# OpenAI Commits to a 2027 IPO - and Puts a Real Number on What > Source: > Published: 2026-08-20 13:04:40+00:00 # OpenAI Commits to a 2027 IPO - and Puts a Real Number on What At an all-hands, CFO Sarah Friar told staff OpenAI will be public in 2027, settling an internal split with Sam Altman. Separately, OpenAI put a number on… OpenAI has finally put a year on the calendar for its long-rumored public offering, and in the same week it disclosed a number the industry usually answers with adjectives. At an all-hands, CFO Sarah Friar told employees the company will be a public company in 2027, or sooner if the business keeps inflecting. That settles a reported internal split: Friar pressing for a 2027 listing at around a trillion dollars, against Sam Altman pushing for a 2026 debut. The cautious path won. ## The S-1 That Hasn't Shown Up The confidential S-1 went to the SEC on June 8. The public prospectus still hasn't appeared on EDGAR, and it must be public at least 15 days before any roadshow begins. A 2027 commitment reads as a deliberate slowdown rather than a delay forced by regulators. The numbers the market is waiting to see audited haven't changed: roughly 2 billion dollars a month in revenue, reported losses of about 1.22 dollars for every dollar earned, an 852 billion dollar valuation from the March round, and the Microsoft revenue-share terms that have never been disclosed. Goldman Sachs and Morgan Stanley are leading the offering. None of this is investment advice - it's a company explaining, with more specificity than before, when it intends to let the public see its books. ## What Safety Actually Costs The more novel disclosure was buried in the same week. OpenAI published its own figure for what safety costs at the frontier: monitoring overhead now runs roughly 20 percent of the [inference](/glossary/inference) [compute](/glossary/compute) being monitored. That covers all [reinforcement learning](/glossary/reinforcement-learning) [training](/glossary/training) and tool-involving evaluations for GPT-5.6 Sol-class models and above, plus all inference on Astra. Twenty percent is a real number on a question the industry usually answers in adjectives. It means one dollar in every five spent on monitoring the most capable models goes not to the model being useful, but to the machinery watching it. That's the price of running a frontier lab under its own safety commitments rather than someone else's. ## The Musk Companies Are Building Something Else Bloomberg reports SpaceX approached Cognition, maker of the Devin coding agent, about a potential acquisition and is weighing a compute partnership. It comes days after closing a reported 60 billion dollar [Cursor](/compare/github-copilot-vs-cursor) deal. Read together with the [Grok](/compare/mistral-large-vs-grok-2) 4.7 delay - xAI is running supplemental training on SpaceX engineering data - a picture forms. The Musk companies are assembling an engineering-specific AI stack: the model, the editor, and now possibly the agent, all trained on proprietary aerospace work no other lab can touch. [Nvidia](/glossary/nvidia) and the Data Layer Separately, The Information reports Nvidia is in talks to invest in Mercor at a 20 billion dollar valuation, roughly double the 10 billion Series C from nine months ago. Mercor's first-half gross revenue reached 614 million dollars against a 2 billion annualized run rate, and Nvidia already paid Mercor tens of millions last quarter. The telling detail: the company being valued at 20 billion dollars doesn't build models. It labels data for the people who do. Nvidia, which sells the shovels, is betting the bottleneck has moved upstream to the raw material. *Sources: AI Tools Recap daily briefing, August 20, 2026; OpenAI internal all-hands reporting on the 2027 listing; Bloomberg and The Information reporting on SpaceX and Nvidia, August 2026.* Get AI news in your inbox Daily digest of what matters in AI.