OpenAI’s $38.5 Billion Hole: The Math That Makes Investors Sweat

The numbers are finally out, and they’re more brutal than anyone predicted. Leaked audited financial documents, independently verified by the Financial Times, show that OpenAI lost approximately $38.5 billion in 2025. That’s not a rounding error from a bad quarter. That’s a company spending $34 billion against $13.07 billion in revenue, with a loss from operations of $20.92 billion. The “revolutionary” business model mostly revolutionized the concept of burning cash.

The $38.5 Billion Question: Where Did All the Money Go?

Let’s start with the raw data, because the numbers tell a story that no amount of PR spin can bury.

The headline $38.5 billion loss includes a non-recurring accounting charge of approximately $30 billion related to OpenAI’s conversion to a for-profit structure. Strip that out, and you’re looking at a more “reasonable” $8 billion loss. But that framing misses the point entirely. The operational loss, the money actually flowing out the door every day, doubled from $8.78 billion to $20.92 billion.

The single largest line item? A whopping $10.59 billion paid to Microsoft for “Research and development” expenses, almost certainly for model training compute. Total payments to Microsoft across all categories hit $17.2 billion. And that’s the company that everyone expects to just “figure out” profitability.

Why ChatGPT Can’t Simply Raise Prices

The obvious fix for a company burning cash is to charge more. But OpenAI faces a fundamental constraint that most traditional businesses don’t: the competitive landscape has shifted dramatically beneath their feet.

While OpenAI was spending $19 billion on R&D, open-source models were getting better, faster, and cheaper. The emergence of models like DeepSeek V4 and Qwen 3.x has created a pricing environment where OpenAI can’t simply 10x their prices without bleeding users. One developer described running a 26B parameter model locally that performs “kind of nuts” for daily tasks. When open-source alternatives can run on consumer hardware, the willingness to pay premium API rates evaporates.

The response from the market has been clear. Enterprise customers are beginning to balk at token-based pricing and demanding measurable ROI. A Forbes survey found that 56% of CEOs see zero ROI from their AI investments. When your customers are questioning whether they’re getting value, and your costs are spiraling, you’re caught in a trap.

The Inference Cost Crisis Nobody Wants to Discuss

This is where things get technical and genuinely scary for the AI industry. Training models is expensive, sure. But inference, the cost of actually running these models for users, is the silent killer.

Training: One-time cost, spread over months, on massive clusters
Inference: Recurring cost, scales linearly with users, requires millions of GPUs

This is the crux of the inference cost crisis that threatens AI margins. OpenAI’s cost of revenue jumped from $2.65 billion to $7.5 billion in a single year, a 183% increase. That’s the cost of generating tokens for hundreds of millions of users. And unlike training, where you can pause and optimize, inference costs are tied directly to user engagement. The more people use ChatGPT, the more money OpenAI loses.

The Circular Cash Model: We’ve Seen This Before

One of the most revealing details from the leaked documents is the nature of OpenAI’s “investors.” In 2025, SoftBank paid OpenAI $867 million. Microsoft paid $303 million. But this isn’t straightforward investment, it’s often vendor financing structured as debt.

Consider this: OpenAI paid Microsoft $10.59 billion for R&D in 2025. Microsoft is also an investor. This creates a circular flow of capital that looks suspiciously like the kind of financial engineering that preceded the 2008 crisis. The WSJ reported that OpenAI has pledged (though not actually spent) $1 trillion on data center buildouts. Where’s that money coming from?

The answer is increasingly from NVIDIA itself through an equity-for-chips scheme inflating the AI bubble. More than 50 equity-for-chips deals were signed this year alone, with NVIDIA holding a $100 billion position in OpenAI. When your hardware supplier is also your largest creditor, the accounting gets creative.

The Open-Source Pressure Cooker

The timing of this leak is particularly brutal for OpenAI. They’re preparing for an IPO that could determine the company’s survival. But the market is already saturated with alternatives that cost a fraction of what OpenAI charges.

DeepSeek’s V4 models, for example, launched with 1.6 trillion parameters and 49 billion active in a Mixture-of-Experts architecture. Google’s Gemma 4 is pushing incredible performance at smaller sizes. The open-weight challenge from DeepSeek isn’t just competition, it’s a direct assault on the economic moat that OpenAI was supposed to have.

The result is a pricing war that benefits nobody except the end user. OpenAI’s projected ChatGPT Plus subscribers are expected to drop from 44 million in 2025 to just 9 million in 2026, driven by a shift to cheaper subscription tiers. Their own internal projections show an 80% decline in premium subscribers. That’s not growth, that’s a managed retreat.

Can OpenAI Survive Until 2030?

The company is telling investors it hopes to be profitable by 2030. But the math is brutal. HSBC estimated that OpenAI needs at least $207 billion by 2030 just to keep losing money at current rates. Their operating losses as a percentage of revenue improved from 237% to 160% year-over-year, but that’s still spending $1.60 for every dollar earned.

The company has just over $50 billion in assets, with about half in cash. That sounds like a lot until you realize they’re burning through it at roughly $20 billion per year in operational losses alone. The structural collapse of OpenAI’s finances isn’t a future concern, it’s happening now.

Consider the SpaceX wild card. When SpaceX IPOs, potentially at a $2 trillion valuation, it will suck up enormous liquidity from the market. That means less capital available for AI companies that can’t demonstrate a path to profitability. OpenAI is racing against an IPO clock that might already be ticking faster than they realize.

The Real Winner of This Crisis

The biggest beneficiaries of OpenAI’s financial distress are the developers who have been building with open-source alternatives. The economic collapse of cloud LLM APIs is accelerating as open-source models hit performance parity. When you can run a model locally that performs 90% as well as GPT-5 for zero recurring cost, the proprietary API model starts looking less like a business and more like a charity.

The question isn’t whether OpenAI will fail. It’s whether the entire “closed-source frontier model” business model was always a house of cards. The leaked financials suggest it was. And the open-source community has been quietly building the scaffolding that will catch the pieces when it falls.

What This Means for Practitioners

If you’re building on top of OpenAI’s APIs, this leak should concern you. A company losing $38.5 billion annually is a company that will eventually have to drastically change its pricing model, or cease to exist. The economic collapse of cloud APIs due to open-source pricing isn’t theoretical. It’s happening right now, and it’s going to reshape how developers think about AI infrastructure.

The smart money is on diversification. Build abstractions that let you swap providers. Invest in local inference capabilities. Watch the open-source ecosystem like a hawk. Because whatever happens to OpenAI, the future of AI isn’t going to be built on $38.5 billion annual losses.

Sam Altman might be shopping for IPO buyers, but the numbers suggest he’s selling something that doesn’t exist yet: a profitable AI company. And in a market where open-source models keep getting better for free, that’s an increasingly tough pitch to make.