OpenAI’s AI-adjusted earnings numbers have echoes of Groupon and WeWork

OpenAI CEO Sam Altman

When Google was about five years old, the world saw its finances for the first time when it filed for an IPO. The company was wildly profitable, making over $100 million in net income in 2003. It never looked back.

When Facebook was roughly seven years old, it shared financial data with prospective investors. I got a look at the document when I was a reporter at MarketWatch. It indicated that the company had already become remarkably profitable, too.

Now, we’ve gotten the first peak at OpenAI’s financials. It’s almost 10 years old, and it’s nothing like Google and Facebook.

The Information obtained OpenAI financial documents and published an analysis this week. Revenue is growing fast and could hit $100 billion in 2029. But the company projects total losses from 2023 to 2028 to be $44 billion.

Excluding AI model training costs

The most concerning part: OpenAI is asking investors to exclude the billions of dollars it spends on training AI models, according to the documents analyzed by The Information.

AI model training is one of the main things that OpenAI does. It’s what everyone thinks of when the startup is mentioned. It launched GPT-4 in early 2023 and rolled out a new o1 model in September, for instance.

Is it OK for an AI company to exclude the cost of training its AI models when it reports earnings? I checked this with a veteran venture capitalist who was a top executive at a Big Tech company for years.

The person said clearly no. These costs are real and it’s a sign of hubris that OpenAI is asking investors to ignore what is a fundamental business expense, the person explained. They asked not to be identified discussing sensitive matters.

‘Not kosher’

I also asked an accounting expert to go on the record about this. I chatted with Francine McKenna, who writes The Dig, a blog focusing on accounting, audit, and corporate-governance issues at public and pre-IPO companies. She used to work at KPMG Consulting and PwC, and implemented accounting and financial systems, including SAP and Oracle ERP software, for big companies.

She knows accounting and technology well. She read The Information story closely on Thursday and had this to say about OpenAI’s efforts to get investors to ignore the cost of training its AI models.

“That is a bridge too far,” McKenna told me. “It’s not kosher.”

For an AI company like OpenAI, training models will be an ongoing process. The world is constantly evolving, and new data is being generated, which will have to be incorporated into AI models’ understanding.

“They are going to have to constantly train their AI models on new information,” McKenna said. “They will never be done. I think it’s a sleight of hand to try to exclude these costs.”

IPO rules

This will become especially important if OpenAI tries an IPO, which some investors expect the startup to do in the coming years.

The Securities and Exchange Commission focuses heavily on alternative measures of profitability and does not allow companies to exclude many of the ongoing costs of doing business.

“If they put that in an IPO prospectus and the SEC looked at it, the SEC would probably say you can’t use these as an adjustment to income,” McKenna said. “To me, that’s more akin to ongoing product maintenance and development.”

I asked OpenAI about the concerns of McKenna and the VC partner on Thursday morning and at the end of that day. The company didn’t respond to my detailed questions.

The VC partner noted that other OpenAI costs could decline in coming years, citing inference as an example. Inference is the stage after training that allows AI models to respond to new information. It’s basically how the models are run.

Those post-training costs could decline dramatically in coming years, which could help OpenAI reduce its losses, the VC partner explained.

However, the VC said you still can’t exclude AI model training, which will be a large and ongoing fundamental cost of providing OpenAI’s services.

Groupon and WeWork

Excluding the cost of the core thing you do has echoes of other high-flying tech startups from a few years ago. McKenna drew comparisons to Groupon, and discussed WeWork, too.

Before Groupon went public in 2011, it suggested that investors look at “adjusted consolidated segment operating income,” or adjusted CSOI, which excluded several marketing-related expenses. The SEC objected, and Groupon had to change its accounting.

Last year, Groupon issued a “going concern” warning, raising doubts about its ability to stay in business. (Its performance has improved a bit this year).

In 2019, WeWork told investors to focus on “community-adjusted EBITDA.” Essentially, the company was asking investors to ignore the cost of doing many of the things that it needed to do to run its business.

OpenAI is doing a similar thing now, according to McKenna and the VC partner. Let’s call it “AI adjusted earnings” for an AI company.

“OpenAI can do whatever it wants and try to spin any story they want with investors. They are a private company,” McKenna said. “Are they producing audited financial statements and giving them to investors? That’s unclear. Right now, they don’t have to. But that will stop at a certain point if they eventually prepare a prospectus for an IPO.”

WeWork went bankrupt in November 2023.