Artificial intelligence

When people say "AI," they usually mean software that has been trained on a huge pile of text, images, or other data — and can then answer questions, write things, recognize pictures, or recommend what you might like.

It feels new, but the recipe is decades old. What changed in the last few years is that the computers got fast enough and the training data got big enough for the answers to suddenly feel useful. ChatGPT in late 2022 was the moment everyone noticed.

AI doesn't come from one company. It comes from a chain of them, each doing one step. Almost every company that touches AI sits somewhere on this chain.

If you want to invest in "AI," the first question is which step you're betting on — the chip makers, the cloud, the model builders, or the apps people actually use.

There are roughly three ways AI companies make money today. Selling chips (NVIDIA, AMD). Renting cloud computers that run AI workloads (Microsoft Azure, Google Cloud, Amazon AWS). And charging subscriptions or API fees for the models themselves (OpenAI, Anthropic, and many smaller players).

Right now, by far the biggest profits go to the chip makers and the cloud providers. The model builders take in enormous revenue but also spend enormous amounts on training and computing — so their profits are still small or negative.

Training a top-tier AI model can cost more than $100 million in compute alone. Running it after training — every time someone asks a question — also costs real money, because each answer needs powerful chips for a few seconds.

This means the economics look more like an electric utility than like a typical software company. Lots of upfront capital, heavy ongoing energy bills, and you need millions of paying users to make it work.

Traditional software companies make a product once and sell it many times — costs barely grow as users grow. That's why software margins are famously fat.

AI is different. Every answer costs the company money in real time. As users grow, costs grow with them. Margins are real but thinner, and they depend heavily on how cheap the underlying chips and cloud get.

AI investments share a few specific risks. Demand could cool if the early hype outpaces what models can actually do reliably. Regulation could limit what AI can be used for (copyright, deepfakes, hiring decisions). And the chip supply chain is concentrated — most leading AI chips are designed by NVIDIA and physically made by one company in Taiwan, TSMC.

On the upside, almost every industry is finding ways to use AI, so the long-term demand for AI compute looks durable even if individual companies stumble.