I was recently quite concerned after researching AI data centers as well as troubling occurrences of rare books being destroyed to feed the data to private AI for my recent articles. There is so much potential for problems with the existing systems.
See Why AI Data Centers Are So Controversial: Energy, Water, and Rising Electricity Costs for more details about all the current issues surrounding AI data centers. Also read The Hidden Costs of Destroying Physical Books for AI Training Data.
So I sat down to have a chat with—appropriately—ChatGPT and asked it specifically about the issue with many new data centers planned and what that means for the nearby homes, businesses and communities. I wanted to know if there wasn’t a better approach than simply constructing massive data centers that use massive amounts of land and resources like water and electricity. Here is what the AI has to say about AI data centers.
There are substantially better engineering approaches than simply continuing to build ever-larger conventional data centers, although some of the alternatives are harder to deploy and require changes in how AI systems are designed, not just where the buildings are placed.
The interesting part is that the problem isn’t really “AI needs too much electricity.” It’s that we’re currently solving a computational problem with a fairly blunt architecture: enormous centralized buildings full of general-purpose accelerators, running continuously, with substantial cooling and electrical infrastructure.
There are several directions that could change that.
1. Stop treating every AI task as a giant AI task
This may be the biggest opportunity.
A huge amount of computation doesn’t require the largest model available. A system could dynamically route a request:
tiny model → medium model → large reasoning model
depending on the complexity of the task.
For something like “What is 17 × 24?” there is no reason to invoke a massive reasoning model. For a complicated scientific problem, there may be.
That means future AI could become much more computationally selective rather than simply becoming bigger.
2. Specialized chips could dramatically reduce the waste
Today’s AI accelerators are already specialized, but there’s still enormous room for improvement.
Instead of throwing increasingly powerful general-purpose GPUs at every problem, hardware can be designed around particular operations:
- sparse computation
- lower-precision arithmetic
- memory-local processing
- neuromorphic approaches
- optical/photonic computation
- application-specific accelerators
The goal isn’t necessarily more computing power. It’s more useful computation per watt.
And this is where your “surely there has to be a better solution” instinct is particularly reasonable. Historically, computing has repeatedly become dramatically more efficient even as demand exploded.
3. The really interesting possibility is doing computation closer to where it’s needed
Instead of every AI request traveling to a giant facility, some processing can happen on:
- phones
- computers
- cars
- appliances
- local servers
- smaller regional data centers
A phone doesn’t need a frontier-scale model sitting inside it, but a surprisingly capable smaller model can handle many everyday tasks locally.
That has another advantage: you don’t have to move all that information back and forth across the network.
4. Data centers themselves can become much less environmentally obnoxious
The conventional picture is:
giant building + enormous electrical connection + cooling plant + backup generators + water + transmission infrastructure.
But there are better possibilities.
For example, facilities can be located where they can use:
- abundant renewable electricity
- naturally cold climates
- reclaimed water rather than potable water
- waste heat for district heating or industrial processes
- existing industrial infrastructure
- electricity that would otherwise be curtailed
Waste heat is particularly interesting. A computer facility doesn’t merely consume energy—it turns almost all of the electricity it consumes into heat.
Instead of treating that heat as something to get rid of, you can treat it as a secondary product.
5. Cooling may be one of the biggest opportunities
A tremendous amount of engineering effort goes into moving heat away from chips.
Air cooling is relatively inefficient at very high densities.
Liquid cooling can move much more heat with much less airflow, and increasingly sophisticated systems can put cooling directly against the heat-producing components.
That can reduce:
- fan noise
- building airflow requirements
- cooling energy
- physical space
- sometimes water consumption.
So your concern about noise isn’t trivial. A data center isn’t necessarily noisy because “computers are loud”; it’s largely the enormous cooling and mechanical infrastructure that creates the noise.
6. And there is a more radical possibility: don’t build the same computer twice
This is where I think the industry has an enormous optimization problem.
Suppose millions of people ask AI essentially the same question.
You don’t necessarily want millions of completely independent computations.
You can cache, reuse, compress, distill, route, specialize, and share computations.
Likewise, a model doesn’t necessarily need to perform the same expensive internal operation for every input.
AI efficiency may ultimately come as much from cleverness in computation as from better hardware.
The catch is that none of this means we can avoid building additional infrastructure altogether.
AI demand is genuinely enormous, and increasingly sophisticated models really do require substantial computational resources. Some workloads simply cannot be squeezed onto a phone.
But I don’t think the long-term answer has to be:
“Build a 10× bigger data center every few years.”
A more sustainable trajectory would look more like:
smaller models + intelligent routing + specialized hardware + sparse computation + local processing + efficient cooling + strategic geographic placement + reuse of waste heat + better utilization of existing infrastructure.
And there’s a fascinating historical precedent here. The early computers occupied rooms. Then they occupied desks. Then they went into our pockets.
The fact that AI currently occupies enormous buildings doesn’t necessarily mean that’s where the technology has to end up.
The most environmentally sensible future may actually be one in which AI becomes simultaneously more powerful and less physically conspicuous, because the computation becomes dramatically more efficient and increasingly distributed.
The next generation of data-center technology is actually proposing right now—nuclear power, geothermal, liquid cooling, underwater/undersea concepts, waste-heat reuse, modular data centers, and edge AI. There are alternatives if we slow down and research them appropriately.
Related:
- What Every Christian Needs to Know About AI and Faith
- Christian AI Alternatives: Gab AI and Other Faith-Aligned Tools for Believers
- Why AI Data Centers Are So Controversial: Energy, Water, and Rising Electricity Costs
- The Hidden Costs of Destroying Physical Books for AI Training Data
- How to Get Accurate Answers from AI (And Avoid Wrong Information)
- How to Fact-Check AI Answers (Step-by-Step Guide)
- Why AI Gives Wrong Answers (AI Errors Explained Simply)
- What Is AI Bias? (Simple Explanation + Examples)
