"Our GPUs are melting": who gets an image on launch day?

transcript

March 2025. Sam Altman posts: our GPUs are melting. Image generation had just reached everyone. So who actually gets an image?

Say a pool of 8,000 GPUs. 16 GPU-seconds per image. That's 500 images a second. Demand: five times that. 2,500 a second. Just queue it? After one hour, 7 point 2 million people waiting.

First, take what's cheap. Pause background work. 4,000 more GPUs in about 10 minutes. Paid users send 20 percent. 500 a second, full quality, on the original 8,000. Free users get a fast mode. Half the GPU time per image. 500 a second.

But free users send 2,000 a second. Fast mode serves 500. That's a quarter. So what do the other three quarters get?

The truth, before they wait. Free accounts get 3 images today, in fast mode. Every request shows an honest ETA, from a bounded queue. Past the allowance? At capacity. Back at 19:00 UTC.

What it never says: something went wrong. Silent failures just come back as retries. More load, not less.

And chat runs on its own pools. The launch can't slow your chat replies. When demand beats capacity, fairness is a choice. Make it, and say it out loud.

Full breakdown in the lecture.