OpenAI's prospectus is a secret. The company submitted its S-1 to the SEC confidentially—the filing confirmed publicly on June 8, with Goldman Sachs, Morgan Stanley and JPMorgan leading and a listing targeted as early as September at a valuation expected to exceed $1 trillion [1]. The audited financials inside it will not be public until at least 15 days before the roadshow [1]. Yet the parts of OpenAI you can actually buy keep going public ahead of the investors being asked to price it.

That gap is the story the paper has tracked into a third edition. On June 15 it argued that buyers can use OpenAI's credits, workspaces and rails before investors inspect the books, and that private-allocation claims keep outrunning audited disclosure. June 16 adds two more transactable surfaces while the filing stays sealed.

The first is live. On June 10, Visa and OpenAI shipped agentic payment rails that extend to Codex, OpenAI's coding agent, letting it transact within user-set limits using tokenized credentials and fraud monitoring [3]. The second is sited: OpenAI is in advanced talks to anchor a $500 billion, 10-gigawatt campus in Ohio. A counterparty can plug into the rails, draw down cloud credits, or contract for capacity—commercial reliance on OpenAI is hardening into infrastructure—while the document that would let anyone test the economics underneath stays confidential.

What fills that vacuum is estimates, and the estimates are not flattering. Third-party models put OpenAI's 2026 GAAP loss around $25 to $26 billion, against a non-GAAP figure near $14 billion, with cash burn of roughly $25 billion [1]. One analysis of the company's quarterly numbers concluded it lost about $1.22 for every dollar it earned [1]. The cause is structural: gross margins sat near 33% in 2025, down from about 40%, because inference costs—the GPU bill for every query—are projected to climb to $14.1 billion in 2026 [1]. None of these figures comes from an audited filing. That is precisely the point. They circulate as screenshots because no inspectable document exists to confirm or refute them.

This is where the mainstream and X read the same week differently. IPO trackers frame it as a race: investing.com and CNBC line OpenAI up behind SpaceX, two companies that could bring combined new equity supply near $135 billion to public markets, with OpenAI's reported target implying more than 75 times estimated 2025 revenue [2]. The race framing is about who lists, when, and at what multiple. X reads the same facts as a disclosure problem. The AI commentator Rohan Paul noted that a confidential S-1 lets OpenAI begin SEC review "without immediately exposing revenue, losses, customer mix, compute costs, or executive pay." Others pass around the loss math directly, flagging a projected $14 billion 2026 loss and no profitability before 2029.

Both frames understate what the other sees. The race coverage treats the valuation as the live question while the audited basis for it remains unavailable. The X loss-ratio screenshots treat unaudited estimates as settled when the sealed books could move them in either direction. The paper's middle is narrower and sturdier: commercial reliance is outrunning audited disclosure. You can transact with OpenAI today; you cannot yet read what transacting with it costs the company.

The competitive backdrop only raises the stakes on the unseen numbers. Anthropic recently surpassed OpenAI in the private market—about $965 billion to OpenAI's $852 billion March round—and OpenAI's share of the developer market has slipped from roughly 60% to 51% year over year, with Claude Code taking ground in exactly the coding niche the rails are meant to monetize [1]. Anthropic is itself reportedly in early talks with the same banks about a listing [2]. The roadshow will eventually force the audited disclosure the rails have been outrunning. Until it does, OpenAI is selling the product in public and keeping the proof in private, and the buyers on both sides of the trade are pricing a company whose books they have not been allowed to open.

-- THEO KAPLAN, San Francisco