AI compute is the processing power needed to train and run artificial intelligence models.

Think of it like electricity. You don't think about kilowatts when you turn on a light, but behind that switch is a whole system of generation, transmission, and metering. AI compute works the same way — except instead of powering a light bulb, it's powering the math behind ChatGPT, image generators, self-driving systems, and every other AI product.

Practically, "compute" means renting time on specialized chips — mostly GPUs (graphics processing units) — that do the enormous number of calculations AI models require.

Why GPUs, Specifically?

Regular computer chips (CPUs) are built to do many different tasks one after another, quickly. AI models need something different: doing the same simple calculation millions of times, all at once. GPUs were originally built for video game graphics, which requires exactly that — updating millions of pixels simultaneously. It turns out that's also the ideal shape for AI math.

Nvidia makes the vast majority of these chips today, which is why Nvidia's earnings reports have become one of the most closely watched events in the entire stock market — they're seen as a read on how much AI compute the world actually needs.

Training vs. Inference: The Two Types Of Compute Demand

This distinction matters for almost everything you'll read about AI compute:

• Training is the process of teaching a model in the first place — feeding it enormous amounts of data so it learns patterns. This is extremely compute-intensive and can take weeks or months on thousands of chips running at once.
• Inference is what happens every time you actually use a model — ask it a question, generate an image, get a recommendation. Each individual request is cheap, but multiplied across billions of users, it adds up to a massive, ongoing compute bill.

Training is a big, one-time (or periodic) cost. Inference is a smaller but constant, growing cost. A lot of the current debate about whether there's a compute "glut" or a compute "shortage" comes down to which of these two you're measuring.

Who Needs Compute, And How Do They Get It?

Four main types of buyers:

• AI labs (OpenAI, Anthropic, Google DeepMind, xAI) — need compute to train and run their models
• Enterprises — companies building AI features into their own products
• Cloud providers (AWS, Microsoft Azure, Google Cloud) — buy chips at scale and rent out access
• Neoclouds — newer, AI-specialist cloud companies (CoreWeave, Nebius, Lambda) that exist almost entirely to rent out GPU capacity

Most AI labs don't own their own chips outright. Instead, they sign long-term contracts to reserve capacity — similar to how an airline might lock in a multi-year jet fuel contract instead of buying oil on the spot market every day.

Why Compute Suddenly Became A Business, Not Just A Cost

For years, compute was purely an expense line for tech companies — something you bought, not something you sold. That's now changing. Some of the biggest buyers of compute have built so much capacity that they have more than they currently need, and they're starting to sell the excess to others. That shift — from compute as a cost to compute as a product — is the single biggest reason it's suddenly in the news.

It's also why Wall Street has gotten interested: if compute can be bought and sold at scale, with prices that move up and down based on supply and demand, it starts to look like a commodity — something tradeable, hedgeable, and investable, the same way oil, wheat, or gold are.

Why Compute Isn't Quite Like Oil (Yet)

A barrel of oil is standardized — a barrel from one producer is essentially interchangeable with a barrel from another. Compute isn't like that yet. A GPU's real-world value depends on which chip generation it is, how much memory it has, how it's networked to other chips, where the data center is located, and how much of that capacity is actually being used at a given moment. Turning that into a single tradeable price is a genuinely hard problem — one that several new companies are racing to solve right now.

Conclusion 

Because compute increasingly gets bought, sold, and resold, a small industry has sprung up to:

• Track and benchmark prices across cloud providers and GPU marketplaces
• Build futures markets, so buyers and sellers can lock in prices in advance instead of guessing
• Package investment products (like ETFs) so ordinary investors can get exposure to the trend, not just the AI companies themselves

None of this changes what compute is — it's still just processing power for AI. What's changing is how that processing power gets priced, packaged, and traded, and who gets to profit from owning it versus just using it.