Leaked Audited Books Show OpenAI Lost $38.5 Billion in 2025

01OpenAI's Audited Books Leaked, Showing a $38.5 Billion Loss for 2025

Ed Zitron saw the documents first. The audited financial statements he describes put OpenAI's 2025 net loss attributable to the company at $38.53 billion. The Financial Times independently verified the figures. On Tuesday, an Ars Technica story pushed them into wider circulation.

The papers themselves stay private. The numbers are specific enough to cut through a year of fuzzy AI financing talk. OpenAI generated $13.07 billion in revenue in 2025, according to Zitron's account of the statements. It also booked $34 billion in total costs and expenses. That left a $20.92 billion loss from operations.

The top-line loss runs larger still. A net loss of $60.35 billion shrank to $38.53 billion once losses attributed to noncontrolling interests were stripped out. Revenue did grow. Research, compute, and other expenses grew faster, then buried it.

For Sam Altman, the timing stings. The former Y Combinator president took over as OpenAI's chief executive in 2019. He has spent the first half of 2026 selling investors one argument: that compute, distribution, capital, and consumer habit compound faster than the cost base. The leaked statements show the cost base ahead.

Most prior loss numbers came from estimates and unnamed sources. These carry an auditor's sign-off and a second newsroom's verification, which makes them far harder to wave away as speculation. The reported revenue and cost lines reconcile to a single operating result.

The leak also lands days after a regulatory step. On June 8, OpenAI said it had confidentially submitted a draft S-1 to the SEC. The company added that it expected the filing to leak and had not decided when to go public. Audited losses of this scale now sit in front of public-market investors before any roadshow opens.

The gap between income and burn is the entire story. Roughly $13 billion came in. More than $34 billion went out. No quarter of revenue growth disclosed so far closes a hole that wide.

02An AI Chemist Improved a Hard Drug-Synthesis Reaction, Running With Little Human Help

OpenAI and Molecule.one say a near-autonomous AI chemist, built on GPT-5.4, improved a difficult reaction used in medicinal chemistry. The companies describe a system that proposes, runs, and refines steps in a drug-synthesis workflow with limited human direction. The claim is about a single reaction, not a pipeline. But it lands the model inside the lab, not the chat window.

Google reported a parallel result on the clinical side. In a study published in Nature, its conversational system AMIE matched primary care physicians on managing complex, long-term health conditions. Chronic disease management is the slow, unglamorous core of primary care: titrating medication, tracking labs, adjusting plans across months. Google says AMIE held its own against doctors on exactly that work.

Two systems, two different labs, the same shift: the output is now expert-level task work, not summaries or first drafts.

That shift forces a measurement problem, and OpenAI moved to fill it the same week. The company introduced LifeSciBench, a benchmark it describes as expert-authored and expert-reviewed, built to score how AI handles real-world life-science research tasks and decisions. The framing matters. General reasoning tests do not capture whether a model can plan a synthesis route or weigh a treatment adjustment. A benchmark written and graded by working scientists raises the bar for what counts as a passing answer.

The three pieces fit together. Production capability and the tools to grade it are maturing in step, which is what separates a demo from a result a lab can act on. For drug discovery, an AI that improves a known-hard reaction shortens one step in a long chain; chemists still own the route. For primary care, a system that equals doctors on chronic management points at triage and follow-up workflows, not diagnosis-by-bot. None of the work claims replacement. All of it claims parity or assistance on tasks that used to require a specialist.

What determines whether this holds is the grading. If LifeSciBench-style evaluations get adopted by people outside the labs that built the models, the parity claims become checkable. If they stay in-house, they stay marketing.

03The IPO-Summer Boom Is Running on 16 Percent Public Approval

Capital is pouring into artificial intelligence faster than the public will vouch for it. A new Pew Research study, released as TechCrunch calls it "a hot IPO summer" for the sector, finds that only 16 percent of Americans expect AI's impact on society over the next 20 years to be positive. Roughly 40 percent expect the opposite.

The gap widens the closer you look. Two-thirds of Americans think AI is advancing too quickly, according to Pew. The same study reports that 67 percent doubt the U.S. government will do anything to meaningfully regulate it, and 59 percent say they don't trust companies to build it safely. The people racing to ship AI products are doing so without the consent of the market they're shipping into.

That distrust runs deepest among the youngest users. Pew says just 14 percent of Americans under 30 believe the technology will help society, the most negative reading of any age group. These are the same cohorts the industry counts on for early adoption and lifetime monetization.

Skepticism has not slowed usage, which complicates the picture. The Verge, citing the same poll, reports that 49 percent of Americans now use chatbots at least occasionally, up from 33 percent in 2024. ChatGPT usage has doubled since 2023, and 44 percent of U.S. adults now reach for it. Pew finds about a quarter of Americans use AI chatbots daily, mostly for research or work.

So Americans are adopting a technology they expect to harm them. They use the tools, distrust the companies making them, and don't believe regulators will step in. Pew's numbers describe a public that has decided AI is both useful and bad for the country at once.

The financial bet assumes that adoption curves eventually translate into goodwill, or at least durable revenue. The polling says adoption and approval have decoupled, at the exact moment valuations depend on the story holding together.