Anthropic's $965B IPO Filing and the S-1 That Will Settle the Bubble Debate

On June 1, 2026, Anthropic confidentially filed its IPO prospectus with the Securities and Exchange Commission. The filing arrived days after the company closed a $65 billion Series H that lifted its private valuation to roughly $965 billion — a number that, on its own, would top OpenAI's $852 billion mark and make Anthropic the most valuable pure-play AI company on the planet that anyone can actually buy shares in. A debut above the trillion-dollar line is now the base case, not the moonshot.

That is the headline, and it is the least interesting part of the story.

The interesting part is the document itself. For five years, the frontier AI industry has run on numbers that no outsider could verify — valuations set in private rounds, revenue figures quoted in press releases without an auditor's signature, gross margins that the labs treat as state secrets, and compute commitments measured in tens of billions of dollars that exist nowhere a short-seller can read them. The entire "is this a bubble" argument has been conducted on vibes, because vibes were the only inputs available.

A registration statement ends that. When Anthropic's S-1 becomes public — and the confidential filing is the step that guarantees it eventually will — the most guarded numbers in technology become audited, itemized, and falsifiable. The IPO is a financing event. The S-1 is an epistemic one. It is the moment the AI bubble debate stops being a matter of taste and becomes a matter of evidence.

This piece is about what becomes knowable, why it matters more than the valuation, and what builders, buyers, and operators should actually do with the information when it lands. The short version: the most important AI document of 2026 will not be a model card. It will be a prospectus.

The Numbers We Already Have

Start with what is on the table before the S-1 fills in the rest. Three companies are converging on public markets in 2026 at valuations that would have been unthinkable a few years ago, and the contrasts between them are where the analysis begins.

SpaceX is targeting a valuation of at least $1.8 trillion in a June debut that could raise as much as $75 billion — more than twice the largest IPO in history. OpenAI sits at $852 billion after a $122 billion round and is preparing its own confidential filing with Goldman Sachs and Morgan Stanley, eyeing a possible fall listing. Anthropic, at $965 billion, has now moved first among the AI labs.

Valuation is the number everyone quotes. Revenue is the number that decides whether the valuation is a thesis or a fantasy. And here the picture inverts in a way that should reframe how you read the whole cohort.

Anthropic's revenue run rate reached roughly $47 billion in May 2026, up from $10 billion in annual revenue a year earlier — a near-fivefold expansion in twelve months. That figure is not just large; it is larger than OpenAI's, which sits around $25 billion annualized, growing at about $2 billion per month. The company that built its brand on caution and safety research is now, by revenue, the biggest pure-play AI business in the world, and it got there spending materially less to train its models than its closest rival.

Put valuation and revenue together and the multiples tell a story that cuts against the lazy "all of AI is overpriced" reflex.

At roughly 20 times its run rate, Anthropic is — astonishingly — the cheapest name in the cohort on a price-to-sales basis, despite carrying the highest valuation among the AI labs. OpenAI trades near 34 times. SpaceX, priced on the future of Starlink and launch rather than today's $18.7 billion in revenue, sits near 96 times. The market is not pricing these three companies on one "AI bubble" thesis. It is pricing them on three different stories, and only the S-1 can tell us whether the cheapest-looking story is actually the soundest one.

Why the S-1 Is the Real Event

A confidential filing is not a public document — yet. But it is the irreversible first step toward one. The Jumpstart Our Business Startups (JOBS) Act lets emerging-growth companies file privately so the SEC review can begin without the numbers being exposed during the back-and-forth. The trade is simple: the company gets to negotiate disclosures in private, but when it moves toward an actual offering, the prospectus must be made public at least fifteen days before the roadshow. The confidential filing is the commitment device. Once it is in, the disclosure is coming.

So the right question is not "what is Anthropic worth." It is "what will we learn that we could not learn before." A frontier-lab S-1 forces daylight onto the exact variables the bubble argument has been starved of.

Consider the first line, because it is the one that decides everything else. The central unknown of the AI economy is whether serving a model at scale is a high-margin software business or a low-margin utility business dressed in software's clothing. If inference gross margins are 70 percent and improving, these valuations are defensible and possibly cheap. If they are 25 percent and compressing under a price war, the entire sector is mispriced and the bears are right. Nobody outside the labs knows which world we live in. The S-1 will say.

The second line matters nearly as much. The labs have signed enormous, long-dated commitments to buy compute — the obligations that fund the data-center buildout everyone can see from satellite photos. In a private company those commitments are a rumor. In a public filing they are a liability schedule with dates and dollar amounts, and they reframe the revenue: a $47 billion run rate looks very different against $10 billion of annual compute obligations than against $40 billion of them. We are about to find out which it is.

The Reversal Nobody Priced In

Spend a moment on the fact that Anthropic passed OpenAI on revenue, because it quietly rewrites the competitive map. Eighteen months ago the consensus was that OpenAI's consumer reach — ChatGPT's hundreds of millions of weekly users — gave it an unassailable revenue lead, and that Anthropic was the smaller, safety-focused sibling that would always trail on the top line.

The crossover happened because the two companies monetize different things. OpenAI's revenue is weighted toward consumer subscriptions and a still-young advertising motion; Anthropic's is weighted toward enterprise and developer API consumption — Claude embedded in coding tools, in customer-facing agents, in the back-office automation that enterprises pay for by the token and that grows with usage rather than with seat count. Enterprise API revenue compounds quietly and without churn drama, and it turns out to compound fast.

The efficiency angle sharpens the point. Anthropic reportedly reached this revenue position while spending roughly four times less to train its frontier models than OpenAI. If that holds up in the prospectus, it is the most important competitive fact in the industry, because it means the revenue lead is not being bought with a proportionally larger cost base. That is the difference between a company growing into its valuation and one inflating toward it — and it is exactly the kind of claim the S-1 will either substantiate or puncture. I have argued before that the AI valuation question is ultimately a margin question, not a hype question, and the filing is where that argument finally meets audited numbers.

Why File Now, and Why First

Timing in an IPO is rarely an accident, and Anthropic's choice to move first among the AI labs is strategy, not impatience. Three forces converge on June 2026 to make this the right window, and reading them tells you as much about the company's posture as the financials will.

The first is the comparable. The first frontier lab to file gets to define the template by which the others are judged. When OpenAI's prospectus lands in the fall, analysts will not read it cold; they will read it against Anthropic's already-public disclosures, line by line. By filing first, Anthropic gets to set the terms of comparison — to make its gross-margin profile, its revenue mix, and its growth curve the reference point against which a rival is measured. In a two-horse race for the same pool of public capital, being the benchmark rather than the challenger is worth real money.

The second is the market window. Equity markets open and close for new issues in ways that have little to do with any single company's quality. Right now the window is open: appetite for AI exposure is high, the indices are near records, and a prominent analyst is publicly calling this "the opening of the floodgates." Windows like this do not stay open indefinitely. A company sitting on a $965 billion valuation and a fast-growing revenue base has every incentive to file while demand is hungry rather than wait for a quarter when sentiment has cooled and the same numbers fetch a lower multiple.

The third is the discipline argument, and it is the one that gets the least attention. Going public is not only a financing event; it is a forcing function. The reporting obligations of a public company — audited statements, quarterly guidance, a board answerable to outside shareholders — impose a kind of operational rigor that fast-growing private companies often defer. For a company that intends to be spending tens of billions of dollars a year on compute and research for the foreseeable future, the governance maturity that comes with public-company status is a feature, not a tax. It signals to enterprise customers, who are betting their own roadmaps on the vendor's longevity, that the company is built to last beyond the current funding cycle.

There is a fourth, quieter force: liquidity for the people who built the company. Employees and early investors holding paper worth a notional fortune cannot pay a mortgage with an illiquid private valuation. After five years and a near-trillion -dollar mark, the pressure to create a real market in the shares is substantial, and it grows with every funding round that raises the paper value without providing an exit. A public listing converts the abstraction into something spendable. None of these forces is unique to Anthropic, which is exactly why the filing reads less like a one-off and more like the starting gun for the cohort.

The Trillion-Dollar Cohort and the "Floodgates"

Anthropic did not file into a vacuum. It filed into a forming wave, and one prominent analyst described the moment as "the opening of the floodgates for the IPO market." Three of the largest private companies in the world are queuing for public markets in the same calendar year.

A wave of mega-IPOs in a single year is exactly the pattern that invites dotcom comparisons, and the comparisons are flying. They deserve to be taken seriously and then taken apart, because the analogy is half right in a way that is more useful than either "it's 1999 again" or "this time is different."

Is It a Bubble? Do the Comparison Properly

The 1999 comparison fails on the one variable that actually defines a bubble: revenue. The canonical dotcom disasters — the pet-supply sites, the grocery delivery dreams — went public with negligible sales and valuations justified entirely by projected eyeballs. There was no there there. The 2026 cohort is the opposite case on that axis: these are companies with tens of billions of dollars in real, recurring, fast-growing revenue going public, not pre-revenue concepts.

Where the analogy does hold — and where the bulls wave it away too quickly — is on two points the headline revenue figures obscure. The first is capital intensity. Dotcom companies were capital-light; you could run one out of a loft. Frontier AI is the most capital-intensive software business ever built, and the compute commitments funding the buildout are a structural liability that a pure software comparison misses entirely. A 20-times-revenue multiple on a business with software margins is one thing. The same multiple on a business that must commit tens of billions to chips and data centers years in advance is a different risk profile, and only the liability schedule in the S-1 will show which one we are actually buying.

The second is circular financing — the web of arrangements in which chip makers, cloud providers, and model labs invest in, lend to, and buy from one another, so that a dollar of "revenue" at one node can trace back to a dollar of investment from another. Some of this is normal ecosystem behavior. Some of it inflates apparent demand in ways that only consolidated, audited disclosure can untangle. This is precisely the dynamic I expect to drive the mark-to-market reckonings I've predicted for the cloud giants' AI stakes, and a public S-1 is the instrument that turns "I suspect circularity" into "here is the related-party transactions footnote."

The honest position on the bubble question is therefore not "yes" or "no." It is "we are about to be able to answer it." That is the genuinely new thing about June 2026: for the first time, the argument has a referee. The numbers that settle it are being filed.

The Margin Question, Quantified

Because the inference gross margin is the variable that decides whether these valuations are sober or silly, it is worth making the stakes concrete. The same $47 billion of revenue supports wildly different conclusions depending on what it costs to produce. Here is the spread the prospectus will collapse into a single audited number.

In the bull case, inference behaves like software: most of the revenue drops through as gross profit, the compute spend is an investment in capacity rather than a cost of goods that scales one-for-one with usage, and a 20-times-revenue multiple looks conservative for a business compounding at this rate. In the bear case, inference behaves like a utility: the cost to serve rises nearly in step with revenue, a price war keeps prices falling faster than costs, and the gross profit left over does not come close to justifying the valuation once you net out the research and compute obligations. The base case sits in between and is, for what it is worth, where most careful observers land — good but not magical margins, improving with scale and silicon efficiency, dependent on the price war not turning ruinous.

What makes this more than an academic exercise is that the answer is not stable. Margins are a moving target in a market where the price of a token has been falling for two years. A lab can report a healthy gross margin in one quarter and watch it compress in the next as a competitor cuts prices to win share. The S-1 will show a snapshot and a trailing trend; the trend matters more than the snapshot, because it tells you which direction the most important number is moving. A 50 percent margin rising toward 60 is a different investment than a 60 percent margin falling toward 50, even though they cross at the same point.

This is also where the efficiency claim does its real work. If Anthropic truly trains its frontier models at a fraction of a rival's cost, that advantage shows up not in the training line — which is lumpy and capitalizable — but in the durability of the inference margin under competitive pressure. A lower cost base is what lets a company survive a price war without bleeding, and the prospectus is where that resilience either appears in the numbers or fails to. Watch the gross-margin trend and the cost-of-revenue footnote together; read in isolation, either one can mislead, but together they are the closest thing to ground truth the industry has ever published.

The Repricing of the Private Market

There is a second-order consequence of all this that will land far from Wall Street, in the term sheets of companies that have nothing to do with Anthropic. For years, private AI valuations have floated free of any public anchor. A Series B startup could justify a rich round by gesturing at the labs: if Anthropic is worth this in a private round, the logic went, then surely a smaller company riding the same wave deserves its own generous multiple. The argument worked precisely because there was no public comparable to check it against.

A public Anthropic supplies that anchor, and anchors are unforgiving. Once there is an audited, daily-priced, liquid benchmark for what frontier AI revenue is worth — a real price-to-sales ratio that updates every time the market opens — every private round in the sector gets measured against it. Late-stage AI startups that raised at lofty multiples on the strength of narrative will find their next round priced against a public number instead of a press release. If the public multiple lands lower than the private market assumed, the repricing ripples outward and downward: down-rounds, recapitalizations, and the quiet re-marking of venture portfolios that had been carrying these positions at the last round's optimistic figure.

This is the mechanism I expect to drive the mark-to-market write-downs I've forecast for the cloud and platform giants holding large AI stakes. A public comparable does not just price Anthropic; it prices everyone who has been valued by reference to Anthropic. The labs were the ceiling that held up the whole structure of private AI valuations. When the ceiling becomes a public, moving number, everything hanging from it moves too.

The direction of that move is not predetermined. If the S-1 reveals genuinely strong economics — healthy, durable margins on real and growing revenue — the public anchor could validate the private market and even pull valuations up, giving every AI startup a richer comparable to point at. The bull and bear cases diverge entirely on the gross-margin line, which is why that single disclosure carries so much weight. But either way, the era of private AI valuations set in a vacuum is ending. From the moment Anthropic trades, the sector has a price, and prices have a way of correcting the stories that were told before they existed.

For founders raising in this environment, the practical takeaway is to assume your next valuation conversation will reference a public multiple you do not control. Build the business so the comparison flatters you — real revenue, a defensible margin story, low customer concentration — rather than hoping the narrative holds. The narrative era is closing. The numbers era, ushered in by a single confidential filing, is opening.

The Mission Tension a Public Anthropic Inherits

There is a second story inside the filing that has nothing to do with multiples and everything to do with what Anthropic is for. The company was founded on a safety-first premise and structured, in part, around the idea that it would sometimes choose caution over growth. Public markets are not built to reward that trade-off. They are built to reward the next quarter.

The friction is real and worth stating plainly. Safety research is expensive, its payoff is diffuse and long-dated, and it does not show up cleanly in a revenue line. A public Anthropic will face shareholders who can read the R&D spend in the 10-K and ask, every ninety days, why a given fraction of it is going to alignment work that does not obviously move the top line. The governance structures the company has built to protect its mission will be tested not by a crisis but by the grinding, ordinary pressure of being a public company that must explain itself four times a year.

This is not a reason to expect the mission to collapse. It is a reason to read the governance section of the prospectus as carefully as the financials. How voting control is structured, what the board can and cannot be forced to do, and how the public-benefit commitments are encoded will tell you whether the safety-first identity is load-bearing architecture or marketing that bends under the first sustained quarter of disappointing guidance. Markets have a way of discovering which is which.

What This Means for Builders and Buyers

Strip away the spectacle and there are concrete moves to make. The filing changes the information environment, and information environments are where leverage lives.

For builders and procurement teams, the most actionable consequence is pricing. A public lab is a lab under pressure to show a path to margin expansion. That pressure cuts two ways for you. In the near term, intense competition among labs racing toward — or already on — public markets keeps inference prices falling, which is good for anyone shipping on top of these models. In the medium term, once a lab is public and must defend its margins to shareholders, the era of selling tokens below cost to win market share gets harder to sustain, and discount-driven pricing can firm up. The window to lock in favorable terms is while the competition for share is still the dominant incentive.

The defensive posture writes itself: do not single-source your intelligence. The same logic that drove Microsoft to build its way off its dependence on OpenAI applies, in miniature, to every company that has wired a single vendor's API into the core of its product. A multi-model architecture is not just resilience engineering; it is negotiating leverage against vendors whose pricing power is about to be re-shaped by the demands of public shareholders.

For investors and observers, the discipline is to wait for the document and read the unglamorous parts. The valuation will be debated endlessly on television. The liability schedule, the related-party footnotes, the deferred-revenue line, and the cohort retention disclosures will be debated by almost no one — and they are where the truth is. The same scrutiny applies to the rest of the wave, including the SpaceX IPO that anchors the trillion-dollar end of the cohort; read each prospectus for what it is forced to reveal, not for what the roadshow wants you to feel.

A Falsifiable Prediction

Analysis that cannot be wrong is not worth much, so here is a claim with a date on it. Given a $965 billion private valuation, a $47 billion and fast-growing revenue run rate, and a base case that already points above the trillion-dollar line, I expect Anthropic's IPO to price at a fully diluted valuation of at least $1 trillion before the end of Q2 2027. The risk to that call is not demand — it is the market window: a sharp repricing of the AI sector, triggered by exactly the margin or circularity disclosures the S-1 will contain, could push the timing out or pull the valuation down. That two sided risk is the point. The prediction is falsifiable precisely because the document that will prove or disprove it is the same document that settles the bubble debate.

Conclusion: The Document, Not the Debut

It is easy to treat the IPO as the milestone — the bell-ringing, the first-day pop, the founder on the exchange floor. That is theater. The substance happened on June 1, when a frontier AI lab committed itself to disclosure. The number that filed was $965 billion. The thing that filed was the end of an era in which the most important economic questions about artificial intelligence could only be argued, never answered.

When the prospectus goes public, the AI industry's most closely held numbers — the cost to serve a token, the size of the compute obligations, the concentration of the customer base, the true shape of the margins — become audited fact. The bears will find their evidence or lose their argument. The bulls will be vindicated by the gross-margin line or quietly abandoned by it. Either way, the debate that has dominated technology for three years gets a referee. The most consequential AI release of 2026 will not have a model card. It will have a registration number.

For continuing coverage of the filing and the broader listing wave, see our news analysis of the trillion-dollar IPO wave taking shape.