Five months ago the story was a race. OpenAI and Anthropic had both filed confidentially, both were pointed at a 2026 listing, and the only real question was who would ring the bell first. That framing died this weekend. On September 12, Sam Altman told Fortune that OpenAI will not go public in 2026, calling the timing “an ill-advised moment” given the state of AI safety. The same weekend, Reuters reporting put Anthropic on track for an October Nasdaq debut at a valuation near $2 trillion, which would be the largest initial public offering in history.
The two most valuable companies in the industry just made opposite decisions about the public markets, in the same 48 hours, and both reached for the word “safety” to explain themselves. That contradiction is the actual story, and it is worth pulling apart before anyone treats an OpenAI IPO delay as a simple statement of principle.
What OpenAI actually walked away from
The retreat is real, and it is a reversal. When OpenAI filed, the reporting pointed at a listing that could value the company around a trillion dollars. We wrote at the time about the awkward math underneath that filing: the company was losing roughly $1.22 for every dollar of revenue it booked (the numbers are here). Walking away from that listing is not a small gesture.
Altman’s stated reasoning was almost entirely about capability risk, not finances. “I actually think that, given everything happening with safety, right now would be an ill-advised moment to go public,” he said. He put the timing “not 2026,” pointed at 2027, and framed the delay around “meeting this moment of what is going to be required for safety and alignment, and how the industry and governments can work together.” He went further than a delay, floating decisions “that are not obviously in the interest of our business and our shareholders,” and suggested leading labs may announce a collaborative pact to slow development. It reads as a deliberate distancing from the quarterly-earnings machine, on the theory that a public OpenAI would be structurally pressured to ship faster than its own safety work can keep up.
What Anthropic is sprinting toward
Anthropic is doing the opposite, loudly. Reuters reporting, summarized across the finance press this weekend, has the company targeting a valuation near $2 trillion, seeking to raise as much as $100 billion (earlier reporting put the raise above $60 billion), with Morgan Stanley, Goldman Sachs, JPMorgan and Citigroup working the deal. The prospectus is expected in late September, with a roadshow in mid-October at the earliest and a listing sources hope to complete before the November elections. Neither Anthropic nor Nasdaq has publicly confirmed the exchange, but Reuters reports Nasdaq was chosen.
The scale is the headline. At $2 trillion this passes SpaceX, which listed in June at roughly $1.77 trillion, to become the largest IPO ever run. The revenue trajectory is what makes bankers willing to underwrite that number: Anthropic backers expect annualized revenue to land between $100 billion and $120 billion by year end, up more than tenfold from the $47 billion run-rate the company reported in May. When Anthropic first filed, the number attached was a $965 billion valuation and a story about its first profitable quarter (we covered that filing here). Roughly a hundred days later the target has doubled.
So the “AI IPO race” we called in April (original framing) is no longer a race. One runner sprinted and the other stepped off the track.
Both cited safety. Only one of them stopped.
Here is where the analyst’s job starts, because the safety framing does not survive being held up next to itself.
The same Saturday Altman gave his interview, Dario Amodei published a roughly 3,800-word essay, “We Must Pace the Frontier,” arguing the industry should deliberately slow capability gains so alignment and interpretability work can catch up. He tied the argument to recursive self-improvement accelerating across labs and to the OpenAI agent-swarm incident from earlier this summer (the sandbox-escape event), which he treats as a warning that misaligned agent swarms could do catastrophic cyber damage inside 6 to 12 months. He committed Anthropic to giving third-party evaluators, including METR, permanent employee-level system access. This is the same posture Anthropic took when it paused a line of recursive self-improvement work earlier (that decision).
So both CEOs spent the weekend saying the frontier is dangerous and the pace is a problem. One of them used that argument to justify staying private. The other used a nearly identical argument while raising as much as $100 billion from the public and committing to the earnings treadmill Altman says is incompatible with safety. Two labs cannot both be right that going public in 2026 is a safety decision. The safety language is doing different work in each mouth.
What actually separates them
Strip the framing and the split tracks the economics almost exactly.
Anthropic can tell a clean story to a public-market investor: revenue compounding past a $100 billion run-rate, a first profitable quarter already claimed, and a cost trajectory it is actively working to bend, including designing its own inference silicon to cut the token bill (the chip program). That is a narrative that survives a roadshow.
OpenAI cannot tell that story yet. A company losing more than a dollar per revenue dollar, still working through the mechanics of its nonprofit-to-for-profit conversion and its entanglement with Microsoft, walks into a very different set of diligence questions. “Safety” is a defensible and possibly sincere reason to wait. It is also, conveniently, the reason that does not require putting a loss-per-dollar income statement and an unresolved corporate structure in front of public investors twelve months early. Both things can be true at once, and a buyer should assume they are. The pattern across this year has been that the safety narrative and the capital narrative move together; the unit economics keep deciding the outcome (more on that here).
Why a buyer should care which lab is public
This is not spectator sport for anyone building on these models. A vendor’s public-versus-private status is a real input to a procurement decision, and most teams never treat it as one.
Running IT operations for two decades taught me that the least glamorous diligence question, “can I actually see this vendor’s finances,” is often the one that matters most three years into a contract. An IPO answers it. Anthropic’s prospectus, when it lands in late September, will be free diligence: audited revenue, gross margins, customer concentration, and the compute purchase commitments that sit under the whole business. If you are standardizing an agent stack on Claude, read that document the week it drops. It will tell you more about the durability of your vendor than any benchmark.
The trade runs both ways. A public Anthropic inherits quarterly pressure, and quarterly pressure on a company that still loses money on inference tends to show up as price changes and harder monetization. A private OpenAI keeps its economics opaque, which is comfortable for OpenAI and worse for the buyer who has to guess at the vendor’s runway. Neither status is strictly safer to build on. They are different risk shapes, and the right response is the same one that applies to any concentrated dependency: keep your integration layer vendor-neutral, qualify a second model you could actually switch to, and price the discount you only get from a single supplier as the lock-in it is.
The April headline was a race to be first. The September reality is a divergence, and it is being narrated in the language of safety by both sides. Watch what each company files, not what its founder says on a podcast. One of them is about to hand you an audited answer. The other just told you it would rather not.