Who Pays for Your AI? Following the Money Down the AI Ladder

“AI is making billions of dollars. So the next question in the AI boom is this. Who is taking the profit home with them?”

That’s the question I asked in my video on 22 September.

Is your AI subsidised?

A token is the small unit of text an AI model reads or writes, roughly a piece of a word. AI companies charge per token, or wrap tokens into a monthly subscription.

In the video I said the tokens being used are cheaper than they should be. The reason is obvious: getting as many customers as humanly possible, to see if there is real demand. Now, we’ve established that there is a demand. There’s an AI assistant for almost everything out there.

There’s evidence for that view. In January 2025, OpenAI’s CEO Sam Altman posted that the company was losing money on its $200-a-month ChatGPT Pro plan, because people used it more than expected. And the demand turned out to be real: OpenAI said it had reached $2 billion in monthly revenue in March 2026, and Anthropic, the company behind Claude, said its revenue was running at more than $47 billion a year in May 2026.

But the honest version is more nuanced. These companies spend enormous sums on training new models and building computing power. Whether each token is sold below cost is disputed. Bloomberg and the Financial Times have both reported that Anthropic made an adjusted operating profit in the second quarter of 2026.

Stage two: cheaper tokens or higher prices

Now we have entered the second stage of the AI boom. Somewhere costs have to come down, and there are two solutions to the same problem.

First, tokens can be made more efficiently. Instead of having to use 10 tokens to complete a task, you might now need five. Second, prices go up. If companies are not able to make models that are efficient, they have no other choice than to raise prices. Can customers afford that? It depends if the businesses that use them get enough out of it.

So far the first path is winning by a lot. The research group Epoch AI found that the price of reaching a given level of AI performance fell between 9 and 900 times per year, depending on the task. In June 2025 OpenAI cut the price of its o3 model by 80%.

As of 1 October 2026, neither OpenAI nor Anthropic is listed on the stock market. Both filed confidential draft IPO papers with the US regulator in June 2026.

The AI ladder, from top to bottom

The model. We have the top level. This is the AI tool. You go and ask your question, or do your work. Think of ChatGPT or Claude.

Cloud computing. Those models run on servers made to produce computing power, like Microsoft Azure or Amazon Web Services. In their latest quarter, AWS grew 37%, Azure 43% and Google Cloud 82%.

The data centre. Huge buildings, but more importantly is what’s inside them. Alphabet, Amazon and Meta alone plan to spend roughly $545–570 billion on capital investment in 2026, much of it for AI.

Chips. Mostly Nvidia, whose chips provide all of this computing power.

Racks, cables and power. Don’t only think about microchips. More importantly is also how you put them together. You put them in racks stacking next to each other. You need different kinds of cables. You also need electricity.

Power: the real bottleneck

A data centre without electricity is an expensive empty building. The International Energy Agency estimated that data centres used about 415 terawatt-hours of electricity in 2024 and could use around 945 by 2030.

Companies like Google are investing to create the power that they need. In 2024 Google agreed to buy power from small nuclear reactors by Kairos Power, with the first targeted for 2030. It also signed a deal to restart the Duane Arnold nuclear plant in Iowa, targeted for 2029.

Who profits first

Who is taking it all home? It starts at the bottom: Nvidia, selling the microchips. In the quarter that ended in July 2026, Nvidia’s revenue was $96.2 billion, up 106% on a year earlier. Its data centre business alone brought in $89.0 billion.

Look up the ladder and you’ll see that cloud computing is next.

What I’m doing

“I myself use AI almost every day for my work. It’s useful, but I don’t think it’s there yet.”

Microsoft and Google Cloud are in the right spot on the ladder, but I personally don’t want to invest in those companies because of ethical reasons.

Where I did put my money is Vertiv. I didn’t mention it in the video title; it was a hidden Easter egg. Vertiv builds the power and cooling systems inside data centres, plus racks and service work, with specialists who set up the technical systems that make a data centre run. In the second quarter of 2026 its sales grew 24%, and it expects about $14 billion in sales for the year.

In the video I said Vertiv has a backlog of two years. The last backlog figure in its earnings releases was $15.0 billion at the end of 2025. That’s roughly one year of its expected 2026 sales, not two. My reason for owning it hasn’t changed, but I want my numbers right.

My bet is that within the next five years tokens will become a lot cheaper: half or less of what they cost today. Is this guaranteed? No. And that’s why I put my bet on different layers of the AI infrastructure.

Always keep an eye out, because you don’t know which companies are overextending. Maybe too many data centres get built. Maybe AI is priced too cheaply. Maybe research costs keep rising. There’s a lot that can go wrong, but it is a technology of the future. And after the AI boom, I think we’ll see robotics coming into play.

Want the opposite view? Read The Devil’s Advocate version of this story.

Keep reading

• The War Was the Headline. The Bond Market Was the Story.
• Gold Is Falling. Here’s What the Big Money Actually Did

I’m not a financial adviser. This is my own investing journey, not financial advice. Any investment decision is your own.