The math on OpenAI's public offering got sharper this week, not softer, and it turns entirely on which loss you print. The paper's July 7 account of OpenAI's trillion-dollar target against $14 billion in losses filed the figure and flagged a possible 2027 slip. Today the slip has firmed, and the numbers have split into two vintages that must never be run together.

Here is the discipline the whole story depends on. OpenAI's internal projections, reported by The Information, imply 2026 losses tripling to roughly $14 billion even as revenue runs toward a $20-to-25-billion annualized rate. [1] That is a projection for the current year. Separately, audited 2025 financial statements — leaked in June and independently verified by the Financial Times — show a $20.92 billion operating loss on $13.07 billion in revenue. [2] That is a larger, historical, audited number. The $14 billion is what OpenAI expects to lose this year. The roughly $20.9 billion is what it did lose last year. They are two different measurements of the same red ink, and conflating them is the single easiest way to get this story wrong.

Even the 2025 figure has layers. The $20.92 billion is the operating loss. The net loss for 2025 widened to $38.5 billion after a $41.55 billion charge tied to the company's conversion from nonprofit to for-profit — a change in the fair value of convertible interests and warrant liabilities, not cash burned on compute. [2] Total costs and expenses reached $34 billion, of which $19.18 billion was research and development. [2] The point of separating these is not pedantry. It is that a reader told "OpenAI lost $14 billion" and a reader told "OpenAI lost $38.5 billion" are looking at the same company through different windows, and only the sealed S-1 reconciles them.

On top of the accounting sits a number OpenAI chose. Sam Altman has called any valuation below $1 trillion a "non-starter." [3] When advisers offered a stark choice — list in late 2026 at a lower price, or wait until 2027 and chase the trillion-dollar mark — Altman rejected the discount. [3] Chief financial officer Sarah Friar has argued for the 2027 timeline, citing roughly $600 billion in future infrastructure spending commitments and the difficulty of meeting public-company reporting standards on a compressed schedule. [1] The trillion-dollar figure, in other words, is a floor Altman set, not a price the market has cleared. The most recent private mark was $852 billion; the trillion-dollar ask sits about 17 percent above it. [3]

The retreat from the tape is the news. OpenAI is not publishing an S-1 to prove the numbers; it is deferring one to protect a valuation. The confidential filing — submitted on approximately June 8 — is the only instrument that would produce audited revenue composition, customer concentration, compute obligations to partners like Oracle and SpaceX, data-center energy contracts, and GPU delivery schedules. [3] It is precisely what stays sealed. The company that would most benefit from showing its books has decided that waiting is cheaper than showing.

That is where the paper's standing rule bites: every IPO claim carries a power, water, and cost receipt. OpenAI's burn is driven primarily by compute and data-center costs, and its projection of cash-flow positivity by 2030 assumes an inference-efficiency curve no auditor has signed. The trillion-dollar number is a bet that the buildout keeps compounding; the losses are the buildout's bill arriving early.

On X, the story is bubble mania — a cash furnace demanding a price it cannot justify, with a mega-listing threatening to whipsaw the chip names that depend on OpenAI's checks. The valuation math travels: The Information's own reporting notes Anthropic's $30-billion run rate may have overtaken OpenAI, making the latter's $850 billion tag "difficult to justify." [4] MSM writes the trillion-dollar headline and the loss figure — sometimes the projection, sometimes the audited number, occasionally both without labeling which is which.

The paper's middle is the vintage. The instrument that would settle the argument is a public S-1, and it is exactly what OpenAI is choosing not to file. Until it does, the trillion-dollar valuation is a floor with no clearing price, the $14 billion is a forecast, the $20.9 billion is history, and the reader who cannot tell those three apart is being sold the bull case as arithmetic. Whether OpenAI files publicly in 2026, formally confirms 2027, or lets the roughly $6.8-billion-a-month burn force a raise first, the answer will arrive as a document — or it will not arrive at all.

-- THEO KAPLAN, San Francisco