AIght_
ToolsCompareLearnUniverseWorkflowsSignalHuman
Take the quiz
← All concepts

Concept

GPUs & Compute

Why AI progress is measured in chips and megawatts.

Moon·Updated Jun 10, 2026
◈ Learn it interactively8 steps · ~5 min →

Where this idea lives

PREREQUISITESGPUs & ComputeBackpropagationBackpropagation — How blame gets assigned through a network.Scaling LawsScaling Laws — Why bigger keeps working — and the question of where it stops.QuantizationQuantization — Why a 70B-parameter model can run on your laptop — and the quality you trade for it.Common misconception: GPUs are for graphics; AI must need something else.Common misconception: Software cleverness matters more than raw compute.
prereqsrelatedtoolsmisconceptions
You might think:GPUs are for graphics; AI must need something else.Software cleverness matters more than raw compute.

Common misconception

“Software cleverness matters more than raw compute.”

The bitterest lesson of the field's history — literally titled "The Bitter Lesson" in a famous 2019 essay — is that general methods plus more compute have beaten human cleverness over and over. The clever ideas that won are mostly the ones that let compute scale. Hardware isn't the boring part of the story. It's the plot.

A neural network's life is multiplication: multiply inputs by weights, sum, repeat — billions of times per training step, trillions per second at frontier scale. None of these multiplications is hard. There are just unspeakably many, and most don't depend on each other, so they could in principle all happen at once.

That word — at once — is why the AI era runs on a chip designed for video games.

CPU vs GPU, in one image

A CPU is a few brilliant workers: a handful of cores, each fast and flexible, built to race through tasks where step two needs step one. A GPU is a stadium of ten thousand interns who can each only multiply — useless for running your operating system, perfect for painting three million pixels at once. Researchers around 2010 noticed that "update every pixel" and "update every neuron" are the same shape of problem. Deep learning's breakout moment — AlexNet, 2012 — was a neural network that wedged itself onto two gaming GPUs and embarrassed two decades of hand-crafted computer vision.

Modern AI chips (NVIDIA's data-center GPUs, Google's TPUs) are that insight industrialized: silicon specialized for matrix multiplication plus enormous memory bandwidth — because feeding billions of weights to the cores fast enough is usually the real bottleneck, not the math itself.

Compute as the field's currency

Training compute is measured in FLOPs (floating-point operations), and frontier training runs are measured in months of thousands of GPUs — costing hundreds of millions of dollars in chips and electricity. This single fact explains much of the industry's shape: why a handful of companies train frontier models and everyone else fine-tunes them, why export controls on chips are foreign policy, why data centers negotiate with power grids, and why quantization and distillation — doing more with less compute — are entire subfields.

When you read AI news, follow the compute. It predicts more than the press releases do.

What to read next

You have the full stack now — data, learning, networks, hardware. Next: the architecture that turned all of it into the modern era. Transformers.

◈ Check yourself

You just read it — now see if it stuck. No score, no penalty for a wrong pick.

Question 1 of 3

Why are GPUs central to modern AI?

Question 2 of 3

Why is a GPU better than a CPU for training neural networks?

Question 3 of 3

What does it mean that compute has become a strategic resource in AI?

← Prerequisite

Backpropagation

Next →

Transformers

by Moon · poet · engineer · third-year CS, Chandigarh University · Updated Jun 10, 2026

← Back to all conceptsBrowse tools →
beginner
Read time5 min read
UpdatedJun 2026
Sources3

Read next

  1. Scaling Laws →
  2. Quantization →