Documents: OpenAI and Anthropic told investors they could be profitable both with and without training costs — and that inference costs eat more than half of revenue

It has been reported that internal investor documents from OpenAI and Anthropic offered two faces of the same math: one that shows paths to profitability when expensive model-training outlays are treated separately, and another that paints a far tighter margin picture when training costs are included. It has also been reported that both firms disclosed inference—or serving—costs that exceed roughly half of their reported revenue, a striking line item that changes the whole story.

What the documents show

According to the report, companies presented investors with scenarios that isolate training as a one-time, capital-intensive phase and then project healthier ongoing margins during inference. In alternative scenarios that fold training expenses into operating results, the profit picture is weaker and slower to arrive. It has been reported that inference costs — the compute and cloud bills required to answer user queries — were disclosed as consuming more than 50% of revenue, a reality that undercuts the classic software margin playbook.

Why it matters

If serving models continues to be that expensive, scaling revenue won’t automatically translate into big profits. Investors tuning up IPO timelines will ask hard questions: can these firms lower per-query costs with model efficiency, custom chips, or better cloud deals? Or will expensive inference turn a high-growth narrative into a capital-hungry slog? Allegedly, those trade-offs were central to conversations in the investor decks.

This is the emotional crux: the AI boom promised software-like economics, but the numbers may look messier in the light of real-world compute bills. Can engineering wins deliver the margin rescue investors expect? Or is the industry about to learn that cutting-edge AI is brilliant — and expensive — for a long time to come.

Sources: wsj.com