The AI boom has introduced an entirely new vocabulary. Just a few years ago, terms like OpenAI, Anthropic, hyperscalers, GPUs, and compute were mostly confined to niche tech circles. Today, they dominate headlines and boardroom conversations.
Most of us now have a basic grasp of what AI can do, yet the engine behind it remains a black box. How do the model makers, the massive compute providers, and the underlying hardware actually fit together to make it all run? Let’s keep things simple and stick to the basics.
- Arguably the two largest models, OpenAI and Anthropic, which respectively created ChatGPT and Claude, are what we can accurately describe as the ‘brains’ of AI.
These companies spend billions of dollars researching models that can understand language, analyze data, reason through problems, generate code, and much more. Without these models, AI data centers are simply expensive buildings packed with tech hardware.
Think of it like this: a personal computer without an operating system like Windows is nothing more than an expensive paperweight. Data centers need the software (the brains) to run all that pricey hardware.
- So, what are hyperscalers, which build data centers? These are massive tech firms that include Microsoft (MSFT, $524, Azure), Alphabet (GOOG/GOOGL, $293/$296, Google Cloud), and Amazon (AMZN, $267, AWS). They own and are continually building data centers around the globe.
OpenAI and Anthropic need hyperscalers, and the hyperscalers need OpenAI and Anthropic. For example, OpenAI primarily runs on Microsoft’s Azure cloud, while Anthropic is a major customer of Amazon AWS and also uses Google Cloud.
- In addition, training AI models requires enormous GPUs (graphic processing units).
Originally created to render 3D graphics for video games, GPUs have become the primary engine behind modern AI and scientific computing because of their ability to process massive amounts of data simultaneously.
While other tech giants are hoping to break into the GPU market, the dominant maker today is Nvidia (NVDA $163).
- Then there is compute. Compute is the raw processing power used to train and run AI models. The greater the compute, the more calculations AI can perform per second, enabling larger models and greater capabilities.
- And one last thing. Some view AI as a glorified search engine. But it’s more than that.
If you ask a search engine to explain something like Gross Domestic Product (GDP), it will reference indexed articles. But if you ask AI to explain GDP to a 12-year-old or a graduate student in economics, it must reason and explain in simple or complex terms rather than simply retrieve.
In short, OpenAI and Anthropic create the AI models that people and businesses want to use, while hyperscalers provide the massive computing infrastructure that makes those products possible. Each side helps justify the other’s enormous investment.
Why investors care
A major debate among investors is whether hyperscalers are spending too much on AI infrastructure, which is fueling sales in an enormous supply chain.
The bullish view: OpenAI, Anthropic, and other AI companies will create enough demand to fill all these new data centers, and AI becomes the next cloud-computing boom.
Today, things like the Internet, streaming, smartphones, and apps have been seamlessly integrated into our lives, generating trillions of dollars in economic output. Businesses that are investing and financing the AI buildout see it as a foundational technology, like the development of electricity and the Internet.
Consumers might hesitate to pay for various AI subscriptions, but hyperscalers are initially banking on revenue from various enterprises, both large and small.
The bearish view: Hyperscalers are building capacity faster than profitable AI applications are emerging. If demand fails to materialize, some data centers and GPUs could end up underutilized.
While the technology itself is clearly moving forward, history shows that massive infrastructure booms, like railroads in the 1800s and telecom in the 1990s, may create overcapacity that can shatter initial profit projections.
But, when early investors lost money in the railroads, the tracks didn’t disappear. That overbuilt network slashed the cost of moving goods and became a critical backbone of the US economy for the next century.
Massive amounts of unused fiber in the early 2000s were eventually absorbed. It’s the backbone of today’s Internet.
Today’s hyperscalers are making a similar bet: whether every dollar spent on AI infrastructure pays off right away, there is a fear of missing out, as whoever controls the tracks will power the next era of human commerce.


