☀️ AI Morning Minute: Compute
AI runs on math. Math runs on hardware. And hardware has become the most fought-over resource in tech.
When people talk about what separates the companies winning in AI from the ones falling behind, they usually land on the same word: compute.
Not algorithms, not data, not talent. Compute.
It’s the raw processing power that trains AI models and runs them once they’re built. The more you have, the bigger the models you can train, the faster you can iterate, and the more users you can serve. It sounds simple. Behind it is a global supply chain, a geopolitical flashpoint, and a cost structure that shapes every decision in the industry.
What it means
Compute (short for computational resources) refers to the processing power used to train and run AI models, measured in operations per second and delivered through specialized chips called GPUs (graphics processing units) or newer AI-specific chips like Google’s TPUs and NVIDIA’s H100s.
Training a large model requires running billions of calculations across thousands of chips simultaneously, sometimes for months. Running the model after training, called inference, requires compute too, just in smaller, faster bursts. Compute isn’t infinite or free.
It’s physical hardware, it lives in data centers, it draws enormous amounts of power, and it costs real money to build and operate.
Why it matters
The amount of compute used to train frontier AI models has roughly doubled every six months for the past decade. GPT-4 required an estimated $100 million in compute to train. Models at the frontier today cost more. That trajectory is why compute access has become the clearest dividing line between labs that can compete and labs that can’t.
Compute is now a geopolitical asset. The US government restricts exports of advanced AI chips to China, and China is spending billions to build domestic chip capacity in response. NVIDIA’s stock price moves on AI policy decisions. The chip supply chain runs through Taiwan. None of that is incidental.
Cost shapes what gets built. A startup with limited compute budget makes different product decisions than a lab with tens of thousands of GPUs. Techniques like quantization and model distillation exist largely to squeeze more out of less compute. The economics of AI are, at bottom, the economics of compute.
Simple example
Think of compute the way you’d think of electricity in the early 20th century. The companies that controlled power generation controlled which industries could scale. Right now, a handful of cloud providers, Amazon, Google, and Microsoft, control most of the world’s AI compute. Renting time on their hardware is how most AI products get built. That concentration is already shaping who wins.