In September, Dario Amodei picked the strangest fight in his industry: he asked rival outfits to collectively slow down. “We must slow the pace at which we improve the capabilities of AI models,” the Anthropic chief executive wrote. For a moment, the most safety-conscious laboratory in artificial intelligence sounded like a man reaching for the brake.
Eleven weeks later, his bankers are reaching for the accelerator. On Wednesday (November 18), Anthropic prices its shares between $70 and $78. On Thursday, they will trade on the Nasdaq under the symbol ANT, in a deal that could raise close to $100 billion.
Between the letter and the listing sits a 335-page prospectus that promises to “bring forth an age of miracles” and then spends 94 of those pages, nearly a third, warning how they could go wrong: models that might “resist shutdown,” “conceal or manipulate information,” or behave in ways “resembling blackmail.”
Markets have priced oil shocks, pandemics and world wars. They have never been asked to price, in a single document, a promise of miracles and a warning of extinction in the same prospectus. This week, they will.
The terms first. Anthropic is offering 825 million shares, worth up to $64.35 billion at the top of the range, plus $9.7 billion if underwriters Goldman Sachs and Morgan Stanley take their option.
Separately, Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC Partners will buy $26 billion of stock at the IPO price, in a placement with no lock-up. The valuation sought: about $1.4 trillion fully diluted, roughly $1 trillion counting only shares outstanding.
But notice what happened to that number. Months ago, the market chatter was a $2 trillion valuation. Last week, CNBC reported that some shareholders call even $1.5 trillion “a stretch.” The markdown is less a defeat than a strategy.
Pricing 20% below expectations, as IPO Boutique’s Jeff Zell told Reuters, lets investors bid the deal up rather than face the take-it-or-leave-it terms that soured SpaceX’s debut. The book, Semafor reports, is already six times covered.
The believers arrive with real numbers. Anthropic’s revenue for the nine months to September was $33.24 billion, up roughly 1,200% on the year. Run-rate revenue reached $81 billion by October’s end, adding $10 billion of annualized business in a month.
Adjusted operating income swung from a $1.8 billion loss to a $1.6 billion profit in the first half. Board members plan to buy up to $303 million of the shares themselves. People selling an IPO story don’t usually do that.
The skeptics, however, arrive with a larger number. The filing records a $104 billion net loss for those nine months, mostly an accounting swing on convertible notes. The cash problem is elsewhere: $895.4 billion of committed spending, “substantially all” for compute, against $81 billion of run-rate revenue.
The counterparties’ run-rate revenue?: Google, $121.3 billion; SpaceX, $120.5 billion, cancellable on 90 days’ notice from March; Amazon, $110 billion over ten years, twice the normal contract length; Microsoft, $44.9 billion; neoclouds Nscale and Lambda, $44.6 and $41.3 billion.
A further $125 to $150 billion of Google chips arrives at Anthropic only from 2028; one data-center lease runs 20 years. “Realistically, anything past three to four years, you’re basically throwing darts in the dark,” Corey Quinn of the Duckbill Group told CNBC. All of it lands as interest rates rise and tech stocks wobble nearly everywhere.
Then the circle, which deserves a gimlet eye. The companies buying $26 billion of Anthropic’s stock are substantially the companies Anthropic pays.
Google will own about 13% after listing. It also sells Anthropic chips and backstops $19 billion of Anthropic’s leases with the data-center firm Terawulf. Nvidia is committing a reported $10 billion to the IPO. Anthropic’s spending flows back to Nvidia’s chips.
Follow a dollar around this table, and you will pass it twice. Watchdogs call it “circular financing”, revenue and equity chasing each other until nobody can tell how much income is real.
The defense of the practice is honest enough: the giants would build this capacity anyway, and equity aligns partners. But in a normal market, your suppliers want you to succeed. In this one, they have paid $26 billion, with no lock-up, to make sure that you appear to.
A fourth perspective is the founding purpose, arguably the strangest tension in the filing. Anthropic exists because five years ago OpenAI researchers concluded the future was being built carelessly, and left. Its chief executive puts the chance that “things go really, really badly” at 25%.
The company that asked the industry to slow down in September released two new frontier models just before its roadshow, wrapped in fresh safety evaluations. Now it answers to quarterly guidance, just as America’s midterms install a Congress inclined to regulate AI and data centers.
Listing will force greater disclosure, and shareholders armed with securities law may police what Congress has not. But if Anthropic’s researchers ever implement a pause, it must now be explained to the owners of 825 million shares.
Then there is OpenAI. Anthropic’s main rival raised $11 billion last week against a reported $20 billion target and says it won’t follow with its own IPO. Sam Altman calls a 2026 listing “ill-advised” and rules one out before 2027.
So Anthropic will go first and write the benchmark. And if Anthropic’s IPO valuation holds, OpenAI’s planned $30 billion private round will be priced against a comparable it didn’t choose; if it breaks, every private valuation in the sector will be marked down by proxy.
One quiet feature: about 20% of the offering will be allocated to retail, against the 5% to 10% typical of big IPOs. SpaceX’s June IPO, with only 4% of shares floated, had a price that was effectively set rather than market-determined.
Anthropic’s offering, priced through an open order book, would be a fairer design, but it would also bring in retail investors who have not read the 94-page prospectus, and the risk that comes with them.
One wild card: Nvidia, Anthropic’s cornerstone and critical supplier, reports earnings on Tuesday, just 48 hours before ANT starts trading on open markets. It will be the chip maker’s first announced results since Jensen Huang’s ouster amid scandal drove Nvidia’s shares down 13% in a week. A weak Nvidia print could be Anthropic’s first market event.
The prospectus leaves out one key number: how concentrated Anthropic’s supply chain is. Most of its commitments are agreements to rent or buy computing power, and that computing depends on hardware.
Every advanced chip Anthropic uses is made by TSMC, which holds 72.5% of the foundry market, according to TrendForce, and is packaged on TSMC’s CoWoS lines. Those chips sit alongside high-bandwidth memory, where two Korean companies, SK Hynix and Samsung, supplied 83% of revenue last quarter, according to Counterpoint.
If a third to a half of Anthropic’s $895.4 billion is in hardware, and roughly two-thirds of an AI chip’s value is Taiwanese fabrication and Korean memory, then $200 to $300 billion of this American IPO is, in the end, an order book for Asia. That is my arithmetic, not the company’s disclosure, and each line item is open to debate.
What cannot be argued with is the structure beneath it: no route to Anthropic’s miracles bypasses Hsinchu and the Korean memory belt. Yet not one share will list in Asia. The continent that builds the boom’s physical layer will buy its financial layer on the Nasdaq, in dollars, at a price fixed on a Wednesday evening in Manhattan.
Wednesday’s pricing will show what institutions will pay; Thursday’s open will demonstrate what everyone else believes. In September, Amodei asked his industry to slow down; this week, the market will tell him what slowing down costs.
The chips will be made in Hsinchu, whether the miracles arrive or not. The warnings were written in San Francisco. The price is set in Manhattan. And on Thursday morning, for the first time in the short and feverish history of artificial intelligence, faith will get a ticker symbol. Everything after that is market-made arithmetic.
Sidhhant Kapai is an Indian writer, journalist and admitted scholar to Columbia University. His subject is the gap between what institutions say and what their records show.
