Question for the MBAs in the audience, is losiing $21 billion a year bad?

 Ed Zirton has a scoop:

Today, I can exclusively report, based on audited financial documents viewed by this publication that have been independently verified by the Financial Times, that OpenAI lost around $38.5 billion in 2025, as well as other crucial details about the financial condition of the company. 

 ...

2025 — OpenAI Had $13.07 Billion In Revenue, $34 Billion In Costs and Expenses, and $20.92 Billion In Losses, with a net loss attributable to the company of $38.53 Billion

• Revenue: $13.07 billion
• Cost of Revenue: $7.5 billion
• Research and Development: $19.18 billion
• Sales and Marketing: $5.73 billion
• General and Administrative: $1.57 Billion
• Total Costs and Expenses: $34 billion
• Loss from Operations: $20.92 billion

Please note that 2025 was the year that OpenAI converted from a non-profit to a for-profit entity, leading to a $41.55 billion loss due to changes in fair value of convertible interests and warrant liability. 

Taking into account other minor factors like interest income and interest expense, OpenAI is left with a net loss of $60.35 billion, which it lowered to $38.53 billion by removing $17.87 billion in costs via that “net loss attributable to noncontrolling members capital” and another $3.95 billion via a “net loss attributable to redeemable noncontrolling interests.” 

Ultimately, the net loss attributable to OpenAI in 2025 was $38.5 billion. 

[Tje conversion from non-profit to for-profit is a one time thing. Focus on the Loss from Operation.] 

I'm not a finance guy—most of you can probably read these numbers better than I can—but I have been following the AI bubble closely (arguably too closely for my mental health), and I can tell you that not only is it bad when a company loses over $30 billion in a year, but the specifics of how OpenAI lost this money are even worse, directly contradicting some of the main assumptions supporting a bubble many times larger than the dot-com bubble.

It's useful to think of these costs in terms of training and inference, and possibly to break down training into training and post-training. Training is where the models are built. Inference is where the models are used to answer questions. Though corporate accounting can be murky, you can generally think of training as falling under research and development and inference as falling under cost of revenue.

One of the key assertions of the AI bulls is that while these companies do lose money (even the most enthusiastic and simple-minded booster will concede that point), they are making money on inference, and since training costs the same regardless of how often the model is used, while inference scales on a per-user basis, the economics should continue to look better as demand for AI increases.

But even if you completely eliminated R&D spending, OpenAI would still have lost money in 2025. Now it's true that the shift to token-based billing will help—it might even tip the training-excluded numbers into the black—but OpenAI is committed to spending more than a trillion dollars over the next four years, and in order to honor those commitments, it has to become not just marginally but enormously profitable in the very near future. Looking at these numbers, it's difficult to see where you get that kind of money.

Of course, the really eye-popping number is the more than $19 billion spent on R&D, which mostly translates to training and post-training. Any way you look at it, these are deeply disturbing numbers for anyone invested in the viability of OpenAI as a business, but things get much more disturbing when you consider the context.

GPT-4, the model that revolutionized the field and permanently changed the way all of us think about natural language processing and the way we interact with computers, cost a fraction of what any of the 5-series models cost. Hyperscalers and AI boosters are extraordinarily adept at coming up with self-serving benchmarks that put each new model in the best possible light, but even there, it's next to impossible to argue that the graph of capability to investment is not looking more and more like an S-curve.

This is hardly surprising given the state of the training data and the increased reliance on post-training. Large language models have already revolutionized quite a few fields and have provided us with powerful tools for a number of applications. This can hardly be called a failed technology at this point (though it is very possible that something better will push it into obsolescence in the near future).

But for the foundational narrative of the AI bubble, this is cataclysmic. The tens of trillions of dollars that OpenAI, Anthropic, Meta, Google, and particularly xAi/SpaceX are promising to pump into the economy in the next decade are based on visions of exponential curves and recursive self-improvement. If the future of AI over the next few years is one of flattening growth curves and minimum-wage workers in developing countries labeling data, things are going to get very ugly very quickly.