OpenAI Filed for a Trillion-Dollar IPO. The Same Week Anthropic Booked Its First Profit.

The two biggest AI labs in the world spent this week telling investors completely opposite stories. OpenAI confidentially filed its S-1 with the SEC on Friday May 22, targeting a Q4 listing at $852B to $1T while still losing $1.22 for every dollar of revenue. Six days earlier, Anthropic told its own investors it expects to book a $559M operating profit on $10.9B of Q2 revenue, the first profitable quarter in the company's history. Same week, same industry, two unrecognizable income statements.

The headlines have treated these as separate stories. They are not separate. They are the cleanest natural experiment we have gotten on what a defensible AI business actually looks like, and the contrast is doing real work.

What OpenAI actually filed

The S-1 went in confidentially, which means the prospectus stays private until roughly fifteen days before the public roadshow. Goldman Sachs and Morgan Stanley are leading the deal, with JPMorgan in the syndicate. The target listing window is Q4 2026, possibly as early as September. The valuation band being floated to anchor demand is $852B at the floor and $1T at the top, which would make it the largest tech IPO ever printed.

The financial picture underneath is the part everyone is going to spend the next four months arguing about. OpenAI is at roughly a $25B annualized revenue run rate, which lines up with 50M consumer subscribers and 9M business users. The Q1 income statement, according to figures circulating to bankers, shows $5.7B of revenue against costs that put the loss-to-revenue ratio at $1.22 lost per $1 earned. The company's own internal projections, reported across several outlets, put the full-year 2026 net loss between $14B and $17B and have it turning cash-flow positive no earlier than 2030.

Sitting behind those numbers are infrastructure commitments that have crossed $1.15T across Oracle, Microsoft, and Amazon. That number is the reason the IPO needs to print soon. The private market can fund a lot, but it cannot fund a multi-decade compute book at this scale without an equity exit in sight. The S-1 is, in part, the public confirmation that the buildout is too big for the capital pool that has carried it so far.

What Anthropic told its investors

On May 20, Anthropic shared updated guidance with investors. Q2 2026 revenue is now projected at $10.9B, a 130 percent jump from $4.8B in Q1. More importantly, the company expects $559M of operating profit in the quarter. That would be the first operating-profit quarter in the company's history.

The number that actually matters is the unit-economics swing underneath the topline. In Q1, Anthropic spent 71 cents on compute for every $1 of revenue. In Q2, that ratio is projected to fall to 56 cents. That is a 15-point gross-margin lift in a single quarter. It is what happens when an enterprise mix dominates the book and a coding product becomes the lead generator. Claude Code crossed $1B annualized within six months of launch. That revenue line lands at substantially higher gross margin than consumer chat because the cost of serving a paid-developer workload is the prompt and the output, not the marketing engine.

Anthropic has caveated this carefully. The company told investors it may not sustain profitability across the full year because of planned infrastructure spending in the back half. Translation: the $200B Google TPU commitment we covered in the TPU math piece is going to start landing on the income statement. But the fact that they can turn one profitable quarter while sitting on that commitment is still the more important fact. As of last summer, the company's own guidance to investors was no full-year profit before 2028. They beat that timeline by at least two years on a single-quarter basis.

The side-by-side

When you stack the public numbers next to each other, the divergence is easier to see than to argue with.

Anthropic is now running a higher quarterly revenue number than OpenAI, at a positive operating margin, with no IPO filed and no obligation to file one. OpenAI is running at roughly half the quarterly revenue, at a deeply negative operating margin, and is actively preparing the largest tech IPO in history. Both companies are valued in trillion-dollar territory by their respective markets. That is the picture.

Why the two stories actually rhyme

The temptation is to read this as a morality play: discipline beats scale, the lean lab beats the burn-it-all lab. That is not what the data says. What the data says is that these two companies are pursuing different optimization functions, and both, at the moment, are being rewarded.

OpenAI is optimizing for distribution depth. 50M consumer subscribers is a moat you can only build by being first and by spending whatever it takes to stay top of mind. The S-1 is rational against that strategy. If your bet is that the consumer category collapses to one default assistant the way Google won search, you spend everything to be the default, you IPO when the private capital runs out, and you trust the public markets to fund the long tail. The infrastructure commitment is not a bug. It is the moat.

Anthropic is optimizing for unit economics on a narrower surface. Claude Code, the API business, and the enterprise plans all share the property that the customer is paying for the inference token directly. There is no consumer subsidy in the middle of the P&L. When the dominant revenue mix is tokens that customers pay for above cost, gross margin moves the right direction every time inference gets cheaper, every time the model gets more token-efficient, and every time a big-spending account opts into Max. The 71-cent-to-56-cent collapse in compute ratio is what that compounding looks like on the income statement.

These are different bets on what AI economics looks like at scale. One says: the value capture is at the distribution layer, so spend on distribution. The other says: the value capture is at the work-product layer, so spend on the products that generate work. Both can be right. Neither resolves the question of whether the buildout numbers (the $1.15T and the $200B) are eventually paid back by the revenue they enable.

What the S-1 is actually disclosing, and what it is not

The confidential filing is not a public document yet. What is reaching reporters is the topline pitch the bankers are using to anchor pre-roadshow conversations. We will not see the real numbers, including the actual loss, the actual compute commitments, the actual customer concentration, and the structure of the Microsoft revenue share, until roughly 15 days before the public roadshow. That is the next data event that matters on this story.

Anthropic, meanwhile, has no obligation to publish anything. The Q2 numbers are guidance to existing investors. Whether the profit actually materializes will be visible only in the retrospective leak cycle when the next round of fundraising happens. Take the projection seriously, but the discipline of checking it against reality only kicks in when the next financing closes.

Both labs are operating in the regime where the numbers we get are the numbers they want us to get. That is the cost of the private-market era in AI lasting this long. The S-1 changes that for OpenAI, partially, in about four months. Anthropic can keep its books inside its investor deck for as long as it wants, or until it files its own.

Our Take

The single-week pairing of these two announcements is the clearest signal we have gotten that the AI lab category is not going to consolidate into one shape. Two trillion-dollar companies just told the market that the path to durability runs through opposite playbooks. Distribution-first burning into the IPO window, and unit-economics-first compounding to a surprise-profit quarter. Both are now real strategies with real numbers behind them.

The thing I would actually watch is what happens to the third and fourth labs in the next two quarters. If a lab cannot tell either story credibly (not enough consumer distribution to justify the burn, not enough enterprise depth to flip the margin), the capital pool gets a lot less patient. The S-1 is going to set the floor on what public markets will accept. Anthropic's profit print is going to set the floor on what private boards will. The labs that cannot match either are the ones that get repriced or absorbed first.

For builders reading this: the practical takeaway is that the price floors on both consumer and API tiers are going to keep tightening. The S-1 forces OpenAI to defend a unit-economics narrative to public investors. The Anthropic profit print forces every other API vendor to explain why their margin structure looks worse. Cheap stays cheap. Expensive has to justify itself. And the comparison spreadsheet that decides where workloads land just got more interesting than the benchmarks for a minute.