Keeping up with AI sometimes feels less like following technology and more like learning a new language.
Foundation model. Inference. LLMs. Distillation. Open weight models. Agentic AI. Generative AI. GPUs. TPUs. CPUs. Tokens. Prompt engineering. Vibe coding.
The funny part is that whenever a new AI term pops up, I usually use AI to explain it to me. It seems fitting that AI has become the instruction manual for understanding…AI.
The good news is that AI isn’t nearly as complicated as the headlines make it sound. Most of the jargon falls into four simple buckets. Once you understand those four pieces, you’ll have a much better idea of what’s actually happening and why investors care.
Every AI system starts with computing power.
Companies like NVIDIA, AMD, and Broadcom design the specialized chips that power AI. The most common are GPUs (Graphics Processing Units), which have become the workhorses of modern AI because they excel at performing many calculations simultaneously. You’ll also hear about TPUs (Tensor Processing Units), Google’s custom-built AI chips that serve a similar purpose inside many of its own products. While GPUs do most of the heavy lifting, CPUs (Central Processing Units) are expected to become increasingly important as agentic AI takes on more complex tasks.
Think of these companies as supplying the engines that power the entire AI ecosystem.
Modern AI requires hundreds of thousands of these chips working together. That’s why companies like Microsoft, Amazon, Alphabet, and Oracle, often referred to as hyperscalers, are investing hundreds of billions of dollars in massive data centers. Those facilities provide the computing power needed to train AI models and process millions of AI requests every day.
This is where most of the new AI terminology comes from.
Companies like OpenAI, Anthropic, and Alphabet have developed AI models such as ChatGPT, Claude, and Gemini. Under the hood, they’re all powered by large language models (LLMs), which are a type of foundation model. Think of a foundation model as a general-purpose AI that serves as the starting point for many different applications, from writing emails to generating software code.
Despite how intelligent they sometimes seem, these models aren’t actually thinking like humans. Instead, they’re continually predicting what piece of text is most likely to come next based on everything they learned during training. They repeat that process thousands of times in a fraction of a second to generate a response.
Once a model has been trained, using it to answer questions is called inference. Training is the expensive process of teaching the model. Inference is what happens every time you interact with ChatGPT or another AI application.
Some companies, like OpenAI and Anthropic, keep their models proprietary. Others, like Meta, release open weight models, allowing developers to build on them and create their own AI applications.
You’ll also hear the term distillation. As larger AI models become more capable, they can be used to teach smaller models to perform many of the same tasks. Those smaller models are faster, cheaper, and require much less computing power.
Much of that cost is measured in tokens, which are simply small pieces of text an AI processes whenever you ask a question or receive a response. As models become more efficient, they require less computing power per token, making AI less expensive to use.
Another buzzword is agentic AI. Unlike today’s chatbots, which mostly answer questions, agentic AI is designed to complete multi-step tasks on your behalf, such as researching a topic, comparing options, creating a presentation, or completing an entire workflow with minimal human involvement.
Finally, there’s generative AI, often shortened to GenAI. It’s simply a broad term for AI that creates new content, whether that’s text, images, video, music, or computer code. And if someone mentions prompt engineering, they’re talking about the art of writing better instructions to get better results from an AI model.
This is where AI starts creating real economic value.
Most companies will never build their own AI model. Instead, they’ll use existing models to improve their products and make employees more productive.
Whether it’s writing emails, analyzing spreadsheets, discovering new drugs, detecting fraud, improving customer service, or helping developers write software, this is where businesses will see the greatest practical benefits from AI.
One of the newest buzzwords is vibe coding, where developers describe what they want in plain English and AI writes much of the code for them. It’s a great example of how AI isn’t replacing people so much as helping them become more productive.
Now the next time you read an AI headline or hear someone mention distillation, agentic AI, or foundation models, you’ll know exactly what they’re talking about.
From an investment perspective, knowing what the buzzwords mean is helpful. Figuring out where the greatest long-term value will ultimately be created is much harder.
Will it be the companies designing the chips? The hyperscalers building the infrastructure? The firms developing foundation models? Or the software companies embedding AI into everyday products?
History suggests transformative technologies rarely create just one winner. The internet generated enormous value for semiconductor companies, networking firms, cloud providers, software developers, and entirely new businesses that didn’t exist before.
AI will likely follow a similar path. Rather than trying to predict a single winner today, investors are probably better served by understanding the entire ecosystem and how the companies within it fit together.
Top Row L to R: Brad Engle, Mike Sullivan, Sebrina Ivey, Christian Lewton, Jason Kitner
Bottom Row L to R: Carin Wagner, Angela Kennedy Lee, Jenny Merges, Brian Friedman, Deirdre Mcguire, Barbara Terrazas, Reed McCoy