Nvidia worth more than countries: the hidden cost of AI

Nvidia, the undisputed queen of graphics cards, has reached a market valuation higher than almost any country. This isn't random; it's a direct result of our reliance on its GPUs for artificial intelligence. Without Nvidia's chips, AI progress would drastically slow down, and everyone knows it.

Training an artificial intelligence model, like those answering your questions or creating images, is a ridiculously expensive endeavor. It requires thousands of GPUs working in parallel for months. For instance, Meta's Llama 2 model needed 6,000 GPUs and millions of dollars just for training. It's like building a power plant just to teach a computer to talk.

The core issue is the cost of computing power. We simply can't produce chips cheap or efficient enough to handle AI workloads sustainably. The original "The Matrix" analogy wasn't about humans as batteries, but human brains as super-efficient GPU clusters. Pretty brilliant metaphor, right?

A human brain consumes about 20 watts to do things that even the best AIs can't dream of. An AI model's GPU cluster can consume megawatts. The difference is staggering. If we could replicate our brain's efficiency, the cost of AI would plummet.

For now, alternatives are scarce, and none are magic wands. Researchers are looking for new chip designs, different architectures, or smarter ways to train AI with less data and energy. But it's a long road. OpenAI, for example, has invested billions in developing increasingly larger, resource-hungry models.

This means artificial intelligence, for now, will remain an expensive toy, reserved for a few large corporations. The GPU gold rush will continue, and with it, Nvidia's fortune. However, whoever manages to "do more with less" in hardware could be the next to reshape the game.