Fynsa
You've probably heard the word “token” used as a unit of measurement for consumption in conversations about artificial intelligence, and you’ve probably run out of tokens because you’ve used up too many of them with some of the most popular models. But what is a token , and why should you care?
Think of a taximeter that doesn't charge per exact kilometer, but rather in units that combine distance and time. The token is somewhat similar: it’s the unit the model uses to measure how much work it performed—both to read what was written to it and to construct its response. And we should care about this because the token is the unit that translates intelligence into cost. Each token has a price, and not all models are equally efficient in their use of it.
Given the above, how is the price of a token? To answer that, we need to think of AI as a production chain, with inputs that generate the results we ask of it. Jensen Huang, CEO of NVIDIA, sums it up in simple terms: AI is a five-layer cake.
At the bottom are the commodities and power generators needed to get the whole machine up and running. Moving up this structure, we find the semiconductors or chips that provide the computing power on which AI operates. These first two layers (energy and chips) are today’s picks and shovels, enabling others to strike gold.
The next stop is infrastructure: data centers, cloud services, fiber-optic cables, and cooling and maintenance providers. This is, so to speak, where artificial intelligence resides. Next come the LLMs—or trained models—the next step up. Here we find companies like OpenAI, Anthropic, and Google, which develop and train the systems that people and businesses interact with. This is where a real race is currently underway to develop proprietary models.
Finally, there are app and software developers—the practical side of the field that seeks to create solutions and charge for them. Here we find well-known companies, such as Spotify and Netflix, which use AI to improve their algorithms and optimize recommendations, or companies like SAP and Salesforce, whose business models are, in many cases, threatened by the growing adoption of AI that allows these services to be replicated on an individual basis, both by companies and by individuals.
Once we understand AI as consisting of five layers, we understand why the price of tokens depends on the cost of energy, water, land, chip wear and tear, and other factors throughout this entire chain.
But it doesn’t end there. Outside this structure are the rest of the people and companies that have incorporated AI into their daily processes, seeking to optimize sales or improve margins across various industries. These secondary beneficiaries are what make the thesis viable: if people and companies weren’t using AI, none of the above would be possible, and we’d be facing something that, rather than a solid trend, would be a bubble. For now, the pace of adoption has been almost as surprising as the advancement of the technology itself.
Our approach to capitalizing on this trend does not rely on a single solution, as there are multiple ways to incorporate this strategy into a portfolio.
The first and simplest option is to invest, for example, in the S&P 500. This allows you to buy shares in the 500 largest companies in the United States—an index that includes companies such as Google, NVIDIA, Microsoft, Amazon, Tesla, and Apple, all of which benefit directly or indirectly from this trend.
Another alternative—one that is slightly more complex to implement—is to select indices with exposure to at least one or more of the five layers described, thereby achieving a diversified approach without losing sight of the central objective.
For those looking for an option that is less spread out across so many stocks, it is possible to select specific companies with greater confidence that they will be the winners in “winner-takes-all” markets (“Winner takes all”), to reinforce the thesis.
In other cases, less experienced investors who might feel overwhelmed by choosing the best mix of options may prefer to delegate the selection of companies or sectors—or rotations among them—to a third party in order to better capitalize on this opportunity.
Over the next few months, we’ll see companies like Anthropic go public—and their IPOs will likely be oversubscribed amid this AI craze. This strategy is reminiscent of the dot-com era, when there were many companies pursuing the same goal, and 25 years later, only a handful remained. That’s why a combination of different “pure players” is usually a wiser move than betting on just one.
The solutions described above seek exposure to the entire AI value chain, but some investors may prefer to gain exposure to only part of it. Rather than trying to pick a winner, this strategy aims to capture value from those who supply the competitors: producers of commodities and rare earths, independent nuclear power or LNG producers, manufacturers of chips, semiconductors, and RAM, and providers of infrastructure, manufacturing, and maintenance for data centers and cloud services. This approach excludes developers of models, software, and applications.
The growing demand for tokens, the rapid pace of technological advancement, investments already made, and the projections of major companies for this year and beyond open up a wide range of opportunities. The key is not to chase the ultimate winner (which is constantly changing anyway), but to identify the industries that are integral to making all of this happen on a day-to-day basis.
Nelson Haase
FOS Senior Advisor at Fynsa
