The OpenAI IPO Is a Bet on One Number You Cannot See Yet

On May 22, 2026, OpenAI quietly handed the Securities and Exchange Commission a draft registration statement on Form S-1, the document a private company files to sell shares to the public. Goldman Sachs and Morgan Stanley are leading it. Nobody outside the company and the SEC has read a page of it.^[2] That is the point of a confidential filing, and it is the first thing worth understanding about this deal.

The reported numbers around it are enormous. OpenAI raised in February 2026 at roughly $730 billion pre-money, closed a $122 billion round at an $852 billion post-money valuation on March 31, and its bankers have floated a debut anywhere from that level to north of $1 trillion.^[2] Then in late June the story shifted again. Bloomberg, citing the New York Times, reported OpenAI is leaning toward waiting until 2027 rather than accepting a lower price now, because Sam Altman apparently treats any haircut to the trillion-dollar target as a nonstarter.^[4] All of these are reported figures, not confirmed by a public filing, and you should hold every one of them loosely.

Here is the thing that actually matters, and it is not the headline. At anything near these levels, you are being asked to pay somewhere around 30 times revenue for a company that lost more money last year than most of the S&P 500 makes. The bull case is real. The bear case is equally real. Whether you buy this depends entirely on which one you think is priced in, and right now the price says the bulls win by a landslide.

What a confidential S-1 actually is

Since the JOBS Act of 2012, a company can submit its IPO paperwork to the SEC privately, go back and forth on staff comments, and revise the whole thing without any of it hitting the public record. The obligation kicks in later: the registration statement must be filed publicly at least 15 days before the roadshow (the pitch tour where bankers walk institutional investors through the deal) begins. At that moment the draft, the SEC's comment letters, and the company's responses all land on EDGAR at once.^[3]

This is why the “is it $730 billion or $852 billion or $1 trillion” argument is mostly noise. None of those are a price. They are anchors. The real price gets set in a roadshow that has not happened, against a book of orders that does not exist, using financials the public has not seen. If you want to know what to watch, watch for the public S-1. Everything before that is positioning.

The bull case, made properly

I want to steelman this, because the lazy bear take on OpenAI is boring and mostly wrong on the thing that matters.

Distribution is the moat, not the model. Altman said in October 2025 that ChatGPT had crossed 800 million weekly active users.^[11]Reporting through 2026 puts it above 900 million. That is not an API business. That is a consumer utility with a habit loop, and consumer habits are the single most durable asset in technology. Google has been the default search engine for twenty years not because its index is unbeatable but because typing into the box is muscle memory. ChatGPT is the only AI product that has achieved that. “ChatGPT it” is a verb now. Claude is not a verb. Gemini is not a verb.

The brand is the category. When a technology becomes a commodity, the brand that owns the default position captures a wildly disproportionate share of the economics. That is the Coca-Cola argument applied to inference. If the models converge (and they are converging), the winner is not the lab with the best benchmark score. It is the one with the icon on nine hundred million home screens.

Agentic products are a real second act. The revenue mix today is heavily consumer subscriptions. The upside case is that agents that actually do work, book things, write code, run workflows, convert a $20-a-month chat subscription into a seat that a business pays hundreds for. If that lands, the revenue multiple you are paying today looks cheap in hindsight, the same way people said Amazon was insane at 100x earnings in 2013.

Growth is genuinely violent. Revenue went from $13.07 billion in 2025 to a reported $2 billion per month by early 2026, and the company has claimed it is growing several times faster than Alphabet or Meta did at comparable stages.^[2,5] That is not a rounding error. That is one of the fastest revenue ramps ever recorded.

Now the part the bankers would rather you skim

Three things bother me, and they compound.

One: the losses are not a rounding error, they are the business model. Leaked 2025 financials reported by Ed Zitron and later corroborated in the press show $13.07 billion in revenue against $34 billion in total costs, producing a $20.92 billion operating loss.^[5,6]R&D alone was $19.18 billion. Cost of revenue was $7.5 billion. OpenAI has reportedly told investors it plans to keep posting large annual losses through 2028 before turning profitable later.^[12] You are not buying a profitable company with a temporary investment cycle. You are buying a company that has told you, in writing, that it will burn cash for years.

Two: the compute obligations are a liability that looks like a strategy. The reported ~$1.4 trillion in multi-year commitments across Oracle, Microsoft, Broadcom, Nvidia, AMD, AWS and CoreWeave is roughly a hundred times the 2025 revenue that was on the books when they were signed.^[9]Set aside whether the deals are structurally binding. The signal is what it is. To hit the revenue that justifies the valuation, OpenAI has to spend an amount of money that only makes sense if the revenue arrives. That is a circular bet, and circular bets are exactly what get repriced when the market's mood changes.

Three: token prices are collapsing and OpenAI does not control that. API pricing across the industry fell roughly 80% between early 2025 and early 2026, and cheap frontier-adjacent models now go for pennies per million tokens.^[16] This is great for anyone building on top of these models. It is brutal for anyone whose valuation assumes model access is a scarce good. Inference is going the way of cloud storage: commoditized at the infrastructure layer, with the value migrating up to the application and the workflow. OpenAI is trying to be both layers at once, which is a defensible strategy and an expensive one.

Anthropic is the uncomfortable comparison

Here is the part that should give any prospective buyer pause. By multiple reports, Anthropic passed OpenAI in annualized revenue run-rate during 2026, hitting roughly $30 billion by late Q1 and, by some accounts, considerably higher by mid-year.^[10]Roughly 85% of that is enterprise and developer revenue, driven heavily by coding and agentic workloads. OpenAI's mix runs the other direction, weighted toward consumer subscriptions where the overwhelming majority of users pay nothing.

Those are two completely different businesses wearing the same costume. Enterprise revenue is stickier, higher gross margin, and easier to forecast. Consumer subscription revenue is bigger at the top of the funnel and far more exposed to churn and to a free tier that eats your own margin. Anthropic filed its own confidential S-1 on June 1, 2026.^[13] Both companies are going to be sitting in front of the same institutional investors in the same quarter, and one of them has the better income statement.

I do not think that settles the argument. Consumer scale is worth a premium, and OpenAI's is unmatched. But the multiple OpenAI is asking for implies it is the clear category winner, and on the metric that public-market investors care about most, revenue quality, it is currently not.

The structure is weird and nobody is talking about it

OpenAI spent its first decade as a nonprofit with a capped-profit subsidiary, an arrangement where investor returns were capped at a multiple and everything above that flowed back to the mission. In October 2025 that ended. The for-profit arm became OpenAI Group PBC, a Delaware public benefit corporation. The cap is gone. The nonprofit, renamed the OpenAI Foundation, holds roughly 26% of the equity. Microsoft holds roughly 27%. And critically, the Foundation still controls the board.^[7,8]

Pair that with the governance track record. In November 2023 this board's predecessor fired the CEO with no warning to its largest commercial partner, and reversed itself in five days under employee pressure.^[15] The people changed. The structure that made it possible, a mission-controlled board sitting above the commercial entity, did not. If you are underwriting a trillion-dollar company, that is a risk factor you should want priced, and I do not think it is.

What I would actually do

I am not going to tell you whether to buy this. I will tell you how I think about mega-IPOs, because the mechanics matter more than the narrative.

• Retail does not get the IPO price. You get the price after the first print, which in a hot deal is frequently 30% to 80% above the price the institutions paid. The pop is the underwriter rewarding its book. You are the exit liquidity for that pop, not the beneficiary of it.
• Read the public S-1 before you form a view. Not the coverage of it. The document. Go to the risk factors, the customer concentration disclosure, the compute purchase obligations table, and the related-party transactions with Microsoft. That last one is where the interesting reading will be.
• Watch the lock-up. Insiders are typically restricted from selling for 180 days.^[2] With this much employee paper outstanding and a secondary market that has already let people cash out once, day 181 is a real supply event.
• Size it like a venture bet, not a core holding. If this works, it is a 5x over a decade. If the AI capex cycle cracks, it is down 70% and takes years to recover. Both are inside the distribution of outcomes. Position accordingly.

The honest summary

Saudi Aramco raised about $29.4 billion in 2019 and Alibaba about $25 billion in 2014. Those are the records OpenAI would be breaking.^[14] A company with a five-year commercial history, a nonprofit board above it, and a fourteen-figure compute bill it cannot currently pay is asking to be the biggest offering ever made.

The way I think about it: OpenAI is a genuinely great business trapped inside a genuinely aggressive price. Nine hundred million people using your product weekly is one of the most valuable assets ever assembled. But the price already assumes the agents land, the enterprise pivot works, the token-price collapse does not eat the margin, the board never does anything mission-driven and stock-hostile, and the competitor currently out-earning them does not compound faster. That is five things going right in a row.

The delay to 2027, if it holds, tells you the bankers know it too. You do not push a hot deal unless you are worried the book is thin. Altman wants a trillion. The market has not yet said it will pay. That conversation, not the S-1, is the actual story here, and it will get settled in public whenever the roadshow finally starts.