Leaked Audit Exposes OpenAI Cash Crisis: 13 Billion Revenue Against a 38.5 Billion Loss

The video opens with a blunt contrast: while OpenAI sells the world on AI as the ultimate money machine, its own books describe a company deep in a cash crunch, planning to keep losing money every year through 2028 before turning cash-flow positive around 2029 or 2030.

Targets keep moving in one direction: projected 2030 revenue rose from about 175 billion dollars to 280 billion by February 2026, while expected spending jumped from near 200 billion to 665 billion; then came the summer-2026 leak, when journalist Ed Zitron obtained audited statements whose authenticity the Financial Times confirmed, showing 13.07 billion dollars of 2025 revenue against 34 billion in costs and a 38.5 billion net loss attributed to the company. The cost split runs to 7.5 billion of cost of revenue, 19.18 billion of research and development, 5.73 billion of sales and marketing and 1.57 billion of administration, leaving a 20.92 billion operating loss, while a 41.55 billion fair-value adjustment tied to the for-profit conversion pushed the headline net loss to 60.35 billion before minority interests brought the attributable figure to 38.53 billion.

Management leans on ratios: 2.25 dollars spent per 1 dollar earned in 2024 fell to 1.60 in 2025, which reads as discipline; early-2026 shareholder documents cited by The Information complicate that story, with first-quarter revenue nearly tripling year over year to 5.7 billion alongside a 3.7 billion cash burn, a 9.3 billion operating loss and a 12.4 billion non-cash restructuring charge, adding up to a 21.3 billion quarterly loss. Gross margin improved from 33 to 39 percent and the company sits on more than 73 billion of cash and securities, yet the Wall Street Journal says OpenAI missed internal revenue and weekly-user goals for the first time under its finance chief, with weekly ChatGPT users peaking near 920 million in February and averaging about 905 million for the quarter.

Governance signals get their own segment: CFO Sarah Friar, once reporting straight to Sam Altman, is said to answer now to Fidji Simo, the chief executive for AGI deployment, amid talk that the company may not be listing-ready; guidance points to roughly 25 billion of cash burn in 2026 and about 57 billion in 2027, so the ratio improves while the dollar hole keeps doubling.

The most technical stretch concerns depreciation: a firm buying tens of thousands of Nvidia GPUs does not expense them on day one but spreads the cost over their useful life, and here estimates diverge — Microsoft uses six years, Meta five and a half, while Michael Burry, famous for calling the 2008 housing crash, argues the realistic span is two and a half to three years given a fresh Nvidia chip every 12 to 18 months, hiding some 176 billion of expense across the industry between 2026 and 2028. Nvidia and CoreWeave counter that older silicon keeps resale value and customer contracts run five years anyway, even as Microsoft's own chief admits slowing some data-center construction for fear of overbuilding for hardware that ages before the buildings open; with no ten-year operating record for chips run this hot at this scale, a one- or two-year assumption error swings hundreds of billions, and the Cisco parallel — an 80 percent share-price fall in two years when spending outran revenue — hangs over the segment.

Funding comes next, framed as two loops; the Nvidia one starts with the September 2025 pledge of up to 100 billion, which lifted Nvidia past 4.5 trillion in market value within weeks, then follows the cash from Nvidia to OpenAI to cloud contracts with Oracle and back to Nvidia as chip orders, with Oracle booking a reported 500 billion of future business while all three cite the same money as booming demand. By January 2026 the Wall Street Journal said the pledge had stalled, with Jensen Huang privately doubting the deal was ever locked, questioning financial discipline and the threat from Google and Anthropic, so the figure under discussion fell toward 30 billion; set against 1.4 trillion of combined compute commitments across Microsoft, Amazon, Oracle, Nvidia and AMD for a business that made 13 billion in 2025, the video labels the pattern vendor financing, the telecom trick from before the dot-com bust, and warns that record spending is a fact while record demand may be the story told with it.

The Microsoft loop runs deeper: backing OpenAI since 2019 with more than 13 billion before the mega-rounds, Microsoft resold its models through Azure for a revenue cut, and filings put OpenAI-linked sales at 24.1 billion for the year to June — roughly 70 percent of the AI revenue used to justify about 190 billion a year of infrastructure outlay. Cash leaves Microsoft, reaches OpenAI with each funding round, and walks back as a cloud invoice booked as AI success; the April 2026 decision to drop the Azure revenue cut and let OpenAI rent Amazon and Google capacity looks like independence, but only helps if the staggering sums needed can be found anywhere else.

The long-range pitch has pivoted too: the 100 billion-a-year 2030 goal once rested on 20-dollar monthly subscriptions and now rests on ads, climbing from 2.5 billion in 2026 to 11 billion in 2027 and 100 billion by 2030, which requires weekly users to grow from some 900 million today to 2.75 billion. That means repeating in a few years inside one app what Facebook, Instagram and WhatsApp needed a combined 17 years to build, and drawing in a third of humanity out of 8.2 billion people including children and the offline; with analysts sizing the whole US chatbot ad market under 6 billion by 2030, the company's slice alone would run more than 16 times that estimate, while Anthropic answers with an ad-free pledge aired during the Super Bowl.

Then comes the power bill: an H100 draws about 700 watts alone and roughly 1,300 watts once server overhead is counted, so a 100,000-GPU site pulls near 200 megawatts and burns a year's electricity for close to 165,000 homes, with the next architecture expected to draw more than three times as much per chip. The Stargate venture with SoftBank and Oracle plans up to 500 billion of data centers in four years, plus a 250 billion Azure commitment into 2032, or about 600 billion of compute spend through 2030 against 13 billion of 2025 revenue; the March 2026 round closed 122 billion at an 852 billion valuation, with Amazon anchoring 50 billion of which 35 billion arrives only on a public listing by the end of 2028 or the arrival of artificial general intelligence. A planned 2-gigawatt Abilene, Texas expansion able to power almost 1.5 million homes was scrapped in March 2026, new grid connections in places like Northern Virginia take four to seven years, some counted-on power may not have broken ground, and the Microsoft exclusivity runs only to 2032 with clauses that fall away if the board declares the big breakthrough.

The closing arc covers structure and precedent: the capped-profit setup limiting investor returns to 100 times capital was loosened years ago with a 20 percent annual uplift and formalized when the business became a public benefit corporation in October 2025, clearing the path for the confidential IPO filing of May-June 2026 aimed at late that year in the 852 billion to 1 trillion band. Klarna's round trip — replacing 700 agents with an OpenAI assistant, then rehiring people after satisfaction sagged — reads as the enterprise cautionary tale; the video ends where it began, listing five bets that must land together, from fundraising cadence to depreciation math, circular funding, a materializing ad market and permanently cheap power, where one slip reprices the next round and shakes confidence in all five.