The Chatbot Has a Building Behind It

At the end of the last chapter, I was beginning to see AI as much more than a chatbot.

But another question was starting to bother me.

If all of this intelligence felt almost weightless on a screen, what was happening behind it physically?

Around the same time, I was taking another elective at UCLA, Leaders in Sustainability, and discovered that I could work toward a sustainability certification that took me into public policy, social impact and choice architecture.

That was one of the things I came to appreciate about UCLA Anderson. You could begin in one programme and find yourself learning from completely different disciplines because a question took you there.

The subjects were different, but a few ideas stayed with me.

One was the importance of holding multiple simultaneous complex truths.

Something can create enormous value and still create costs somewhere else.

The person receiving the benefit is not necessarily the only person affected by it.

And small changes in how choices are presented can influence behaviour.

Those ideas began to collide with a world I already knew.

I had been involved with data-centre projects since around 2012, mostly from the build side.

I had walked through server rooms and construction sites. I had seen the electrical systems, cooling, fire protection, backup power and the many layers of infrastructure required to make something we casually call “the cloud” actually work.

The cloud, of course, was never really a cloud.

It was buildings.

So when conversations around AI began turning toward electricity, water and data centres, the issue did not feel abstract to me.

I knew there was a building behind the chatbot.

That led me to a simple question:

What actually happens, in resource terms, when I type a question into an AI model?

I made that the focus of my sustainability project.

I explored whether unnecessary AI output could be reduced without reducing the usefulness of the outcome. In a small pilot, output-token use fell significantly under the conditions we tested while participants were still able to reach useful outcomes.

The result interested me, but its limitations were probably more important.

A token is not a unit of electricity.

Fewer tokens do not translate neatly into the same percentage reduction in energy, water or emissions. Between the words appearing on my screen and the resources consumed sit model architecture, chips, servers, utilisation, cooling, data-centre design and the electricity system itself.

That made me look beyond the model.

Once I did, even a simple data-centre announcement began to raise a string of questions.

Why this location?

Where will the power come from?

Is there enough capacity?

Who is the customer?

Is the demand already there?

How much AI is actually being adopted, and how deeply?

Then the workforce question became personal.

Someone I knew was directly affected by layoffs associated with AI.

That made me more careful about the stories I was hearing.

Were companies genuinely replacing work with AI?

Or were restructuring, cost pressure and other factors sometimes being folded into a simpler story about AI replacing people?

The more I read, the less comfortable I became with any single explanation.

The flood of competing claims eventually pushed me to start tracking adoption and infrastructure for myself.

Not because the dashboard was the point.

I wanted to understand whether the different parts of the story were actually moving together.

And that was when the picture changed again.

AI was not one thing.

It was a chain.

Semiconductors.

Compute.

Data centres.

Construction.

Electricity.

Water.

Networks.

Software.

Corporate adoption.

Workers.

Consumers.

A change in one part can create consequences somewhere else.

Some industries may be disrupted.

Others may see enormous new demand.

Some communities may benefit from investment.

Others may carry more of the infrastructure or environmental cost.

What had started for me as a question about a chatbot had become something much larger.

AI was connected to chips, data centres, electricity, companies, workers and users.

It was not one industry.

It was a system.

And once I saw it that way, another question became harder to avoid.

The infrastructure behind AI was expanding quickly.

So was its presence in everyday life.

We were beginning to use it to write, analyse, search, decide, learn and solve problems.

That brought the question much closer to home.

If AI was becoming part of the way we do so many things, what might it do to the way we think?

That is the question I want to turn to next.

I want to start with the one that worries me most personally.

Thinking itself.