Three S-1s. Ninety days. Two hundred and forty billion dollars of fresh equity. The 2026 AI listing season is not a calendar coincidence. It is an industrial event. Anthropic in October. OpenAI in Q4. SpaceX — carrying xAI inside it like a Trojan horse with a chatbot — in summer. Combined market capitalisation crossing the listing threshold: north of three and a half trillion dollars before any of them trade their first share. There has never been a quarter like this. The dot-com peak does not come close. Saudi Aramco does not come close. The 1880s railroad rush is the only honest analog, and the railroads took fifteen years to do what these three companies are about to do in twelve weeks.
The numbers are easier to write than to absorb. OpenAI is mid-flight on a $122 billion private round at a post-money of $852 billion — the largest financing in the history of Silicon Valley. Anthropic, a research lab that did $1 billion of revenue in December 2024, just printed $30 billion in annualised revenue and passed OpenAI for the first time on April 7th. SpaceX, having absorbed xAI in an all-stock deal that valued the AI lab at $250 billion, is preparing a June roadshow at $1.75 trillion — more than double the largest IPO ever priced. Together, the three listings will deliver up to a quarter of a trillion dollars of public-market capital in a single autumn.
I want to make a stronger claim than “this will be a big year.” The 2026 listing season is the moment artificial intelligence stops being a private-market parlour game and becomes a Wall Street story priced quarter-by-quarter against the rest of the index. Once these companies are listed, every subsequent decision — every capex commitment, every model release, every safety disclosure, every hire — happens under a forward P/E and a sell-side consensus and an activist letter waiting to land. The frontier AI roadmap is about to acquire, for the first time, a share price.
And the three companies walking through that door are walking through it with three radically different theories of what an AI company even is. They will price differently. They will govern differently. They will be coveted by different investors and feared by different regulators. Inside ninety days, the public market is going to perform a price-discovery event the private market has been postponing for four years. By 2028, two of the three will look obviously correctly valued in retrospect. One of them will look catastrophically wrong, and the resulting bear case will be the most-cited finance case study of the decade.
1 · Three bets, three trillion-dollar wagers
OpenAI’s wager is that the company that wins AI looks like Microsoft circa 1995. ChatGPT now reports 810 million monthly actives and one million enterprise customers. Annualised revenue closed 2025 at $20 billion-plus, growing in triple digits. The company will lose roughly $14 billion in 2026 alone and does not expect profitability until 2030. None of which has prevented it from preparing an IPO at approximately $1 trillion. The brand has crossed the threshold — alongside Google, Kleenex, and Photoshop — where the company name is the verb. The bet is that distribution wins, that the consumer funnel feeds the API, that the API funds the next model, and that the next model defends the consumer brand. It is a flywheel argument. Flywheels are wonderful when they spin. The catastrophic-misvaluation case for OpenAI is what happens if the flywheel slows.
Anthropic’s wager is the opposite. Where OpenAI plays the consumer brand, Anthropic plays the enterprise substrate — the model the CFO doesn’t lose sleep over. Eight of the Fortune Ten are Claude customers. Claude Code alone crossed $2.5 billion in annualised revenue, a number most listed software companies would happily take for an entire product line. The headline figure that broke in early April was almost embarrassing in its tempo: $30 billion annualised, up from $9 billion at year-end — a clean three-bagger in 100 days. On April 7, 2026, Anthropic passed OpenAI’s revenue figure for the first time. That bullet point, when it lands in the prospectus, will be one of the most consequential single lines in the financial press of the decade. VC offers have hit $800 billion. Goldman and JP Morgan are the bookrunners. October is the target. Anthropic’s safety-first charter, often dismissed as marketing, is about to be tested by quarterly earnings calls during which the safety stance will need to be defended as a moat rather than a headwind.
xAI’s wager — which is now, functionally and legally, SpaceX’s wager — is that AI does not win as a standalone industry, but as one load-bearing column in a vertically integrated empire. In February 2026, SpaceX absorbed xAI in an all-stock transaction that valued the AI lab at $250 billion and the combined entity at $1.25 trillion. Five months later, that cap table is preparing to come public at $1.5 to $1.75 trillion, with Bank of America, Citi, Goldman, JP Morgan, and Morgan Stanley joint-bookrunning what would be the largest IPO ever priced — more than double Aramco.
The compute story inside the SpaceX hull is, on its own, the most aggressive infrastructure build in the history of computing. Memphis Colossus was assembled in 122 days — a number the data-centre industry literally did not believe was possible until Musk did it — and was doubled in 92 days thereafter to 200,000 GPUs. A second Memphis site is coming online with another 110,000 GB200s. The publicly announced roadmap targets one million GPUs by the end of 2026 — an order of magnitude beyond what any other AI lab has built or committed to build. Investors, by Musk’s structural design, will not be able to buy xAI as a separate equity. They will be buying the entire AI-and-aerospace conglomerate as a single instrument: rockets that recover themselves, a satellite ISP with seven million subscribers, the largest AI supercluster on Earth, a chatbot growing faster than ChatGPT did at the equivalent age, and the world’s most public CEO. He is, in effect, building Berkshire Hathaway with rocket boosters and a one-megawatt language model where the insurance float used to be. Whether the public market accepts that bundle — and at what discount — will be the most interesting price an S-1 has produced in a decade.
2 · The numbers that break the comparable-set
It is worth pausing on the absurdity of the financial profiles these companies will present. Anthropic went from approximately $1 billion in annualised revenue in December 2024 to $30 billion in April 2026 — a thirty-fold expansion in sixteen months. There is no public-software comparable. The closest precedent is Snowflake 2018-to-2020, which expanded ten-fold and was treated as a generational outlier. OpenAI has done roughly the same trick at a higher base — from $4 billion in late 2024 to $25 billion in early 2026. These are growth rates normally seen in the first three years of a company that finishes its first decade at $200 million in revenue. They are being printed at the multibillion mark.
The comp question gets stranger on the cost side. OpenAI’s $14 billion 2026 loss is not a Q1-of-life capital burn. It is a structural compute cost driven by the pricing of inference at frontier scale. The conventional path-to-profitability framework — designed for SaaS companies with $200 million ARR and 80% gross margins — has nothing useful to say about a company that grows tenfold in a year while losing fifteen billion dollars at an $852 billion valuation. The framework needs a rewrite. The rewrite happens during the IPO season.
The compensation line is the dramatic one. OpenAI’s average employee stock-based compensation in 2025 was $1.5 million — the highest figure ever reported by a private US technology company at scale, and approximately ten times the typical Big Tech average. Roughly 46% of the company’s revenue, on a fully-loaded basis, is recognised as SBC. The day the IPO prices, that line item begins washing through public earnings. Meta has been topping the talent war ceiling for a year: a reported $200 million package to lure Apple’s Ruoming Pang to the Superintelligence Lab; $100 million signing bonuses to OpenAI and Anthropic researchers; multi-year retention grants stacked on top. OpenAI is responding by nearly doubling headcount, from 4,500 at the start of 2026 to a target of 8,000 by year end. By 2027, the price of a senior AI researcher will be set not by recruiter offers but by the daily moves of three tickers.
And the dollar flow has a closed loop. Big Tech AI capex in 2026 runs to roughly $700 billion across Amazon, Alphabet, Meta, and Microsoft, of which approximately $450 billion is direct AI infrastructure. The IPO proceeds — on first look a freshly capitalised attack on the hyperscalers — are, on second look, a recapitalisation of the same Nvidia order book the hyperscalers are filling. OpenAI’s recent multi-year Stargate-class commitments to Microsoft, Oracle, and CoreWeave run into the hundreds of billions. Each commitment is a forward contract on Nvidia silicon. The IPO is the financing event that lets the contract get paid. By H2 2027, the AI labs will be returning a large fraction of their IPO proceeds to the hyperscalers in inference credit and bare-metal rental fees, who will return a large fraction of that to Nvidia in chip orders. The cash-flow loop closes tightly enough to be diagrammed on a napkin. IPO investors fund AI labs. AI labs fund compute. Compute funds Nvidia. Nvidia funds TSMC. TSMC funds ASML. The 2026 listing season is the upstream funding event for the entire two-year capex cycle of the sector.
3 · The trial that could detonate the season
Sitting underneath every other story is a single legal proceeding the financial press has, in my opinion, dramatically under-priced. On April 27, 2026 — nine days from the time of writing — jury selection begins in the Northern District of California for Musk v. Altman. The case Elon Musk filed in February 2024 against the company he co-founded, the man currently running it, and the structural conversion that turned a non-profit research charter into the largest for-profit IPO in history. Damages on the table: between $79 billion and $134 billion. Relief sought: the literal removal of Sam Altman and Greg Brockman, and the structural restoration of OpenAI to its original non-profit form. There is no analogous lawsuit anywhere in the modern history of public-company formation.
The trial calendar matters because it sits directly upstream of the IPO calendar. If Musk prevails on any of the surviving claims — fraud, breach of charitable trust, unjust enrichment — the consequences for the OpenAI S-1 are material in a way prospectus drafters do not enjoy contemplating. Damages of even the lower end of the expert range would be the largest pre-IPO contingent liability in capital-markets history. A successful unjust-enrichment finding could give the OpenAI Foundation expanded claims on the for-profit’s economic interest at exactly the moment the for-profit is trying to explain to public investors that the Foundation’s post-conversion 26% stake is the right number. A finding that the for-profit conversion itself was procedurally defective would be an extinction-level event for the listing timeline.
Musk’s pleading posture is one of the more curious in modern litigation. He has amended the complaint such that any damages won are paid not to him personally but to the OpenAI Foundation. He is, in effect, suing for $134 billion that, if won, will be deposited in the bank account of the very entity he is trying to liberate from Altman’s control. Whatever you think about the merits, that posture is the cleanest possible answer to the predictable accusation that the suit is competitive sabotage. It is a near-perfect signaling move. It is also the kind of move that only makes sense if the plaintiff actually believes the founding mission was real, was breached, and is worth the litigation he is paying his own lawyers tens of millions of dollars to pursue.
And there is a structural irony worth naming. If the SpaceX-via- xAI listing prices in June at $1.75 trillion and the OpenAI listing prices in Q4 at $1 trillion, the largest passive index in the world will, within months, hold both stocks simultaneously. The CEO of one position is, at the same moment, suing the CEO of the other for damages exceeding the GDP of seventy-five percent of the world’s nation states. There is no precedent for an index fund holding two positions whose principals are in active multi-billion-dollar federal litigation against each other. The portfolio-construction question this raises — and the fiduciary-duty question underlying it — will be one of the most quietly important threads in institutional asset management of 2027.
4 · The Grok-clause: the strangest underwriter covenant ever printed
While the trial sits upstream, the most-discussed alleged shakedown of the season sits midstream — on the underwriting syndicate of the SpaceX deal itself. According to reporting from the New York Times, Forbes, Reuters, and a dozen others, with no public denial from Musk or SpaceX, advisers vying for a position on the bookrunner ladder were told that participation in the deal required, among other things, that the bank purchase enterprise Grok subscriptions and integrate them into the bank’s internal IT environment. Several syndicate banks — Bank of America, Citi, Goldman, JP Morgan, Morgan Stanley — reportedly agreed, with some now spending tens of millions of dollars annually on the chatbot. Musk also reportedly asked the banks to advertise on X, though press accounts suggest he was somewhat more flexible on that ask.
“It is the strangest covenant any of us have seen on a syndicate invitation. It is also, on the dollars involved, the cheapest thing we have ever bought.”
Whatever you think of the ethics — and the SEC will, in the fullness of time, almost certainly think something — the clause is structurally fascinating. Musk has discovered that the underwriting fee on a $1.75 trillion deal is a greater asset to the banks than any individual line item on their balance sheet, and he has converted that scarcity into Grok ARR. The fee on a $75 billion bookrun runs to roughly $1 billion split across five banks. A $20 million annual Grok contract per bank costs each of them less than 10% of their fee. They take the trade. Musk turns a one-time fee outflow into a recurring revenue line. It is one of the most efficient revenue-engineering manoeuvres in the history of capital markets, and it works only because the IPO fee pool is so concentrated and so contested.
It also tells you something darker about the bargaining power differential between a soon-to-list Musk vehicle and the Wall Street syndicate desks. The institutional infrastructure of a $25-trillion equity-capital-markets business was, on this report, prevailed upon to subscribe to a competing AI vendor as a condition of access to a single bookrun. Whether other AI vehicles attempt similar plays in the August-September kick-off period — and whether the SEC has anything to say about it before the SpaceX shares price — is one of the things to watch.
5 · The day the music starts
The ripples from a successful listing season run in directions that the bookrunners’ pitch decks tend to underweight. Index inclusion is the first wave: S&P 500 and Nasdaq 100 committees will need to decide on these names within months of listing, and inclusion is essentially certain for OpenAI and SpaceX given market capitalisation. Forced passive-flow buying will run into tens of billions in the first month alone. The Aramco precedent, where the Tadawul listing was followed by months of mechanical demand and considerable price distortion, suggests these names will trade away from fundamentals for at least two quarters.
The pre-IPO secondary market dies the day the public listing prices. The opaque, fragmented secondary market in OpenAI and Anthropic shares — estimated at $20 to $40 billion in cumulative volume during 2024-2026 across Forge, Hiive, EquityZen, and direct dealer flow — collapses overnight into a single transparent public price. Hundreds of LP positions across hundreds of funds will need to be marked to the new price the same week. Several growth funds carry meaningful concentration risk to a single name in this category. Endowments, sovereign wealth funds, and family offices will rebalance portfolios that have, for the first time, a verifiable market price for what was previously a thinly-traded private mark.
The comparable set explodes. Public-market software valuation has operated, since roughly 2022, with Microsoft, Google, and Meta as the closest pure-play AI proxies. After the listings, analysts will have three pure-play AI labs to weight against the hyperscalers. Every existing software company gets re-marked relative to the new comps. Snowflake, Databricks (when it eventually lists), Palantir, ServiceNow — the entire enterprise-software stack — gets revalued in reference to a fresh price for the underlying intelligence layer.
And quarterly cadence enters the AI roadmap. Frontier research has been operating on twelve-to-eighteen-month development cycles, with launch timing dictated by capability gates and safety review. Once these companies are publicly traded, the calendar acquires a new master — the sell-side consensus model, with its quarterly revenue line and forward guidance. Investor- relations decks become a strategic document. Capacity expansions need to be pre-announced. Model launches start to align suspiciously well with the start of new fiscal quarters. The tenure of a research VP is no longer evaluated on paper output alone; it is evaluated on whether the work-stream meaningfully moved the needle on the next-quarter print. Before the IPO, an AI lab is a research organisation with a revenue model. After the IPO, it is a revenue model with a research department. That transformation is the most consequential thing that happens to these companies in 2026, and the founders all know it.
6 · The civilizational register
I want to end at a higher altitude than the analytics. The three companies preparing to list are not just commercial enterprises seeking capital — they are, by their own founders’ repeated public statements, the institutions that will determine whether artificial general intelligence arrives in 2028, 2030, or 2035; whether the systems that arrive are aligned, partially aligned, or unaligned; whether the centre of gravity of the technology lives in San Francisco, in Memphis, or in some jurisdiction we have not yet named; and whether the people who get to decide those questions are public-equity shareholders or founder-CEOs whose super-voting stock keeps them in the chair.
The IPO season is the single largest transfer of decision-making authority over a frontier technology since the 1947 atomic-energy debates handed nuclear power to a civilian commission. The choice the public market makes in this autumn will define what AI is allowed to become for at least the next decade. There will be a winner. There will be a corpse. Both will be obvious by 2028, and not before.
I have been writing about the intersection of mathematics, AI, and markets for years now, and I have not before written a sentence as confidently as the one that ends this essay. By the time the calendar turns to 2027, the way the world thinks about artificial intelligence — as an industry, as a research agenda, as a regulated activity, as a class of tradeable assets, as a civilizational project — will have been permanently restructured by what happens in three SEC filing rooms over the next ninety days. Sit close to a screen. Read every prospectus. Note which voices in the press are sober and which are breathless. The titans are about to walk through the door, and they are not coming out the same way.
“The 2026 AI IPO season will be remembered the way the 1880s railroad listings are remembered, the way the 1999 dot-com flood is remembered, the way the 2007 LBO peak is remembered. It will be remembered correctly only by people who watched it happen, wrote it down at the time, and had the discipline not to flinch.”
Get the next essay in your inbox.
Tuesday weekly. Mathematics, finance, and AI — written like an engineer, not a marketer.
Free. Weekly. One click to unsubscribe. Hosted on Buttondown.