;
• environmental groups;
• conservative fiscal watchdogs;
• labor groups that want real local benefits, not just tax giveaways.

Special interests win when the opposition is fragmented. This issue becomes powerful when people realize the same project can raise electric bills, consume water, receive tax breaks, add pollution, and create relatively few permanent local jobs.

4. Make transparency the first political demand

Before arguing about bans or limits, require disclosure.

Every proposed large data center should publicly report expected electricity demand, peak load, water use, backup generation, emissions, noise, tax incentives, permanent jobs, and who pays for new infrastructure. The EU is moving toward requiring large data centers to disclose metrics such as water use and clean energy use, which is a useful model.

Transparency is politically harder to oppose. Once the numbers are public, communities can make better decisions.

5. Give local governments a model ordinance

Most towns do not have the staff to evaluate a 500-megawatt data-center proposal. So advocacy groups, universities, and state attorneys general should provide a standard ordinance that local governments can adopt.

That ordinance would require:

• environmental and grid-impact studies;
• water-impact studies;
• limits on diesel backup generators;
• noise rules;
• community benefit agreements;
• proof that residents will not subsidize the project;
• independent review paid for by the developer.

This turns a vague concern into a ready-to-use governance tool.

6. Force the issue through “no subsidy without conditions”

Many data centers depend on tax abatements, fast-track permitting, special utility arrangements, or public infrastructure support. That is the pressure point.

The rule should be simple:

No public benefit unless there is a public return.

That means no tax breaks unless the company agrees to clean power, water protections, local infrastructure payments, public reporting, and enforceable penalties.

7. Use moratoriums strategically, not permanently

Temporary moratoriums can be useful when a community is being rushed. They should not be framed as “no data centers ever.” They should be framed as:

Pause approvals until we know who pays, where the power comes from, how much water is used, and what protections residents have.

That is reasonable and defensible. State and local governments are already scrutinizing data centers through proposed bills, reporting rules, energy-cost protections, and in some cases moratoriums.

8. Make companies compete on responsibility

The goal is to divide the industry. Some companies will fight all regulation. Others may accept standards if it gives them faster approval and reputational advantage.

Create a “responsible data center” certification tied to:

• no residential rate increases;
• new clean power;
• low or no potable-water use;
• transparent reporting;
• local benefits;
• flexible computing that reduces peak grid stress.

Then governments can say: meet the standard and your permitting moves faster; refuse and you face deeper review.

9. Use litigation when disclosure is missing

Where projects hide water use, energy demand, emissions, or noise impacts, legal challenges can slow or improve approvals. That is already happening: Equinix’s proposed Cape Town data centers have faced objections over missing information about water, power, emissions, and noise.

Litigation is not the whole strategy, but it is a useful backstop when communities are denied basic facts.

10. Turn this into a pocketbook issue

The most effective political campaign would not lead with climate. It would lead with utility bills.

The message:

AI companies can build, but they must bring their own clean power, pay for the grid upgrades, protect water, and guarantee that local families do not pay more because of them.

That is how I would get it implemented: not by waiting for ideal federal legislation, but by creating a repeatable state-and-local playbook that makes irresponsible projects expensive, transparent, and politically difficult — while giving responsible projects a path forward.

But then, isn’t there a way AI could come up with a way to do AI without requiring as many resources?

Turns out that intelligent consideration of options could decrease demands by at least 50%.

Yes. AI can help make AI less resource-intensive. But it cannot solve the problem by technology alone.

There are at least five promising paths:

1. Smaller models for most tasks

A lot of AI use does not require the biggest frontier model. Drafting an email, summarizing a document, classifying support tickets, tutoring on routine content, or answering internal policy questions can often be done with smaller, specialized models.

One 2025 paper estimated that better model selection — using a smaller model when it is sufficient — could reduce AI energy consumption by about 27.8% worldwide in 2025, with savings varying widely by task.

2. Smarter routing

Instead of sending every request to the largest model, an AI system can first ask: “How hard is this task?”

Simple request → small model.
Medium request → mid-sized model.
Hard reasoning or high-stakes task → larger model.

That is probably one of the most practical near-term solutions. It is the AI equivalent of not using a tractor-trailer to pick up a loaf of bread.

3. More efficient inference

A lot of energy is used not just to train models, but to run them again and again for users. Researchers have found that inference optimizations can sharply reduce energy use, but the gains depend on the workload, software stack, hardware, and how the model is served. One 2025 paper found that relevant inference optimizations could reduce total energy use by up to 73% compared with unoptimized baselines.

These improvements include batching, caching, quantization, better decoding strategies, shorter prompts, shorter outputs, and avoiding repeated work.

4. Better hardware and data-center design

AI can help optimize cooling, workload scheduling, chip utilization, grid interaction, and water use. There are also hardware changes, such as photonics, that may reduce heat and energy losses from data moving inside data centers. Recent reporting notes that copper interconnects are becoming a bottleneck in large AI clusters and that optical approaches may reduce heat, complexity, and power losses.

5. AI to improve the grid

AI can also help forecast electricity demand, integrate renewables, manage batteries, optimize cooling loads, and shift non-urgent computing to times when clean power is abundant. In principle, data centers could become more flexible grid participants instead of constant stressors.

But here is the catch: efficiency can increase total demand. When AI becomes cheaper and faster, companies put it into more products, users make more requests, and developers build more AI workflows. This is the Jevons paradox problem: efficiency lowers cost, which can increase total consumption. Recent research on AI and rebound effects argues that technical efficiency alone is unlikely to guarantee lower environmental impact without governance and economic incentives.

So my answer would be:

Yes, use AI to make AI more efficient — but pair that with rules requiring lower total environmental impact, not just better efficiency per query.

The standard should be something like:

1. Use the smallest adequate model.
2. Disclose energy, water, and carbon per major service.
3. Require new data centers to use additional clean power.
4. Charge large AI users for grid and water impacts.
5. Reward companies that reduce total resource use, not just resource use per token.

Otherwise, we may build a system where each AI interaction becomes more efficient, but the world uses so many more of them that total electricity, water, and emissions keep rising.

Okay humans, what are you going to do about it?