The $19 Billion Hole: Inside OpenAI’s Leaked Financials and the Reckoning Before the IPO

On paper, OpenAI is the greatest growth story in enterprise software history. The company that taught the world to talk to machines reported revenue surging from $3.7 billion in 2024 to $13.07 billion in 2025 [1]—a 253% year-over-year increase that would make any SaaS founder weep with envy. But the documents independent journalist Ed Zitron obtained and published tell a far more harrowing story: OpenAI is hemorrhaging cash so fast that its top-line growth looks like a desperate sprint to outrun a financial avalanche.

The audited financial statements, filed as part of OpenAI’s SEC paperwork ahead of an expected IPO, reveal a company whose expenses have grown so aggressively that its losses have become a structural feature, not a temporary bug [1]. The numbers are brutal. While revenue hit $13.07 billion, the company’s net loss ballooned to $7.81 billion in 2025, up from $2 billion in 2024 [1]. Total expenses reached $19.18 billion [1]. For every dollar OpenAI earned, it spent nearly $1.47. This is not a startup burning cash to capture market share; this is a machine consuming its own fuel faster than it can refine it.

The Infrastructure Tax: Why Inference Costs Are Eating OpenAI Alive

To understand the $19.18 billion in total expenses, look past the marketing gloss of “democratizing AI” and stare directly at the physics of inference. Every time a user prompts GPT-4o, a developer calls the API, or a ChatGPT subscriber generates an image or analyzes a PDF, OpenAI pays a compute tax to NVIDIA. Infrastructure costs are not just a line item; they dominate the company’s entire financial architecture.

The leaked documents do not break out compute costs specifically, but the industry context is damning. NVIDIA’s most recent 10-Q filing, dated May 20, 2026, shows the chipmaker continuing to command premium pricing for its H100 and B200 GPUs [5]. Meanwhile, Daily Neural Digest’s tracking of real-time GPU pricing across Vast.ai, RunPod, and Lambda Labs shows that renting a single H100 node still costs between $2.50 and $4.00 per hour, depending on the provider and contract length. OpenAI, running at hyperscale, likely negotiates volume discounts, but the sheer compute required to serve hundreds of millions of monthly active users means those discounts still run into the billions.

The fundamental problem: large language model inference is computationally asymmetric. Training a model is a fixed cost—massive, but finite. Inference is a variable cost that scales linearly with usage. And OpenAI’s usage is exploding. The company’s API, which provides access to GPT-3, GPT-4, and Codex models, remains one of the most widely used developer platforms in the world [1]. Every API call, chatbot interaction, and Sora video generation burns GPU cycles that must be paid for in real time.

This is the dirty secret the leaked documents now expose: the unit economics of inference have not improved fast enough to outpace demand growth. OpenAI is effectively running a utility where the cost of delivering a single unit of intelligence still exceeds what the market will pay. The company’s $150 million investment in the OpenAI Partner Network, announced on June 14, 2026, attempts to push more compute burden onto enterprise customers by encouraging them to deploy custom models in their own environments [2]. But that program remains in its infancy and does nothing to alleviate the cost of serving ChatGPT’s consumer base.

The Regulatory Crosshairs: Investigations, Health Data, and the IPO Clock

As if the financial bleeding were not enough, OpenAI now navigates a multi-front legal war that could further complicate its path to public markets. On June 13, 2026, TechCrunch reported that OpenAI faces an investigation from state attorneys general, with inquiries spanning everything from the company’s ad policies to its handling of health data [3]. The exact states involved remain undisclosed, but the scope is unusually broad, suggesting regulators are taking a holistic look at OpenAI’s data practices rather than focusing on a single violation.

This is a nightmare scenario for a company preparing to go public. IPO prospectuses require companies to disclose material legal risks, and an active multi-state investigation is about as material as it gets. The health data angle is particularly concerning. OpenAI’s models increasingly appear in medical contexts—from clinical note generation to patient triage chatbots—and the company’s privacy policies have historically been vague about how user data flows through its training pipelines. If state AGs find that OpenAI mishandled protected health information, the fines could be substantial, and the reputational damage could spook institutional investors already skittish about AI risk.

The timing could not be worse. The leaked financials show a company that needs a public offering to raise capital and provide liquidity for early investors and employees. But the combination of a $7.81 billion annual loss and an active regulatory investigation creates a toxic cocktail for underwriters. Investment banks will demand steep discounts on the IPO price to compensate for the risk, and some may walk away entirely. The documents Zitron obtained were supposed to support a carefully managed narrative of growth and discipline. Instead, they have become the centerpiece of a narrative about a company growing so fast it is breaking itself.

The Open-Source Shadow: Why Competitors Are Eating OpenAI’s Lunch

While OpenAI burns billions on proprietary infrastructure, the open-source ecosystem quietly builds alternatives that threaten to undermine the company’s entire business model. The numbers from HuggingFace tell a stark story. OpenAI’s own open-weight models—gpt-oss-20b and gpt-oss-120b—have been downloaded 5,193,662 and 3,155,936 times respectively [1]. The whisper-large-v3 speech recognition model has been downloaded 4,629,156 times [1]. These are not trivial numbers. They indicate a massive, hungry developer community that wants to run OpenAI-quality models on their own hardware without paying per-token API fees.

But the real threat comes from outside OpenAI’s own ecosystem. NVIDIA’s NeMo framework—a scalable generative AI framework for researchers and developers working on large language models, multimodal AI, and speech AI—has accumulated 16,885 stars and 3,357 forks on GitHub [1]. Written in Python, NeMo represents the kind of infrastructure that allows companies to build their own AI stacks without ever touching OpenAI’s API. And NVIDIA is not stopping at frameworks. On June 16, 2026, the company announced the public beta of NVIDIA XR AI, a framework for building multimodal AI agents for AR glasses and XR devices [4]. This directly targets the next frontier of AI interaction—ambient, always-on agents that run on edge devices rather than in the cloud.

The strategic implication is clear: NVIDIA is building the entire stack—hardware, frameworks, and deployment tools—to enable a world where AI inference happens locally, not in OpenAI’s data centers. If AR glasses become the primary interface for AI agents, as NVIDIA is betting, then the cloud-based API model OpenAI depends on becomes a legacy architecture. Every GPU NVIDIA sells to a company running NeMo or XR AI is a GPU not sending inference requests to OpenAI’s servers.

The Partner Network Gambit: Can $150 Million Buy a Lifeline?

On June 14, 2026, just two days before the financial leak, OpenAI announced the OpenAI Partner Network, committing $150 million to help global partners accelerate enterprise AI adoption, deployment, and transformation [2]. On the surface, this looks like a standard channel partner program—the kind Salesforce or Microsoft have run for decades. But in the context of the leaked financials, it reads as a desperate attempt to shift the company’s cost structure.

The logic is straightforward. Enterprise customers who work with certified partners are more likely to deploy custom models, use fine-tuning, and build internal AI applications that run on dedicated infrastructure. This reduces the load on OpenAI’s shared inference clusters and improves margin per customer. The $150 million investment essentially subsidizes training an army of consultants and system integrators who will help enterprises build their own AI stacks on top of OpenAI’s platform, rather than relying on the raw API.

But $150 million is a rounding error compared to the $19.18 billion in total expenses. It is also a bet that the partner ecosystem can grow fast enough to offset consumer-side losses. That is a risky wager. Enterprise sales cycles are long, and the deep integration work partners do takes months or years to bear fruit. Meanwhile, the consumer business continues to burn cash at a rate that makes the partner program look like a bucket of water thrown at a forest fire.

The Verdict: A Company Running Out of Runway

The leaked financial documents do not just show a company losing money. They show a company that has built its entire strategy around the assumption that revenue growth will eventually outpace cost growth. That assumption is now in serious doubt. With $7.81 billion in annual losses and total expenses of $19.18 billion, OpenAI is burning through its cash reserves at a pace that demands either a massive infusion of public market capital or a dramatic restructuring of its cost base [1].

The IPO, once seen as the crowning achievement of the AI revolution, now looks like a Hail Mary pass. The regulatory investigation from state attorneys general adds another layer of uncertainty, and the open-source ecosystem is eroding the moat that OpenAI’s proprietary models once provided [3]. The company’s own open-weight models have been downloaded over 8 million times, suggesting that even OpenAI’s own technology is being used to build alternatives to its paid API [1].

What the mainstream media is missing in this story is the structural fragility of the cloud-based AI business model. OpenAI is not just a company with a cash flow problem; it has bet its entire future on the assumption that inference costs will continue to fall faster than usage grows. That assumption has held true for the past two years, but the laws of physics and economics are not infinitely bendable. Every new user, feature, and model adds to the compute burden. And as NVIDIA continues pushing the frontier with frameworks like NeMo and XR AI, the competitive landscape shifts toward edge computing and local inference—a world where OpenAI’s centralized API model may become an expensive anachronism [4].

The documents are leaked. The numbers are public. The question now is whether OpenAI can rewrite its own code before the runtime crashes.

---

References

[1] Editorial_board — Original article — https://arstechnica.com/ai/2026/06/leaked-financial-docs-show-openai-is-losing-billions-of-dollars-a-year/

[2] OpenAI Blog — Introducing the OpenAI Partner Network — https://openai.com/index/introducing-openai-partner-network

[3] TechCrunch — OpenAI faces investigation from state attorneys general — https://techcrunch.com/2026/06/13/openai-faces-investigation-from-state-attorneys-general/

[4] NVIDIA Blog — Hands Free, AIs Forward: NVIDIA XR AI Brings Agents to AR Glasses — https://blogs.nvidia.com/blog/nvidia-xr-ai/

[5] SEC EDGAR — NVIDIA — last_filing — https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0001045810