OpenAI's Two Trillion Dollar IPO Math: Can Nvidia's Holdings Close the Gap?

A Two Trillion Thought Experiment (0:00)

Hey, chibis! I'm Aiwee, and today we're talking about the two trillion dollar AI IPO the math refuses to reach. If you enjoy stories like this, hit the like button and subscribe if you haven't already — let's go! Imagine paying a hundred dollars today for the right to collect a single dollar a year, every year, forever. That is the math of a perpetuity at one percent.

Now imagine paying that same hundred dollars when the going rate is more than five percent. Your perpetual dollar, discounted at today's yields, is suddenly worth less than twenty dollars. This is not an abstract exercise. It is the arithmetic the public market is being asked to perform on a single artificial intelligence company, and the difference between the two answers is roughly three quarters of a trillion dollars. A frontier AI company is anticipated to become a publicly traded company in the next several months.

Its reported target is around two trillion dollars, a figure that, as The Economist has noted, is roughly equal to the combined value of the ten largest technology initial public offerings in history. The filing was expected as early as late August and had not appeared at the time of recording. A peer lab has pushed its own listing to the following year. If ordinary methods of valuation cannot reach a number this large, what is filling the gap, and who is on the other side of the trade? The setting is unusual.

A purchasing managers survey showed United States business output growing at its fastest pace in five years, and the Nasdaq sits at a record high. Yet initial public offerings are being pulled.

Who Is on the Other Side (1:45)

Jay Ritter of the University of Florida told Bloomberg the pullback is especially surprising. The reason is the rate environment. The ten year Treasury yield reached five point two three percent on a Thursday evening, the highest since two thousand four. Matt Kennedy of Renaissance Capital told Bloomberg rising rates are a double whammy for growth stage issuers, shrinking the present value of distant profits and raising the cost of borrowing to build physical infrastructure. Holtec postponed its listing the prior week, and five of the ten largest listings of the year are trading below their offer prices.

The last comparable rate hike into boom episode was nineteen ninety nine. So what does the company actually look like on a spreadsheet? Its annualized revenue was running at approximately sixty five billion dollars in August, a figure characterized by The Financial Times Lex column as unofficial and to be handled carefully. Revenue had grown more than tenfold year over year. Eighteen months earlier, internal forecasts predicted twenty twenty seven revenue of twelve billion dollars, a number that, in a story full of optimistic projections, has turned out to be on the low side.

To translate that revenue into a price, it helps to revisit a frame set down in two thousand two by Scott McNealy, the co founder of Sun Microsystems, the company whose servers ran much of the early internet. Speaking to BusinessWeek after the dot com crash, McNealy explained what it meant to pay ten times revenues for a stock.

Rates, the Nasdaq, and 1999 (3:30)

To return the investment in ten years, he said, a buyer would need one hundred percent of revenues, paid out as dividends, for ten straight years, with zero cost of goods, zero expenses, zero taxes, and zero research. At the two thousand peak, the ten year Treasury yield was about six and a half percent, and discounted at that rate, even the rosy ten year payback took roughly seventeen years in present value terms. The number on the table this week is roughly thirty one times revenue. Assume no costs, no staff, no taxes, and no electricity bill, and pay out every dollar to shareholders forever. Discount that perpetuity at today's Treasury rate and the present value is about one point two seven trillion dollars, roughly three quarters of a trillion short.

If conventional math will not reach two trillion, the markets turn to a different tool, the total addressable market. The metric took on its modern form in the late nineteen nineties when the internet analyst Henry Blodget argued that Amazon, then at two hundred and forty dollars a share, was worth four hundred. He was later permanently barred from the securities industry. The Financial Times Lex column recently walked through how the AI industry's TAM has been expanding. In May, SpaceX told investors the addressable market for enterprise applications, including AI, was twenty two point seven trillion dollars, roughly three quarters of United States gross domestic product.

Last week, Morgan Stanley estimated generative AI could tap sixty trillion, about half of world output. In four months, the cited addressable market grew by thirty seven trillion dollars. Uber cited a twelve point three trillion dollar TAM at its twenty nineteen listing, and its current annual revenue is under sixty billion.

Thirty One Times Revenue (5:15)

WeWork cited three trillion dollars and later went bankrupt. Behind the headline number sits a network of cross holdings that is harder to see. SoftBank, the parent of SB Energy, is marketing more than eleven billion dollars of new bonds to fund its next investment in OpenAI. Those bonds priced at nine point seven five percent. The Financial Times reports OpenAI expects to burn through nearly two hundred and eighty billion dollars by the end of twenty thirty.

OpenAI is SB Energy's main customer, leasing the Ohio campus for twenty years, and OpenAI is also an investor in SB Energy, with warrants that pay out if SB Energy reaches an eighty billion dollar valuation. SB Energy's filings show Nvidia bought one point five billion dollars of shares in August at a ten percent discount to the eventual IPO price. Nvidia has agreed to guarantee up to one hundred and five billion dollars for the Ohio campus, and the campus will run exclusively on Nvidia hardware for twenty years. The Wall Street Journal reports SpaceX is renting compute capacity to the AI lab for one point two five billion dollars a month. Aswath Damodaran of New York University has put it simply.

To value one company now, you have to value four others. The buildings behind these circular flows have a curious property. SB Energy, SoftBank's United States data center developer, is targeting a roughly fifty billion dollar valuation despite not having switched on a single data center. It has never built one itself, though it recently acquired a consultancy that has built fifteen. At that mark, it would trade at roughly four hundred times annualized earnings before interest, taxes, depreciation, and amortization.

The TAM Keeps Growing (7:15)

Its own prospectus states it will need one hundred and seventy four billion dollars to fulfill what it has already promised. Its new debt priced at a nine point seven five percent yield, described as the largest junk bond offering on record, and lenders will not permit dividend payments until twenty twenty nine. The Danish academic Bent Flyvbjerg is cited for the iron law of megaprojects, that they come in over budget, over time, over and over. Paul Kedrosky of Man Group has observed that lenders are financing these as long life infrastructure, while the chips inside will be obsolete in one to two years. The one public company at the heart of this chain is also the cheapest on conventional metrics.

Bloomberg reports Nvidia's annual sales went from about twenty seven billion dollars four years ago to an estimated four hundred and ten billion this year. Nvidia is now the most valuable company in the world, and shares are up more than one thousand six hundred percent over four years. On forward earnings it trades at less than seventeen times, near its cheapest level in over a decade. Investors are paying roughly half as much per dollar of Nvidia profit as a year ago. Jensen Huang said at a Goldman Sachs event this month that Nvidia is the first and only growth of value stock.

Gross margin was seventy five percent last quarter, with analysts expecting it to slip below seventy two percent by year end. David Russell of TradeStation told Bloomberg Nvidia's margins are more likely to compress than expand.

Cross Holdings and Warrants (9:00)

Eli Horton of TCW told Bloomberg Nvidia's multiple only makes sense if AI spending slows. The finance academics Lubos Pastor and Pietro Vernazza have argued that uncertainty about a young company's future can raise its value. Edward Miller wrote in nineteen seventy seven that when pessimists cannot short, optimists set the price. At the same time, the price of what the labs sell is collapsing. The Financial Times reports that cheaper open weight models from Chinese labs, including DeepSeek and Moonshot, are approaching frontier performance and taking share.

On a single Tuesday, two leading labs released models forty percent and fifty percent cheaper than the versions they replaced. A study by Epoch AI estimates the cost of achieving a fixed level of model performance has fallen about thirteen fold per year since twenty twenty three, faster, the researchers say, than electricity, computing, or DNA sequencing. The price of inputs is rising. The price of outputs is falling. Eric Glyman, the co chief executive of Ramp, told the Financial Times his firm cut its AI bill by forty percent by using routers that direct tasks to different models.

There is, however, a real counterweight. Data from OpenRouter shows weekly model usage up about twenty five thousand percent since the start of the prior year. Alain de Vinsky at Yale found twenty two and a half percent of the AI lab's users were still using its models a year later, versus thirteen point two percent for the peer. Ben Thompson has argued that integrating model and tooling can produce lock in.

SB Energy's Curious Books (10:45)

None of this means the underlying technology is a fad. The internet certainly changed the world, but it is entirely possible that AI ends up more like email, something people use every day that nobody makes much money selling. The listing itself is therefore a forensic event. SoftBank's stake in OpenAI was last marked privately at eight hundred and fifty two billion dollars. Bloomberg reported in August that SoftBank has taken out a ten billion dollar margin loan secured against those OpenAI shares.

The peer lab is reportedly in talks to raise privately at one point two trillion. Sam Altman told Fortune that, given AI safety considerations, this would be an ill advised moment to go public. Aswath Damodaran has compared the present situation to the dot com correction, a series of trees falling in the forest until a threshold is crossed. Jim Chanos expects the year to break the record for stock issuance. The academic research on that is not encouraging.

Loughran and Ritter found that issuers of new shares tend to underperform for years afterward. The investor Mike Paulus told the Financial Times that the profit motive is overwhelming executives' calls for caution. So what is an honest viewer to watch for, when the prospectus finally lands? The four numbers that matter are gross margin trajectory, customer concentration, the related party transaction schedule, and the first quarterly disclosure of compute spend as a share of revenue. If those four land cleanly, the math starts to behave.

Nvidia at the Center (12:30)

If they do not, the present value of every dollar of revenue, discounted at five point two three percent forever, is exactly the figure this hour began with. The offering will set the price. The market will do the rest. Thank you for watching. If you found this analysis useful, please subscribe and share it with one person who reads footnotes.

See you next time.