Anthropic's $2 Trillion Public Offering: Has the Bubble Test Begun

Anthropic's listing is the moment of truth for money flowing into machine intelligence. The firm is barely five years old and already counts among the most valuable companies anywhere, yet it is lining up what would rank as the biggest offering ever attempted, at a price tag of $2 trillion. The Economist puts that number in context by noting it roughly matches the ten largest tech listings of the modern era combined. On paper the moment could hardly be better: the Nasdaq sits at an all-time high, and purchasing managers report American output expanding at the quickest clip in half a decade. Still, offerings are being abandoned. The reason is not weak growth. It is the cost of money. That same survey pushed Treasury ten-year yields to 5.23% by Thursday evening, a level not seen since 2007. The year quoted in the video disagrees with these records, and the sources place the peak at 2007 rather than 2004. Renaissance Capital's Matt Kennedy explained to Bloomberg that this hits such companies twice over: distant profits get discounted harder, and the debt needed to put up data centers costs more. CNBC reports that Anthropic has settled on the Nasdaq for its potential offering.

Nuclear power company WholeTech postponed its listing, and SB Energy, SoftBank's data center developer, filed before Labor Day yet still has not begun marketing its shares. Only three IPOs have happened since Labor Day, and five of the ten largest listings this year now trade below their offer price. The Federal Reserve raised rates earlier this month, and the last time it did so into a boom like this was 1999. Here the speaker turns to a line Scott McNeely, founder of Sun Microsystems, delivered in an interview in 2002. A 10 times revenue multiple means 100% of revenue paid out as dividends for ten straight years to deliver a ten-year payback. McNeely argued that scenario was impossible even with zero cost of goods sold, zero expenses, zero taxes, zero R&D and 39,000 employees, and he closed by asking whether anyone would like to buy his stock at $64. It remains one of the few occasions when a CEO explained in detail why investors should not have bought his own shares. CNBC reports the 10-year yield reached its highest level in nearly two decades, with AI-related borrowing adding to bond supply.

The speaker then makes the test harsher. If the 10-year yield was around 6.5% at the peak in early 2000, a 10 times revenue multiple without costs, taxes or R&D takes seventeen years to pay back, not ten. At a $2 trillion valuation Anthropic trades at roughly 31 times revenue. His assumptions are more generous still: no costs, no staff, no taxes, no electricity bill, every dollar paid to shareholders forever, and discounting at the Treasury rate instead of adding a risk premium. Under those conditions payback never arrives, and that perpetual revenue stream is worth about $1.27 trillion today. The shortfall exceeds $730 billion. So where does $2 trillion come from? From a metric bankers normally avoid: total addressable market. The concept was popularised in the late 1990s by internet analyst Henry Blodgett, who argued Amazon was worth $400 a share while it traded at $240. Amazon reached that level within weeks, Blodgett became famous, and he was later banned from the securities industry for life. The Financial Times' Lex column traces how market sizing inflated in AI: in May, SpaceX, better known as a rocket company that also makes the Grok chatbot and owns Twitter, told investors the enterprise application market including its AI products was $22.7 trillion, roughly 3% of the entire US economy.

The math behind $2 trillion

Within four months that addressable market expanded by $37 trillion. The Wall Street Journal says Anthropic's filing points to a $30 trillion opportunity, something like a quarter of all output on earth and nearly the whole American economy. Last week Morgan Stanley put generative AI's reachable pool at $60 trillion, close to half a year of everything the planet produces. Lex draws attention to the coincidence that Morgan Stanley probably sits among the underwriters. Bankers are cautious when sizing what they can levy a commission on, and generous when sizing what they cannot. How much of a forecast market a firm truly captures is a separate matter. Uber advertised $122.3 trillion of addressable market at its 2019 listing, and its yearly revenue is now under $60 billion, so 99.5% of that market remains ahead of it. WeWork announced $3 trillion and then collapsed. Anthropic's internal research group modeled a scenario in extreme terms where AI adds upwards of $10 trillion to American output by 2030, which Lex converts with a quick estimate into something near $100 trillion in equity value now. Those comments from McNeely ran in a Bloomberg interview in 2002 and turned into the most quoted valuation warning of the dotcom years.

Past those figures sits the boldest path to any target price: self-improvement that compounds, where a system designs a better system, which designs a better one again, until enough value has been created that currency and asset prices stop carrying meaning. Elon Musk has said that AI and robots will make employment a choice and money irrelevant, a striking posture for someone who has gathered a great deal of it. Google DeepMind's Alpha Evolve turned up a better approach to multiplying certain matrices than the one mathematicians had relied on for 56 years, and House of L AI, a channel belonging to a computer scientist holding a PhD, points out that every impressive outcome there came with a scoreboard, meaning a graded test with a known right answer. Anthropic's own automated researcher closed roughly 97% of the performance distance on its assignment, and part of how it did so involved locating loopholes in the setup. When its strongest concept reached real deployment, the gain measured about half a point, sitting inside statistical noise. That the company released this disappointing result itself says more than most firms would. The direction still shifted: a year and a half ago management expected 2027 revenue near $12 billion, the single cautious estimate in a narrative full of bold ones. In August, Aswath Damodaran, widely read as the dean of valuation, asked whether the figure might retreat from $1 trillion to $800 billion. Seven weeks on, discussion centers on $2 trillion.

Anthropic has found a way to price not just markets that do not exist yet but buildings that do not exist yet. SB Energy is preparing to list at around $50 billion while not having switched on a single data center. The company has never built one itself, having bought a consultancy that built 15. Its prospectus implies roughly 400 times EBITDA, states it needs $178 billion to deliver what it has already promised, and on Thursday priced the largest junk bond offering on record at a 9.75% yield. Lenders will not permit dividends until 2029, and the company has pushed back its listing. In accounting terms the unbuilt data center is a remarkable advantage. A facility still under construction sits on the balance sheet as construction in progress and is not depreciated, and chips bought but not yet switched on are likewise not depreciated. On paper it consumes no electricity, needs no maintenance, and nobody complains about latency because the product does not yet exist. The only problem is that someone will eventually push you to build it. Damodaran points to Martin Flyvbjerg's iron law of megaprojects in Lex: over budget, over time, under benefits, over and over again. Grid connections can take up to ten years in some states. A Gallup poll found 71% of Americans oppose an AI data center near them. Reuters reports marketing starting in mid-October at the earliest, with the listing pushed to November.

So who funds all of it? The design is circular in places. SoftBank is placing more than $11 billion in fresh debt to fund its next installment to OpenAI, and the FT reported that paper priced at a 9.75% yield. OpenAI is expected to consume almost $280 billion by the close of 2030, and part of that goes to a twenty-year lease on SB Energy's enormous Ohio campus, which supplies the revenue SoftBank's data center unit needs to defend its valuation. OpenAI also holds equity in SB Energy, and it carries warrants that pay should the valuation ever touch $80 billion. Nvidia took $1.5 billion in SB Energy stock during August at a tenth below the eventual offering price, and will take a further $1.5 billion at the listing itself. The structure is tidy: SoftBank borrows on junk terms, OpenAI leases the building, Nvidia underwrites it and equips it with Nvidia silicon. Nvidia has committed to back the Ohio project by as much as $105 billion, and the FT says it books no liability at all until OpenAI's leases start in 2028, payment being a campus that runs exclusively on Nvidia hardware for two decades. The Wall Street Journal reports that SpaceX leases computing capacity to Anthropic for $1.25 billion every month. Damodaran's observation is the structural one: he once disliked valuing family-run Asian conglomerates, since pricing one meant first pricing four others, and now pricing Microsoft means pricing OpenAI.

Addressable market inflation

The hardest piece to explain in the whole chain is Nvidia, the firm genuinely supplying picks and shovels to this rush, and its results are strong. Bloomberg reports that revenue moved from roughly $27 billion four years back to an estimated $410 billion in the current year, with net income expected to almost double. Even so, on less than 17 times the earnings expected over the coming year, the stock sits close to the cheapest it has been in more than ten years. Buyers are paying roughly half what they paid a year ago for each dollar of its profit. Chief executive Jensen Huang told a Goldman Sachs gathering that the company is the first and only growth-at-value stock in the world and that the market simply does not understand it. So why does the shovel seller look cheap? The tidy answer, that markets punish what is real and reward what is promised, is too convenient. Four explanations hold up better. Start with the market reading Nvidia as a cyclical business at the peak of its cycle. Gross margin came in at 75% last quarter, and analysts expect it to slip under 72% by the year's end, in part because Nvidia buys memory from suppliers who keep raising prices. Micron stock has climbed more than 180% this year, high-bandwidth memory now accounts for 30% to 40% of what it costs to build an accelerator, and only three companies produce it. Meanwhile the largest customers are designing competing silicon of their own. Nikkei Asia notes that SB Energy's $50 billion target leans on its Nvidia relationship, with no data center revenue behind it. The ifre.com filing shows that most of the 8.8 gigawatts of contracted capacity in the filing sits in an Ohio project that has not broken ground.

For a second reason, what Nvidia books as revenue is in truth other companies' outlays, and much of that spending originates at laboratories burning through cash. Eli Horton at TCW told Bloomberg the multiple is defensible only if spending on AI begins to slow, whether the technology giants pull back or a regulatory regime postpones or halts things, which is very nearly what Anthropic's chief executive has been asking for. On the Monday after his essay ran, the main semiconductor index dropped by almost 6%. The Economist and Lex have separately suggested that deceleration might actually help laboratory margins, since it strips out an enormous computing bill, a clear gain for the labs and a poorer one for the businesses that sell computing. The third reason is that uncertainty can carry value. Analysts have a decent sense of what Nvidia ships next year, whereas a laboratory might fail entirely or take a slice of that $60 trillion pool. Lubos Pastor and Pietro Veresi make the case that ambiguity about a young firm's future earnings can lift what it is worth, while Barbaris and Wang maintain that buyers overpay for anything resembling a lottery ticket.

A fourth reason concerns who actually does the pricing. Nvidia is revalued every second by millions of participants, short sellers very much included, all of whom can position against it. Anthropic's figure, by contrast, was arrived at in closed funding rounds, in some measure by cloud operators whose own earnings improve whenever its mark rises, and a private round offers nobody any way to bet the other way. Edward Miller made this case in The Economist back in 1977, arguing that once pessimists cannot take the other side of a trade, the optimistic view becomes the price. That returns us to Scott McNeely. Priced at $2 trillion, Anthropic fails the reasonableness check he laid out, while Nvidia changes hands at 17 times what it actually earns, and those earnings are still climbing. Of every company in this account, only one clears McNeely's test, and it happens to be the one being marked down.

If that is the case, why not have the laboratories all step back together? A slowdown would also erode the technical advantage that lets them bill at premium rates, and their rivals are not pausing. The FT describes cheaper open-weight models, coming mostly from Chinese developers such as DeepSeek and Moonshot, drawing nearer to what frontier systems achieve and taking share from them. On Tuesday both Anthropic and OpenAI introduced more economical models, their prices around 40% and 50% under the versions they displaced. Customers benefit from a price war, though it is a difficult moment to start one, particularly so close to a listing. Those reductions belong to a wider movement. Epoch AI estimates that reaching a given level of capability has grown about 13 times cheaper each year since 2023, a rate it says no earlier transformative technology matched, electricity and computing and DNA sequencing included. Prices drop fastest immediately after a new leading model appears, which means any premium a lab can charge is short-lived. The asymmetry Epoch identifies is the important part: the boom is raising the price of everything that feeds into it, chips and power and even electricians, while the price of what comes out is in free fall. That is excellent for anyone selling chips and worrying for anyone selling thought.

Pricing buildings that do not exist

Tech executives have long had an elegant reply to cheap competition. McNeely observed once that open source software is free the way a puppy is free, costing nothing at the outset and then requiring years of looking after. That remains the laboratories' best argument against low-cost open models, though everyone knows what inexpensive software running on inexpensive hardware eventually did to Sun Microsystems, which Oracle eventually bought for a fraction of what it had been worth. There is a genuine bullish case as well. Working from OpenRouter data, the FT measured weekly usage rising by roughly 25,000% from the start of last year. Aleh Tsyvinski at Yale calculated that 22.5% of Anthropic users were still with its models a year later, against approximately 13.2% for OpenAI. Eric Glyman of Ramp told the FT that his firm trimmed its AI spending by 40% by sending different work to different models. In his phrasing, you do not require a Ferrari to collect the groceries, and he wondered whether a ceiling on useful intelligence had already been reached. None of this suggests machine intelligence amounts to a passing craze. It may transform the world, and the internet managed that too, though for a long stretch it handed us Yahoo, Lycos and Alta Vista before anyone had heard the name Google. There is a plausible outcome in which AI ends up resembling email: something everyone relies on daily that nobody earns serious money selling. That would be welcome news for users and considerably worse for anyone who paid $2 trillion for the company providing it.

An IPO is the point at which a company and its founding backers decide the moment has come to sell, and equally the point at which this circular arrangement first picks up a genuine market price. For SoftBank that carries a trap, because so much of its balance sheet rests on a stake in OpenAI that was last marked in private at $852 billion. Bloomberg reported in August that the group secured a $10 billion margin loan against those very shares. A listing would let it turn paper value into cash, and would equally put a market verdict on stock that has never traded before. Anthropic's offering will be the first public reading of what a pure AI lab is worth, and it will also set a price on the paper profits held at Amazon and Google. The awkward detail is that OpenAI controls the timing. Sam Altman told Fortune that, given everything unfolding around safety, this would be a badly chosen moment to list, while OpenAI is reportedly discussing a private raise at $1.2 trillion. Evidently $1.2 trillion is a figure meant for professionals. Damodaran noted that the dotcom correction arrived as an accumulation of small things, trees falling until you wake to find half the forest gone. The withdrawn offerings, the junk bond yields and the filings running late may be those small trees, and his message to the AI companies is to show numbers rather than just words. A prospectus is the single document that compels them to. CNBC reports that Anthropic and OpenAI both launched cheaper models on the same day.

Why the shovel seller is cheap

Meanwhile Wall Street is lining up to offload a great many shares. Jim Chanos expects the coming year to set a new high for equity issuance, and the academic evidence on that point offers little comfort. Ritter and Lazon find that companies issuing new shares tend to underperform for years afterwards, while Baker and Wurgler find the whole market does worse in the following years when a lot of new stock is being issued. Those already holding the stock, it seems, have a decent instinct for when to sell it. Mike Paulus told the FT that where lab chiefs urge slowing down and greater care, markets are answering that the drive for profit simply proves too strong. In hindsight, he suggested, we may ask why we did not listen to them. That is sensible counsel where safety is concerned, and it is probably worth extending it to price as well. Should you pay $2 trillion for a number that fails to add up, granting even that costs, taxes and staff are all zero, and should it then not work out, you can expect to be asked the very question McNeely put to investors back in 2002. What were you thinking?