QWE AI Academy By the end of this guide, you’ll be able to open any leaked or filed AI company financial document – including the OpenAI numbers Ed Zitron just published – and tell within five minutes whether the company is actually dying, just burning cash on purpose, or hiding something in the footnotes. The OpenAI leak is the worked example. The skill is portable.
The story dropped on June 16 and the news cycle has been screaming about a $38.5 billion loss. That number is technically correct and also misleading. We’ll get to why.
The takeaway, in one paragraph
OpenAI’s leaked financials show a real problem, but not the one the headlines say. The $38.5B net loss is mostly a one-time accounting ghost from the nonprofit-to-for-profit conversion. The number you actually want to stare at is the $20.92 billion operating loss on $13.07 billion of revenue – meaning the core business spent roughly $1.60 to make every $1 in 2025. The cash burn is real but smaller than the headline. The dependence on Microsoft is the part nobody is pricing in correctly.
What actually leaked (the brief version)
Tech journalist Ed Zitron published audited 2024 and 2025 financials on June 16, 2026. The Financial Times independently verified the documents. OpenAI declined to comment. The leak landed eight days after OpenAI confidentially filed its S-1 with the SEC on June 8, aiming at a valuation up to $1 trillion.
That timing matters for the tutorial below. Because the S-1 is confidential, you cannot pull the full document off EDGAR yet – confidential drafts only become public at least 15 days before the IPO roadshow, per SEC rules for emerging growth companies. So for now, the Zitron leak is the only window we have.
Method A: Read the headlines. Method B: Read the four lines.
Most tutorials and most news write-ups about this leak follow Method A. They quote the $38.5B figure, mention the $17.2B Microsoft payment, gesture at the IPO, and stop. Method A gives you something to tweet. Method B gives you something to bet on.
Here’s how they compare on the same OpenAI document:
| What you look at | Method A (headlines) | Method B (the four lines) |
|---|---|---|
| Net loss | $38.5B – “catastrophic” | $38.5B but $30B+ is non-cash one-time |
| Operating loss | Not mentioned | $20.92B – the actual core business hole |
| Cash burn | Not separated | ~$8B real cash out the door (per FT source) |
| Compute dependency | “Big Microsoft deal” | $17.2B of $34B costs flow to one vendor |
Method B is what we’re going to walk through. It works on any AI company filing – Anthropic, xAI, Perplexity, whoever leaks next.
The four lines that actually matter
Line 1: Revenue growth (top line)
OpenAI did $13.07B in revenue in 2025, up from $3.7B in 2024. That’s 253% growth. For an AI company at this stage, anything under 100% growth would be alarming. Tripling is healthy.
What to check: is the revenue growing faster than costs? For OpenAI in 2025, revenue grew ~3.5x while total costs grew ~2.7x ($12.48B → $34B). The ratio is improving. Per Fortune’s math, OpenAI spent $2.37 to make $1 in 2024 and $1.60 to make $1 in 2025. Still upside-down. Less upside-down than before.
Line 2: Operating loss (not net loss)
This is the single most important number in any AI company document. Net loss includes accounting noise. Operating loss is closer to “did the actual business of selling AI to customers make or lose money?”
OpenAI’s 2025 operating loss: $20.92B. That means the company spent about $21B more running its product than it brought in. This is the number to compare across years and across competitors. Not the splashy net loss.
Line 3: Cost of revenue vs R&D split
An AI company’s cost structure tells you whether it’s a research lab or a product business. From the leaked docs:
- Cost of revenue (serving the model – mostly inference): $7.5B in 2025, up from $2.65B
- R&D (training the next model, salaries, experiments): $19.18B, up from $7.81B
- Sales & marketing: $5.73B, up from $1.11B (5x in one year – that’s the part that should raise eyebrows)
R&D bigger than revenue is the AI lab tell. A normal SaaS company has cost of revenue dominant and R&D maybe 20-30% of revenue. OpenAI’s R&D alone is 147% of revenue. That’s not a software business profile – it’s a frontier research operation that happens to also sell a chatbot.
Line 4: Single-vendor concentration
This is the line every news article mentions but nobody analyzes correctly. OpenAI paid Microsoft $17.2B in 2025 – that’s 50.6% of total expenses going to one supplier who also owns roughly 27% of the company. The Register’s reporting on internal Microsoft documents shows Azure inference alone hit $8.7-9B in the first three quarters of 2025, much of it as in-kind Azure credits rather than cash.
Pro tip: When a company’s biggest expense line is also a major equity holder, the “cost” number is partly a bookkeeping fiction. Microsoft can keep extending Azure credits because it’s effectively paying itself. The day that arrangement changes – through the IPO restructuring, a contract expiry, or a new cloud deal – that $17.2B becomes real cash leaving the building.
The accounting trick you need to spot
Here’s where Method B saves you from looking dumb on the internet.
The $38.5B headline net loss includes a $41.55B one-time charge tied to the fair value of convertible interests and warrant liability. Plain English: when OpenAI converted from a nonprofit to a for-profit, all those early investor IOUs got revalued at the company’s much higher current worth. About $30B of that charge is just “our valuation went up so our owe-it-to-investors number went up.” No money left the building.
Strip that out and an anonymous source told the FT the true out-of-pocket cash burn was around $8B for the year. That’s the real burn rate. With ~$25B in cash reserves, that’s roughly three years of runway at current pace – assuming Microsoft keeps eating compute costs in-kind.
Both numbers are real. They measure different things. The $38.5B is GAAP net loss. The $8B is cash burn. Anyone quoting one without the other is selling you something.
How to find this stuff yourself next time
You probably won’t get a leaked PDF emailed to you. Here’s the public path:
- Go to SEC EDGAR full-text search
- Search the company name or ticker – every U.S. public filing since 1994 is there, free
- For an IPO, filter by form type S-1 (or S-1/A for amendments – always read the latest amendment, not the original)
- Jump straight to “Management’s Discussion and Analysis” and the financial statements. Skip the marketing fluff at the top
- For a private company like OpenAI right now, you’re stuck waiting until the confidential filing goes public (15 days before roadshow at minimum)
Foreign companies file Form F-1 instead. Already-public companies use S-3 for follow-on offerings. The shape of the financial statements is the same. The four lines above work on all of them.
Edge cases the news coverage missed
The mystery $3.74B line. In the 2024 statements, OpenAI marked $3.74B as “net loss attributable to noncontrolling members capital,” which moved that loss off the headline figure. Zitron himself wrote that it’s unclear what this means or how OpenAI reconciled removing those costs. Translation: there are probably consolidated entities – subsidiaries, joint ventures, special-purpose vehicles – that haven’t been publicly disclosed. When the S-1 goes effective, this line should get explained. Until then, treat the 2024 “net loss attributable to OpenAI” of $5.09B with mild suspicion.
The in-kind compute trap. The $8B real cash burn figure is the floor, not the ceiling. A chunk of Azure usage is being paid in Microsoft credits, not dollars from OpenAI’s bank account. If the IPO restructures the Microsoft relationship – which is one of the things the S-1 needs to address – those credits could convert into actual cash obligations. The cash burn number could double overnight without a single new dollar of spending.
The S&M jump. Sales and marketing went from $1.11B to $5.73B in one year. That’s a 5x increase against a 3.5x revenue increase. Either OpenAI is paying a lot more to acquire each customer (worse unit economics), or it’s front-loading enterprise sales spend that will pay off later. The leaked documents don’t separate enterprise from consumer, so this one is genuinely unknown until the S-1 breaks it out.
What to do with this in the next 24 hours
Open Zitron’s full breakdown in one tab and a blank doc in another. Write down the four numbers – revenue, operating loss, cost-of-revenue, single-vendor exposure – for OpenAI 2025. When the next AI company leak or filing drops (Anthropic, Mistral, xAI), pull the same four numbers and compare. That’s it. You’ll know more than 95% of the people posting about it.
FAQ
Is OpenAI actually going bankrupt?
No. With roughly $25B in cash and ~$8B real annual burn, there’s multi-year runway even before the IPO closes.
Why did the leak land right before the IPO?
Probably not a coincidence. OpenAI filed its S-1 confidentially on June 8, 2026, and the documents leaked to Ed Zitron eight days later. The confidential filing route is designed to let management control the narrative – leaks like this one strip that control away. Whoever passed the audited statements to Zitron arguably wanted the public to see the $20.92B operating loss before OpenAI’s bankers could frame it as a “path to profitability” story in the prospectus. The fact that the FT independently verified the documents tells you the source had real access, not screenshots from a forum.
Does losing money this fast mean AI is a bubble?
Not on its own. Amazon lost money for years building infrastructure that later became hugely profitable. The honest answer is that we won’t know until 2027 or so whether per-query inference costs fall faster than per-query revenue. If the cost curve bends down – through smaller models, better hardware, or efficiency gains – OpenAI’s numbers look like Amazon circa 2001. If it doesn’t, they look like WeWork. The leaked docs don’t settle the argument either way. They just confirm the bet is enormous and the timeline is short.