23/23 tldr: - selling tokens is a ~10 to 25% gross margin business, and that's for the labs that own the model and pay nothing for weights. minimax ~9% implied. Z ai 24.6% stated. a reseller serving someone else's open weights starts below that, and kimi k3 now takes up to 30% on top. - both labs spend 2.5x to 4x their api revenue on training. - volume is fickle. minimax lost 75% of its openrouter tokens in ten weeks. Z ai took 31% of the whole platform in six days with a free model. - price goes up when the model gets smarter. Z ai doubled its average api price while volume grew 40x. capability is the only pricing power. - the margin lives at the compute layer. minimax rents. Z ai runs 100,000 chinese chips and is buying hardware. both prepay gpu rent by the hundreds of millions. - inference only providers are real businesses (baseten ~$600M arr, morph $6M with 2 people) but they sit between commodity tokens and expensive gpus, with model labs starting to take a cut. - openai and anthropic will show the same shape at 20x the price. the s-1s will tell us if closed weights hold that price. if you liked this, let me know. dms open if you want to talk gpus, or the economics of the ai space!

we’re vc funded and i still spend half an hour trying to save $100 on a flight. i grew up in central europe and spending money is still uncomfortable, even when it would save me time. how did you get past this as a founder?