Their revenue in the entire first half per the article was $4.3B, even if we assume it was backloaded in Q2 that’s a -500% profit margin. I’m very excited to review their S-1. Like, for every dollar you give them they spend $5 to deliver the product. They’re selling $10 bills for $2. They seem to have taken the joke about making up losses with scale pretty seriously.
What Theo claims on YouTube is that each model they make is profitable over its lifetime and the weird numbers for the company are because they keep training more models, having to pay up front to buy a bunch of future profit each time basically
The problem is that OpenAI will have to keep doing this basically indefinitely, as otherwise open-source commodity models will catch up with them and offer an equivalent product at a much lower cost. If the company is reliant on having to keep paying more than it's making on its current model to train new models, I don't see how it can ever become a sustainable business.