The GPUs Are Melting

OpenAI added a million users in an hour this week. The GPUs are, in Sam Altman's words, "melting."

Not from some frontier research breakthrough. Not from a leap toward artificial general intelligence. From anime fan art.

On Tuesday, OpenAI launched native image generation in ChatGPT, powered by GPT-4o. By Thursday, the entire internet had decided the only reasonable use for this capability was turning everything — selfies, memes, pets, historical atrocities — into Studio Ghibli-style illustrations. The dreamy watercolors of Spirited Away, mass-produced at the speed of a text prompt.

Altman, to his credit, seemed genuinely delighted. He updated his own profile picture to a Ghibli rendering. Then he announced rate limits. Free users get three generations per day. The company is "working on making it more efficient." The classic Silicon Valley sequence: launch a product that creates infinite demand, discover your infrastructure can't handle infinite demand, ration access while calling it a temporary measure.

This is the oldest pattern in industrial economics, replaying in silicon. Build the factory before you've secured the supply chain. Create desire at a pace your fulfillment can't match. Then act surprised when the shelves are empty.

OpenAI spent years building language models. They spent months integrating image generation. They apparently spent approximately zero time modeling what would happen when 200 million users all wanted to generate pictures at once. A million new users in a single hour — compared to the five days it took the original ChatGPT to hit that mark in 2022. Nobody at OpenAI thought to ask what happens when the viral moment arrives and the demand curve goes vertical?

The infrastructure math is brutal. Every image request burns through GPU clusters that cost thousands per hour to operate. Multiply that by a usage spike that added a million users in sixty minutes, and you've got a company running its compute at redline because everyone discovered they prefer Ghibli aesthetics to productivity tools.

Which brings us to the other pattern hiding in the noise: whose aesthetic is everyone consuming?

Hayao Miyazaki, co-founder of Studio Ghibli, watched an AI-generated animation in 2016 and called it "an insult to life itself." He said he would never wish to incorporate the technology into his work. Nine years later, millions of people are using that exact technology to replicate his studio's visual signature — the lush backgrounds, the dreamlike color palettes, the expressive character work that took decades of human craft to develop.

OpenAI, threading the needle with the precision of a company that has lawyers, now refuses to generate images in the style of living individual artists. But studio styles? Fair game. The distinction is legally convenient and aesthetically meaningless. The whole point of Studio Ghibli's look is inseparable from the artists who created it. But when you can replicate the aesthetic at scale for free — until the GPUs melt, anyway — the legal nuances become a rounding error.

So here's where we are: a company valued at over $300 billion built the machine that creates the desire, can't fulfill the desire it created, and is rationing access to a feature that reproduces an art style whose creator considers the technology an insult to life itself.

The GPUs aren't melting from the weight of human ambition. They're melting from the weight of a system that learned to generate demand faster than capacity. Every industrial revolution has this moment — the gap between what you promised and what you can deliver, papered over with rate limits and cheerful blog posts.

Altman says the limits "hopefully won't be long." Start the timer.

Sources:

• Sam Altman says ChatGPT's Studio Ghibli-style images are 'melting' OpenAI's GPUs — Fortune, 2025-03-28
• OpenAI's GPUs "melt" as AI Ghibli art surge hits 1M users in an hour — Interesting Engineering, 2025-03-28