I have previously noted I am not a fan of Ed Zitron’s writing on A.I., which I think is driven more often by adherence to narrative than by genuine skepticism. Even so, his newsletter is extremely popular, and occasionally that pays off with honest-to-goodness scoops. Yesterday, he got a big one — a smattering of OpenAI financial documents revealing the company’s spending and earnings for the past two years. In 2024, it made billions of dollars less than it spent — including over a billion dollars on sales and marketing alone — and its 2025 numbers look even worse:

The financial condition of OpenAI is deeply concerning. $38.53 billion in losses are astronomical, and far higher than most believed it would be. Losses also appear to be mounting year-over-year at a dramatic rate, and I’m not sure how this company finds a way toward any kind of sustainability or profitability.

Zitron shared these documents with the Financial Times, which independently verified them, and added some much-needed context. In particular, that whopping $38.5 billion loss accrued in 2025 and highlighted by Zitron — including in his headline — seems far less dramatic:

Before OpenAI’s switch late last year to become a public benefit corporation, investors in the company received convertible interest rights rather than conventional equity. Under US accounting rules, those interests were treated as liabilities and periodically revalued as the company’s valuation increased.

As OpenAI’s worth rose, the increased value of those investor rights created a roughly $30bn charge, added the person. The charge is not expected to recur following the restructuring, they said.

That expense is not something that can be waved away, of course, but it does not seem to be materially related to the company’s actual costs of creating and selling its products. Losses without including that charge were, according to the Times’ “person familiar with the matter”, $8 billion, or roughly 60% more than in 2024. But that is against revenue of $13 billion in 2025, a significant increase over 2024’s $3.7 billion. (OpenAI, in the first three months of 2026, earned $5.7 billion.)

These juicy numbers were republished by outlets like Reuters, the Next Web, Stocktwits, Benzinga, and Startup Fortune. Shamefully, all attributed them solely to the Times without mentioning Zitron’s critical role. These publications — particularly Reuters — should be giving full credit to the original source.

That Zitron now has actual, verified numbers also allows us to check some of his own reporting. For example, in April, he was quite upset that “every outlet has continued to repeat that OpenAI ‘made $13 billion in 2025,’ despite that being very unlikely given that it would have required it to have made $8 billion in a single quarter”. It is unclear to me which outlets Zitron is referring to as I could find just one — Russia Today — using that quoted phrase verbatim.

Even so, Zitron goes on to write about some apparently conflicting numbers reported by Anthropic before concluding:

Though I cannot say for certain, both of these situations suggest that Anthropic and OpenAI are misleading their investors, the media and the general public. If I were a reporter who had written about Anthropic or OpenAI’s revenues previously, I would be concerned that I had published something that wasn’t true, and even if I was certain that I was correct, I would have to consider the existence of information that ran counter to my own. I would be concerned that Anthropic or OpenAI had lied to me, or that they were lying to someone else, and work diligently to try and find out what happened. I would, at the very least, publish that there was conflicting information.

Two days after this article, he again claimed that “every single story about OpenAI’s revenue other than my own reporting (which came directly from Azure) massively overinflates its sales”, which are more like “a mere $2.27 billion in the first half of last year”.

The numbers Zitron now has for OpenAI suggests this narrative is complete hogwash. Yes, these companies leak overly-optimistic annualized run rates, but if that was a factor in the audited financials Zitron obtained, he likely would have mentioned that. He does not — and neither does the Times, for that matter. “Due to the seriousness of this story”, Zitron wrote, “I am not going to do very much editorializing”, so we will see in a later issue of this newsletter whether he acknowledges this self-induced frenzy was all in his head.

This is why I read Zitron’s work in the framework of conspiracy thinking. He accused OpenAI of “massively overinflat[ing] its sales” and “misleading their investors” based on his own calculations using leaked Azure figures. But it turns out OpenAI did, apparently, have that $13 billion in real non-ARR-fudged revenue for last year, and its operating loss is shrinking. Real analysts, not me, can figure out whether this company is on a path to a functional business. Zitron conjured a whole fictional narrative out of misreading some numbers and then, it would seem with this latest update, misunderstanding them again because it is useful for the story. Still, he should be credited for this scoop.

Jon Keegan, of Robinhood’s Sherwood News:

A Sherwood News analysis shows that the breaks afforded to Meta on just the sales tax of GPUs would come out to more than $3.3 billion — enough to build 33 new high schools, pay the salaries of all the state’s public school teachers for more than a year, or pay for more than seven years of the Louisiana State Police budget. (The secretary from the Parish committee that approved the financing plans declined to comment, and the chair of the committee didn’t respond to requests for comment.)

This is the very same project where Jonathan Weil, of the Wall Street Journal, found “aggressive accounting” that “strains credibility”. Neither of these advantages would be possible for a less-resourced competitor. Meta is a company so rich it benefits immensely without carrying nearly as much risk as the scale of this project would imply.

⌥ Permalink

Tabby Kinder in New York and George Hammond, Financial Times:

OpenAI has signed about $1tn in deals this year for computing power to run its artificial intelligence models, commitments that dwarf its revenue and raise questions about how it can fund them.

Emily Forgash and Agnee Ghosh, Bloomberg:

For much of the AI boom, there have been whispers about Nvidia’s frenzied dealmaking. The chipmaker bolstered the market by pumping money into dozens of AI startups, many of which rely on Nvidia’s graphics processing units to develop and run their models. OpenAI, to a lesser degree, also invested in startups, some of which built services on top of its AI models. But as tech firms have entered a more costly phase of AI development, the scale of the deals involving these two companies has grown substantially, making it harder to ignore.

The day after Nvidia and OpenAI announced their $100 billion investment agreement, OpenAI confirmed it had struck a separate $300 billion deal with Oracle to build out data centers in the US. Oracle, in turn, is spending billions on Nvidia chips for those facilities, sending money back to Nvidia, a company that is emerging as one of OpenAI’s most prominent backers.

I possess none of the skills most useful to understand what all of this means. I am not an economist; I did not have a secret life as an investment banker. As a layperson, however, it is not comforting to read from some People With Specialized Knowledge that this is similar to historically good circular investments, just at an unprecedented scale, while other People With Specialized Knowledge say this has been the force preventing the U.S. from entering a recession. These articles might be like one of those prescient papers from before the Great Recession. Not a great feeling.

⌥ Permalink