Horton Hears a Hyperscaler: What a bedtime story taught me about AI investing

My kids and I have read Horton Hears a Who! more times than I can count. It’s become the go-to bedtime pick in our house, partly because the story is wonderful and partly because it clocks in at about 30 minutes flat. My wife recently confessed that she reads every third page when it’s the kids’ turn to choose. Well played, dear.

For anyone who hasn’t had the pleasure of reading the book approximately 400 times, Horton the elephant discovers an entire civilization living on a tiny speck of dust. No one else can hear the Whos, and because the other animals can’t see or hear them, they assume the Whos don’t exist. Horton refuses to accept that conclusion. His famous refrain is that “a person is a person, no matter how small,” and the story ultimately becomes a lesson about recognizing value even when it is easy to overlook.

Recently, I found myself thinking about Horton while trying to make sense of AI investing, specifically, the enormous investment required to support artificial intelligence. In a conversation on the Invest Like the Best podcast, Patrick O’Shaughnessy asked investor Gavin Baker, founding partner and chief investment officer of Atreides Management, to explain how he thinks about the economics of AI as increasingly capable models become cheaper and more widely available. Baker’s answer was memorable because of how simple it was: “a token is a token.”

The more I thought about it, the more useful I found the idea. Not because it tells us which AI company will ultimately win, but because it offers a framework for thinking about where value may actually reside as the technology evolves.

As AI gets cheaper, what happens to the infrastructure?

One of the most important developments in artificial intelligence isn’t simply that models are becoming more capable. It’s that the cost of using them continues to fall. Open-source models, improvements in computing efficiency, and intense competition among AI developers are making sophisticated AI increasingly accessible.

A token is a token

To understand why that matters, it helps to understand what a token actually represents.

A token is essentially a small unit of text that an AI model processes or generates. Depending on the model, a token might be a whole word, part of a word, punctuation, or another piece of information. Every time an AI system answers a question, summarizes a document, writes code, analyzes information, or performs another task, it’s processing tokens.

And regardless of which model generates those tokens, producing them requires physical resources. Computing power, memory, and electricity are still required. The infrastructure doesn’t particularly care whether the token came from an expensive, proprietary frontier model or a less expensive open-source model.

The economics of the businesses can be very different. One company may charge a significant premium for access to its model, while another may offer a model at a fraction of the cost. Their margins, competitive positions, and valuations may have very little in common. But underneath those differences is a basic physical reality: someone still has to provide the computing power necessary to process the token.

“A token is a token.”  — Gavin Baker, Atreides Management

If the price of an individual token falls, that doesn’t necessarily mean the infrastructure required to produce tokens becomes less important. It may simply mean that more tokens can be produced and consumed economically.

Cheaper AI could mean more AI

This is where the economics get particularly interesting.

We’ve seen this pattern before. As computing power, data storage, and internet bandwidth became cheaper, consumption didn’t decline. Instead, those technologies became embedded in more of the economy. Activities that once seemed too expensive or impractical become routine because the underlying technology is suddenly affordable enough to support them.

This phenomenon is often referred to as the Jevons paradox: The idea is that improved efficiency can increase total consumption because lower costs encourage greater use.

AI could follow a similar path. If the cost of generating and processing tokens continues to fall, businesses may not simply use the same amount of AI for less money. They may use AI in places where it previously wasn’t economical at all.

A company might analyze millions of documents rather than thousands. A software developer might use AI throughout the development process rather than occasionally asking it to write a piece of code. A business might incorporate AI into customer service, research, operations, and products in ways that would have been difficult to justify when the technology was more expensive.

That distinction matters for the infrastructure supporting AI. The question isn’t simply whether an individual token becomes cheaper. It’s also whether the number of tokens the world wants to consume grows enough to offset that decline in price.

Where does the value go?

This brings us to a broader question about investing in emerging technologies: if the economics of an industry change, does the industry itself become less valuable, or does the value simply move somewhere else?

Consider what could happen if AI models become increasingly commoditized. Competition could drive prices down, making it harder for model developers to maintain the extraordinary margins investors may currently associate with the technology. That could be a meaningful challenge for companies operating at the model or application layer.

But lower margins at one layer of the technology stack don’t necessarily eliminate demand at the layers underneath it.

If cheaper AI leads to dramatically greater adoption, someone still needs to provide the computing power. Someone needs to manufacture the semiconductors, build the data centers, provide the networking equipment, supply the memory, and generate and transmit the electricity required to keep everything running.

The economics may shift away from the companies selling the intelligence toward the companies providing the infrastructure around the intelligence.

That doesn’t mean every semiconductor company, data center operator, or utility will benefit equally. Nor does it mean today’s enormous investment in AI infrastructure will necessarily produce attractive returns. Capital can be misallocated, technology can become obsolete, and investors can overpay for a good story.

It simply means that when we evaluate an emerging technology, we shouldn’t assume that the company with the most recognizable brand is necessarily the only place where economic value is being created.

What Horton gets right about investing

And this is where my children’s bedtime story comes back into the picture.

The animals in Horton Hears a Who! make a fairly understandable mistake. They judge the Whos based on what they can see and hear. Because the Whos are tiny and invisible to them, they conclude that the Whos aren’t important. Horton, on the other hand, understands that their size and visibility have nothing to do with whether they exist or whether they have value.

Investors can make a similar mistake when evaluating new technologies. We naturally gravitate toward what is visible: the companies with the biggest brands, the most recognizable products, the most impressive demonstrations, and the most attention from the market. Those things can be important, but they can also obscure the less visible pieces of the economic system that make everything else possible.

AI makes a great example: We interact with the model, so it’s natural to focus on that part of the technology. But, behind every AI interaction is a physical infrastructure we don’t see: chips, memory, networking, data centers, and energy. Those resources don’t disappear when one model becomes cheaper or when another model becomes more popular.

In fact, greater competition at the model layer could ultimately make the infrastructure more important by allowing AI to spread into more applications and reach more users.

The bigger lesson for AI investing

I don’t think anyone can say with certainty how the AI investment story will unfold. There are enormous questions still to be answered about competition, capital spending, technological obsolescence, energy availability, utilization, pricing, and ultimately how much AI the world will actually use.

But uncertainty is precisely why I find Baker’s framework useful. It encourages us to separate the story we’re hearing about an industry from the underlying economics of that industry.

That’s a useful discipline well beyond AI. Markets often encourage us to use labels as shortcuts for value. A company is considered “the AI company,” so we assume it will capture the value created by AI. A product is expensive, so we assume it must be more valuable. A technology becomes cheaper, so we assume demand must be weakening.

Sometimes those assumptions are correct. Sometimes they aren’t.

The more useful question is often what lies underneath the headline. What’s actually being consumed? What resources are required to produce it? Where does the money flow through the system? And if one part of the value chain becomes more competitive, which other parts might benefit from the resulting increase in adoption?

That’s the lesson I took from “a token is a token.” It isn’t a prediction about which AI company will win, and it isn’t an argument that AI infrastructure investments are guaranteed to succeed. It’s simply a reminder that the most visible part of an industry isn’t always the part that captures the most enduring economic value.

Horton understood that a long time ago: A person is a person, no matter how small.

And perhaps, when it comes to AI, a token is a token, no matter which model produced it.

Technology can change quickly, but the principles we use to evaluate investment opportunities don’t have to.

Connect with us if you’d like to discuss how developments in AI and other emerging technologies fit into your broader financial plan.