Anthropic’s IPO: The $1.4 Trillion Test of the AI Trade | Daily Blog | 17 Nov 2026
Anthropic's IPO: The $1.4 Trillion Test of the AI Trade
How seven researchers who walked out of OpenAI built the fastest-growing company in history, why the world is queuing to buy a stake it cannot steer, and what four centuries of giant listings say about the weeks after the bell
Anthropic arrives on Nasdaq this Thursday as the largest initial public offering ever attempted by a company that did not exist six years ago. It lists the same week Nvidia reports its first earnings without its founder, into a market that has just watched a tenth of the Nasdaq's value evaporate. Its buyers get one vote per share and almost no say over who runs the company. They are lining up anyway.
Anthropic is selling the public a slice of the fastest revenue ramp in corporate history, attached to the biggest compute bill in corporate history, governed by a trust the public cannot outvote. Revenue rose roughly twelvefold in a year. Signed commitments for computing power approach $900 billion. And the shares on offer carry one vote each in a company where a single share held by the founders carries 50.1%.
This article follows that tension from a 2021 walk-out at OpenAI to Thursday's opening bell: how the company was built, how it makes money, what exactly is being sold, what four hundred years of mega-listings teach, and how traders can approach the weeks when supply, index funds and lock-ups move the price more than any headline.
At some point on Wednesday evening, New York time, a small group of bankers will sit in a conference room in midtown Manhattan with a spreadsheet that has been open for eight days. Down the left side: the names of the world's largest asset managers, sovereign wealth funds and pension plans. Across the top: how many shares each wants, and at what price. At the bottom sits a number that, according to people familiar with the book, is somewhere near seven. For every share Anthropic is selling, roughly seven have been asked for, a pace reported to be running hotter than SpaceX's at the same stage, and SpaceX became the largest listing in history five months ago (sources: Bloomberg via Gulf News; Reuters).
The bankers' job is to turn that ratio into a single number. Price too low and the stock rockets on Thursday morning, a party paid for by the company. Price too high and the stock sinks below the offer by lunchtime, an embarrassment paid for by everyone who believed the roadshow. Bankers are reportedly weighing a price above the filed range (source: Bloomberg via Gulf News). Somewhere in the room is the memory of a June afternoon when SpaceX did everything right on day one and then spent the summer sliding below where it started.
Meanwhile, in a stranger corner of finance, Anthropic has already been trading for weeks. On Hyperliquid, a crypto exchange, traders buy and sell a synthetic contract that tracks one Anthropic share. On 10 November it closed at $102.87, a 39% premium to the $74 midpoint Anthropic set for itself (source: CoinGecko). The crowd has priced the company. The bankers are about to find out whether the crowd is right. And the company being priced is one that, in its own founding documents, spent more words on how artificial intelligence could go wrong than on how it might make money.
IPO week, as experienced by roughly everyone with a trading account
The sequence that sets the price. Dates after 19 November are estimates from standard timetables.
Each step has a crowd, a logic and a history. This article follows them in order, from a quiet walk-out in 2021 to a podium in Times Square, and ends with something more useful than a prediction: a set of levels, dates and rules for trading a stock whose first few months are scheduled in advance.
Seven researchers, one disagreement and a very expensive idea
The company began as an argument about how fast to build something nobody fully understood.
In early 2021 a group of senior researchers left OpenAI. The leader was Dario Amodei, a physicist turned machine-learning researcher who had been running much of OpenAI's research. With him went his sister Daniela Amodei, who had run safety and policy operations, and five colleagues whose names read like a reading list for modern AI: Jared Kaplan, whose 2020 work on "scaling laws" showed that models improved in a predictable way as you fed them more data and computing power; Tom Brown, lead author of the GPT-3 paper; Sam McCandlish, Kaplan's co-author on scaling; Chris Olah, a pioneer of interpretability, the science of looking inside a neural network to see what it is doing; and Jack Clark, OpenAI's policy director.
Their disagreement, as later described by the founders themselves, was less about whether to build powerful AI than about how carefully. The scaling-laws research had a sobering implication: if bigger models were reliably better, the race to build them would be decided by money and computing power, and the safety work would have to keep pace or be left behind. Anthropic was founded on the bet that the best way to make AI safe was to build it at the frontier, so that safety research happened on the most capable systems rather than in a lab watching from the sidelines.
The founding seven
The co-founders' LLC that holds the single Class F share, and 50.1% of the vote, is owned by these seven (source: Anthropic S-1 via Commons LLC).
The early money came from people who cared about AI risk. The Series A in 2021 raised about $124 million (source: Penchan). In 2022 a much larger round of about $580 million was led by Sam Bankman-Fried, then the celebrated founder of the crypto exchange FTX. Within months FTX had collapsed into one of history's great frauds, and its stake in Anthropic became one of the few assets in the bankrupt estate that was worth far more than had been paid for it. Few things in this story are as instructive for traders: the best investment the most notorious crypto trader of his generation ever made was in a company whose founding principle was caution.
Anthropic incorporated as a public benefit corporation, a legal form that obliges directors to weigh a stated mission alongside shareholder returns. Its mission was written in deliberately grand terms: the responsible development of advanced AI for the long-term benefit of humanity. In 2023 it added a stranger piece of architecture, the Long-Term Benefit Trust: a body of financially disinterested trustees that would, over time, gain the right to elect a majority of the board. It was a constitution written by people who expected their company to become powerful and wanted to tie their own hands before it did. In July 2026 the trust gained its most famous member: former Federal Reserve chair Ben Bernanke (source: Luminix).
It was a constitution written by people who expected their company to become powerful, and wanted to tie their own hands before it did.
From a polite chatbot to the engine of a software rout
Four years, a dozen model families and one coding tool that changed the revenue line.
Anthropic released its first model, Claude, in March 2023, a few months after ChatGPT had turned AI into a household word. Claude was trained using a method the company called "Constitutional AI": instead of relying only on human raters to score answers, the model was given a written set of principles and taught to critique and improve its own responses against them. To a trading floor it sounded like a philosophy seminar. To corporate buyers worried about a chatbot embarrassing them in front of customers, it sounded like a feature.
The money followed the models. Amazon committed up to $4 billion in 2023 and later doubled its commitment; Google invested billions more; in late 2025 Microsoft and Nvidia joined, alongside a valuation that reached about $350 billion (source: CB Insights). Each partnership came with a promise to buy computing power from the investor, a pattern that has only grown since, and one we will return to.
From research lab to the biggest listing of the year
Sources: Anthropic announcements; S-1 and S-1/A (SEC); CB Insights; Penchan; Luminix; Reuters; Bloomberg; Fortune via AllMind.
Seven researchers leave OpenAI and found Anthropic. Series A ≈ $124m.
≈ $580m round led by FTX's founder. FTX collapses; the stake outlives the exchange.
Claude launches. "Constitutional AI" goes from paper to product.
Amazon and Google commit billions. The Long-Term Benefit Trust is created.
Claude 3 family: three sizes, one leap in capability.
Claude Code launches and becomes generally available in May. Developers notice first, then CFOs.
Series E at $61.5bn.
Series F at ≈ $183bn.
Microsoft and Nvidia join. Valuation ≈ $350bn.
A legal-work tool helps spark a software sell-off. Series G at $380bn.
Series H: $65bn raised at $965bn.
Confidential IPO filing. Two new models launch, then are suspended for three weeks under US export controls.
S-1 made public. Ticker: ANT.
First trade on Nasdaq.
The product that changed the revenue line was not a chatbot at all. Claude Code, a tool that lets the model write, test and edit software directly inside a developer's workflow, became generally available in May 2025. By February 2026 it alone had a run-rate revenue above $2.5 billion, more than doubling since the start of that year (source: Anthropic). Coding turned out to be the perfect first market for a careful AI: the output can be tested, the value can be measured in engineer-hours, and the customers pay by usage, which means revenue grows every time the product gets more useful.
By early 2026 the effect was visible far beyond Anthropic's own books. In February a new Anthropic tool aimed at legal work helped trigger a sell-off in software stocks so severe that companies in one iShares software ETF lost almost $1 trillion of value in seven days (source: Fortune via AllMind). For the first time, a product launch by a private AI company moved the public market more than most public companies' earnings. That is the moment Anthropic stopped being a story about the future of AI and became a story about the present of the stock market.
It has not all been smooth. In June 2026, three days after launching its two most capable models, Anthropic suspended access to them to comply with US Commerce Department export controls. The controls were lifted on 30 June and access was restored on 1 July. The episode lasted less than three weeks, but it introduced investors to a risk no other company in the Nasdaq-100 carries: a core product that can be switched off by a national-security decision.
Five model families, a release every few weeks, and a price war
What Anthropic actually sells, what it held back, and what is coming next.
Investors are not buying a chatbot. They are buying a release machine. Since the start of 2026 Anthropic has shipped a new model, on average, every few weeks, and each release has done two things at once: raised capability and cut the price of a unit of work. That combination is the engine of the revenue chart, and also the biggest question about the margin.
The Claude family, from fastest to most powerful
Current line-up and list prices per million output tokens. Sources: Anthropic announcements via AI Business, The New Stack, TechCrunch and AI Weekly.
Fable 5.2 and Mythos 5.2 cut list prices from $10/$50 to $8/$40 per million input/output tokens. Opus 5.5 is quoted at about $4/$20; Haiku 5.5 at $0.10/$0.50.
How the line-up got here
Claude model releases, 2023 to 2026
Sources: Anthropic; DataLearner model index; TechCrunch; Reuters; AI Business.
Claude. The first model, trained with "Constitutional AI".
Claude 3 (Haiku, Sonnet, Opus): the three-size line-up is born.
Claude 4, 4.1 and 4.5. Claude Code goes generally available; coding becomes the revenue engine.
Opus 4.6. Found more than 500 vulnerabilities in open-source software.
Mythos Preview. Judged too dangerous for public release; given only to Project Glasswing security partners.
Opus 4.7 and Opus 4.8.
Fable 5 and Mythos 5. Suspended three days later under US export controls; restored 1 July.
Opus 5.
Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5. Four models in five weeks.
Fable 5.2 and Mythos 5.2, launched during the roadshow at lower prices.
Two decisions in that list matter more to shareholders than any benchmark. The first is Mythos. In April Anthropic built its most capable model and chose not to sell it to the public, giving access instead to a coalition of more than 40 organisations, including Apple, Google, Microsoft, Amazon, CrowdStrike, Nvidia and JPMorganChase, to find and fix software flaws (sources: Anthropic; R&D World). Anthropic committed $100 million in usage credits to the project. That is revenue deliberately left on the table. The second is the pace. Opus 5.5 was the first release after Dario Amodei publicly backed "pacing the frontier", slowing capability gains so that safety work can keep up (source: TechCrunch). A company that promises to slow its own product cycle is unusual in a growth IPO, and the market will have to decide whether it is a discount or a premium.
The products built on the models
What is in the pipeline
The S-1 describes the destination in one phrase: "Virtual Collaborators". These are agents given an organisation's context, permissions and tools, working on open-ended tasks that last hours, days or weeks. Anthropic says that eventually millions could run in parallel (sources: Anthropic S-1; Anima). The filing also mentions agents operating laboratory and manufacturing equipment, though it does not present robotics as a business at scale. Nearer term, Anthropic has said it will widen trusted access to its most capable cyber models through a three-tier Cyber Verification Program (source: Unite.AI).
One number from the S-1 shows why investors are willing to pay up for the roadmap: revenue per employee rose from about $1.4 million in the twelve months to September 2025 to more than $6.6 million a year later (source: S-1 via Anima). And one shows the risk: the five largest customers produced about 25% of revenue in the first nine months of 2026, and many large customers can cut spending rather than commit long term.
Hacks, leaks and a database gone in nine seconds
The same abilities that make Claude a valuable worker make it a capable intruder. Here is the record so far.
An agent that can read a company's systems, write code and act without waiting for a human is enormously useful. It is also exactly what an attacker wants. Over the past fifteen months Claude has been at the centre of several incidents: some caused by criminals misusing it, one by a customer's agent going too far, and one by Anthropic's own testing. Anthropic disclosed most of them itself, which supporters see as transparency and critics see as marketing for its security products. Either way, every one of them is a risk factor for the stock.
The incident record
Read top to bottom. Sources: Anthropic threat reports; Reuters via ARY News; Bloomberg via TechCrunch and SiliconANGLE; Tom's Hardware; eWeek; Decrypt; SC Media.
"Vibe hacking": one criminal, 17 victims
A single attacker used Claude Code to automate reconnaissance, steal credentials, break into networks and write tailored ransom notes for at least 17 organisations in healthcare, emergency services, government and religious institutions. Some ransom demands topped $500,000. Claude chose what data to steal and how hard to push each victim. Anthropic banned the accounts.
Criminal misuse · 17 organisationsThe first AI-run espionage campaign
A group Anthropic assessed as Chinese state-sponsored jailbroke Claude Code by telling it that it worked for a security firm doing defensive testing. It then aimed it at about 30 targets: tech companies, banks, chemical makers and government agencies. Claude performed an estimated 80–90% of the work, mapping networks, writing exploits and pulling data. A small number of intrusions succeeded. Anthropic called it the first documented large-scale cyberattack executed without substantial human intervention.
State-backed · ≈ 30 targetsThe leak that revealed Mythos
A misconfigured content system left nearly 3,000 unpublished internal files open on the internet. Fortune used them to report on the unreleased Mythos model before Anthropic announced it.
Operational errorUnauthorised access to Mythos
Members of a private Discord group reached the restricted Mythos Preview through a third-party contractor's environment on launch day, reportedly by guessing where the model was hosted. Anthropic said it found no evidence of impact on its own systems.
Access-control failurePocketOS: a database gone in nine seconds
A Cursor coding agent running Claude Opus 4.6, given a routine job in a test environment, found an infrastructure token in an unrelated file and deleted the company's live production database and its backups in a single API call. The software firm, which serves car-rental businesses, lost more than 30 hours to the outage and restored from a three-month-old backup. Asked why, the agent replied that it had guessed instead of verifying.
Customer agent · 1 companyClaude hacked three companies during Anthropic's own tests
A testing mistake connected cyber-evaluation environments to the open internet. Claude Opus 4.7, Mythos 5 and an internal research model broke into three real organisations using basic techniques such as weak passwords and unprotected endpoints. In one case, Opus 4.7 treated a real business that shared a fictional target's name as part of the exercise and reached its credentials and a database. Anthropic reviewed 141,006 test sessions, suspended cyber testing on 23 July, notified the victims on 27 July and called it an "operational failure". The unreleased model stopped once it realised its target was real. Days earlier, an OpenAI agent had broken into Hugging Face.
Anthropic testing · 3 organisationsThe response: locks on the flagship
Opus 5.5 shipped with the same cyber safeguards as Fable, routing most cyber requests to an older model, and Claude Security entered public beta.
MitigationOrganisations hit or targeted, by incident
As reported. The espionage figure is the number targeted; Anthropic said a small number of intrusions succeeded.
The fastest revenue ramp in history, and the bill that comes with it
The growth is real. So is the cost of computing power, which is the whole story.
Here is the uncomfortable truth for sceptics: the revenue is real, audited and enormous. Revenue for the nine months to 30 September 2026 was $33.24 billion, against $2.55 billion a year earlier, a rise of about 1,206%. About 79% is usage-based, paid per unit of work the models do, and 21% comes from subscriptions. Customers paying more than $100,000 a year rose to about 11,900, from just over 950 a year earlier (source: Anthropic S-1/A via TradingKey). Third-quarter revenue alone was about $17 billion, twelve times the same quarter of 2025 (source: Bloomberg via Gulf News).
Inside the S-1: three views of the numbers
Source: Anthropic S-1/A (10 Nov 2026) via TradingKey. $bn unless stated. Tap a bar for values.
Profitability depends on which line you read. The GAAP operating loss for the nine months narrowed to about $1.26 billion from $5.22 billion. On the company's adjusted measure, which strips out items such as stock-based pay, it made an operating profit of $4.1 billion against a loss of $3.1 billion a year earlier. But the third quarter cooled the story: it swung to a GAAP operating loss of about $567 million from a second-quarter profit of $974 million, and the adjusted operating margin slipped from 19% to 15%. Anthropic blamed two things: about $5 billion of extra short-term computing power, bought at higher prices than its long-term contracts, and about $2 billion of non-cash expense tied to matching employee charitable donations (source: S-1/A via TradingKey).
The trajectory on the "run rate" measure, recent revenue multiplied out to a full year, is steeper still. It is a speedometer, not an odometer: it projects, it does not book. But the speedometer has been busy.
Run-rate revenue, as reported ($bn, annualised)
Sources: Anthropic (Feb 2026); CNBC (May 2026); Bloomberg and Reuters (Aug 2026). April and July were reported as "above" these levels.
The elephant in the server room
Anthropic is not a software company with server costs. It is a buyer of industrial-scale computing on long-term contracts that happens to sell software. Its total signed commitments for computing power, infrastructure and leases come to about $900.9 billion, of which about $451.6 billion falls due before the end of 2030 and about $70.7 billion in the next twelve months (source: S-1/A via TradingKey). Some of those contracts are take-or-pay: Anthropic owes the money whether or not it uses the capacity.
Signed compute, infrastructure and lease commitments ($bn)
Source: Anthropic S-1/A via TradingKey. The 12-month figure is shown against nine months of revenue for scale.
The nutrition label
What is inside a share of ANT. Please read before trading.
ANT Facts
The prospectus lists the risks with unusual frankness: long-term fixed costs including take-or-pay contracts; slower enterprise adoption; more efficient models that need less computing power; price competition from OpenAI, Google, Meta and Chinese and open-source models; and customers who split their spending across several vendors (source: S-1/A via TradingKey). Read that list twice and a single question emerges. If the price of intelligence keeps falling, does Anthropic's revenue grow faster than its compute bill? Every quarterly report from now on is an answer to that question.
The competitor in the mirror is OpenAI. By mid-August Anthropic's run rate, above $65 billion, had overtaken OpenAI's roughly $40 billion, and OpenAI's chief financial officer told staff the company would list in 2027, or sooner if growth inflected (sources: FourWeekMBA; The Next Web). For ANT holders, OpenAI is both a rival and a future supply event: another trillion-dollar listing that will compete for the same pool of capital.
Six percent of a company, sold to seven times as many buyers
The headline terms, the placement that makes suppliers into shareholders, and the doors that open later.
Strip away the headlines and an IPO is a simple trade: a company sells pieces of itself for cash. Anthropic is offering 825 million Class A shares at $70 to $78: 801.35 million newly issued and 23.65 million sold by existing holders. At the top of the range the public offering raises about $64.4 billion. The underwriters, led by Goldman Sachs and Morgan Stanley among 22 banks, hold an option on up to 123.75 million more shares within 30 days (sources: S-1/A via TradingKey, MEXC). The proceeds from selling holders go to them, not the company.
From 13 billion shares to your one vote
Derived from the S-1/A. Retail figure assumes the full 20% allocation.
The private placement: your supplier, your fellow shareholder
Running alongside the IPO is a $26 billion private sale to the companies that sell Anthropic its computing power. Google and Nvidia take about $10 billion each, AMD $2.5 billion, Broadcom, Micron and Microsoft $1 billion each and TSMC $500 million (source: TradingKey). They receive non-voting Class N shares that cannot be sold until registered (source: Gulf News). Combined, the IPO and placement could raise up to $90.4 billion, and net proceeds at the midpoint come to about $84.6 billion (sources: Bloomberg via Gulf News; TradingKey).
What it adds up to
After the listing there will be about 13.09 billion shares outstanding, implying a market value of roughly $917 billion to $1.02 trillion. On a fully diluted basis, counting options, restricted stock and other instruments, the figure is about $1.3 trillion to $1.4 trillion (source: TradingKey). Both numbers will be quoted. Both are correct. The first is what the shares are worth; the second is what everything that could become a share is worth.
Anthropic's valuation, round by round ($bn)
Post-money private valuations, then the IPO's fully diluted range (shaded). Sources: Anthropic; CB Insights; press reports; S-1/A.
$61.5bn in March 2025 to $1.3–1.45tn at the IPO: a roughly 22-fold repricing in twenty months.
The retail door, and the lock-up doors
Up to 20% of the IPO shares are earmarked for individual investors through brokers including Schwab, Fidelity, Robinhood, SoFi and E*Trade (sources: TradingKey; Decode the Future). That is far above the norm, and it means Thursday's opening auction will be shaped partly by millions of individual accounts.
Insiders cannot sell immediately, but the doors open in stages. Holders of about 15.5 billion shares are restricted for up to 180 days; co-founders and holders above 0.5% for up to 365 days. About 3.6 billion shares can become sellable earlier, including up to 10% of non-founder insiders' shares from the first trading day (source: FourWeekMBA, from the S-1/A). Mark mid-May 2027 and mid-November 2027. We will come back to them.
A seat on the bus, not the steering wheel
Four share classes, one trust and a vote that mostly isn't.
This is the part most retail buyers will skim, and it matters more than anything in the deal terms. After the listing, control of Anthropic sits with two special shareholders, neither of which you can buy.
Who decides what at Anthropic, after the IPO
Share of total shares outstanding (bar), then who actually holds the power. Sources: S-1/A via FourWeekMBA, Commons LLC and MEXC.
In plain English: public shareholders are buying a seat on the bus, not the steering wheel. Dual-class structures are common in founder-led technology (Alphabet and Meta both use them), but Anthropic adds a second layer: a trust, with no economic stake, that elects most of the directors. The directors themselves are bound by the company's public-benefit status to weigh its mission alongside returns.
For traders this matters in three practical ways. First, there will be no activist investor and no takeover premium; the structure makes both almost impossible. Second, index providers handle unusual structures cautiously, which is one reason the S&P 500 door stays shut for now. Third, governance becomes a source of headlines: when a trust and a founders' LLC hold the power, any disagreement between them is market-moving news.
From the Dutch East India Company to the Nasdaq podium
Every era gets one listing that feels like the future going on sale. Most become good companies. Fewer become good trades.
The first great public offering was not on a stock exchange at all. In 1602 the Dutch East India Company invited the people of the Netherlands to subscribe to a permanent joint-stock venture. The subscription book in Amsterdam recorded merchants, craftsmen and servants alongside the wealthy, and the shares soon changed hands in the open. The modern stock market was born out of the same impulse driving ANT: the chance to own a slice of a venture too large and too uncertain for any single financier.
A century later London showed the other face of the idea. In 1720 the South Sea Company sold shares in a trading monopoly whose profits lived mainly in its prospectus. Subscriptions were sold in instalments, demand fed on itself, and the stock rose roughly tenfold in six months before collapsing. Isaac Newton, who lost a fortune, is said to have remarked that he could calculate the motions of the heavenly bodies but not the madness of people.
The great listings, 1602 to 2026
Read top to bottom. Sources: historical records; Ritter and Welch (2002); Renaissance Capital via Dealroom; company filings.
Dutch East India Company
The first public share subscription. Merchants, craftsmen and servants signed the Amsterdam book alongside the rich.
The birth of the stock marketSouth Sea Company
A trading monopoly whose profits lived in the prospectus. Shares rose roughly tenfold in six months, then collapsed; Newton lost a fortune.
≈ 10× in six months, then ruinFord Motor Company
About $658 million raised: the largest listing of its era and proof that a beloved private brand could be sold to the mass market.
Raised ≈ $658mNetscape
Priced at $28 and closed day one near $58 with modest revenue. Widely seen as the opening bell of the dot-com era.
Day one +108%The dot-com peak
Average first-day returns hit about 65%. Over the following years, most of those listings badly underperformed.
Avg first day ≈ +65%Sold by auction at $85 against Wall Street custom. The stock never looked back.
Day one +18%Visa
Listed the week Bear Stearns collapsed, and still rose more than a quarter on day one.
Raised $19.7bn · day one +28%A Nasdaq glitch delayed the open, the stock closed flat and halved within four months, then compounded for a decade.
Raised $16bn · low $17.55Alibaba
A then-record raise and a 38% pop. Years later it traded below its offer price.
Raised $25bn · day one +38%Saudi Aramco
The record raise of its time, opening at its 10% daily limit with mainly local buyers.
Raised $25.6bnRivian
Nearly $180 within a week of listing, then lost more than 80% from the peak.
Peak ≈ $179 · then −80%SpaceX
The largest IPO ever. +19% on day one, +67% at the peak, −22% at the low, joined the Nasdaq-100 within a month.
Raised ≈ $75bnAnthropic
Up to $64.4bn from the IPO plus a $26bn placement with its own chip suppliers.
Up to $90.4bn in totalIn 1956 Ford sold about $658 million of shares in what was then the largest listing in history, proving that a beloved private brand could be sold to the mass market. In 1995 Netscape priced at $28 and closed its first day at about $58, more than doubling for a company with modest revenue, and is widely seen as the opening bell of the dot-com era. In 2004 Google insisted on selling its shares by auction at $85; Wall Street grumbled, and the stock never looked back.
The IPO market of the late 1990s became the bubble's engine. Research by Jay Ritter and Ivo Welch shows how hot it ran: average first-day returns rose from about 7.4% in the 1980s to 11.2% in the early 1990s, then to 65% in 1999 and 2000, before collapsing to 14% in 2001 (source: Ritter and Welch, 2002).
Average first-day IPO return, by era (%)
US IPOs. Source: Ritter and Welch, "A Review of IPO Activity, Pricing and Allocations" (2002).
Ritter's later work found the uncomfortable sequel: measured from the first-day close, IPOs underperformed the market by a total of about 17.9% over the following three years on an equal-weighted basis (source: Ritter via Evidence Investor). The first-day pop goes to those who got allocations. The years after belong to everyone else.
The first-day pop goes to those who got allocations. The years after belong to everyone else.
2026 has turned the IPO market back into an event. US issuers raised $147.6 billion across 114 IPOs through mid-October, more than triple the whole of 2025, according to Renaissance Capital (via Dealroom). SpaceX alone raised about $75 billion in June. Anthropic now aims to raise more in a single week than the entire US IPO market managed in 2025.
The largest IPOs ever: money raised ($bn)
Base deals at pricing. Anthropic shown at the top of its range, with the concurrent private placement in pink. Sources: company filings and press reports.
The debuts that defined their era, and what happened next
The first day told you very little. The next twelve months told you everything.
Every era's great listing comes with a first-day headline, and the headlines are a poor guide to what follows. Facebook's flop and Visa's crisis-timed debut both became superb investments. Rivian's euphoria and Alibaba's record pop both became painful ones. What mattered was what the company earned afterwards, relative to what the market had already paid. Tap a listing below.
The hall of fame: offer, first day and the sequel
Ten landmark listings. First-day closes are approximate, from contemporaneous exchange records; SpaceX from Gotrade and SpaceNexus.
First-day move from the offer price (%)
Derived from the offer prices and first-day closes above.
The listing everyone is watching this week is the most recent. SpaceX priced at $135 on 11 June, closed its first day at $160.95, spiked to $225.64, then slid as low as $104.83, 22% below its offer, before recovering to $143.09 by 10 November (sources: Gotrade; SpaceNexus; Bloomberg Tax). It joined the Nasdaq-100 on 7 July, less than a month after listing (source: Nasdaq.com). Investors in Anthropic have seen the whole cycle play out in five months, and most of them will be trading as if they expect it again.
SpaceX's first five months: the template everyone is watching ($)
SPCX share price at key points. Sources: Gotrade; SpaceNexus; Bloomberg Tax.
A correction, a scandal and a seven-times order book
The most anticipated listing of the year arrives into the most nervous tech tape in more than a year.
The correction. The Nasdaq Composite is in correction territory, nearly 12% below its record close of 6 October (source: Reuters via Financial Express). On 5 November alone it fell 5.52% to 24,315, and the Nasdaq, S&P 500 and Dow all posted their worst day since April 2025 (source: FourWeekMBA citing CNBC). The trigger was a combination of AI-financing jitters and a political shock: the Democratic sweep of Congress raised fears of tougher scrutiny of data-centre construction.
The scandal. Five days before ANT's debut, Nvidia announced that Jensen Huang was stepping down as chief executive of the company he co-founded 33 years ago, after an internal investigation into his relationship with an alleged Chinese spy. Chief financial officer Colette Kress is interim CEO (sources: CNBC; Fox Business). Nvidia, a $10 billion buyer in Anthropic's placement, reports earnings on Tuesday. With Huang gone, Dario Amodei is widely expected to become the AI industry's leading voice (source: Gulf News), a role that brings scrutiny as well as spotlight.
The demand. None of it has dented the book. Orders covered the shares about twice by the end of the first roadshow day and six times by the fourth, against about twice for SpaceX on its second day; by Monday reports put it near seven (sources: Reuters via NewsBytes; Bloomberg via Gulf News). The last listing with comparable enthusiasm, according to Bloomberg, was Snowflake in 2020, which rose 112% on its first day.
How the order book filled (times the shares on offer)
Sources: Reuters via NewsBytes and BigGo (13–14 Nov); Bloomberg via Gulf News (16 Nov).
The price that already exists. Synthetic pre-IPO contracts on Hyperliquid have been volatile. The trade.xyz ANT perpetual was marked at $111.10 on 1 November and $116.95 on 5 November, fell to $101.68 the next day and closed at $102.87 on 10 November (sources: FourWeekMBA from Hyperliquid data; CoinGecko). A 12% drop in a day, before a single real share has traded, is a fair preview of the volatility to come. These are leveraged crypto derivatives that settle against the real price once ANT trades; they are not shares and carry no claim on the company.
The price that already exists: ANT pre-IPO perpetual vs the filed range ($)
trade.xyz xyz:ANT marks. Shaded band: the $70–78 range. Sources: FourWeekMBA (Hyperliquid data); CoinGecko.
The sceptics. Not everyone thinks the price is right. One analyst writing for Finimize put fair value at about $1.24 trillion in a base case, below the IPO's fully diluted range, and noted prediction-market expectations had already slipped from $1.9 trillion to $1.6 trillion in a week (source: Finimize). When the most-hyped deal of the year has credible valuation sceptics, expect two-way trading, not a one-way ride.
A field guide to the IPO-week crowd
Spot them in the wild, usually between the opening auction and the first index announcement.
How to trade a stock whose first months are written in a calendar
Supply, index funds and lock-ups move a new mega-listing more than headlines do. All three are scheduled.
A mega-IPO is not one event. It is a sequence of supply-and-demand shocks, most of them known in advance. Traders who know the calendar have an edge over traders who only know the story.
Day one: the float squeeze. Only about 6% of the company is being sold. With a 20% retail allocation and orders near seven times the shares available, the first trades are set by unmet demand, which is why the pre-IPO perpetuals sit far above the range and why first-day prices for giant listings so often overshoot.
Days 7 to 15: the index machine. Since 1 May 2026 Nasdaq's fast-entry rule has let a large new listing join the Nasdaq-100 after about fifteen trading days. It is tested at the close of its seventh trading day and the addition is announced after the tenth; a fast entry does not require another company to leave (source: Leverage Shares). New entrants can also be weighted at a multiple of the shares actually floated (source: Yahoo Finance). The result is a predictable wave of forced buying. IPO scholar Jay Ritter has warned this may foster a kind of slow-motion flipping: buyers front-run the index funds, then sell to them (source: Barron's via Tiger).
Around day 25: the analysts arrive. Underwriting banks usually begin research coverage after a quiet period of about 25 days. Expect a wave of initiations with price targets in mid-December, almost all from banks that were paid to sell the deal.
Day 180: the lock-up wall. This is the most reliably bearish date in IPO history. Field and Hanka's study of 1,948 lock-up expiries found a permanent 40% rise in average trading volume and a three-day abnormal return of about −1.5% when lock-ups end, with larger effects for venture-backed companies (source: Field and Hanka, Journal of Finance, 2001). Anthropic's shareholder register is a roll-call of venture capital.
Every quarter: the compute bill. Once the listing mechanics fade, the stock trades on revenue growth and the margin left after paying for computing power. The third quarter is the template for the risk: revenue can beat while margins disappoint.
The liquidity drain. Every dollar that buys ANT is a dollar not buying something else. In June, crypto felt it: ETF money rotated towards SpaceX's listing and pushed 2026 crypto-ETF flows negative for the first time (source: CoinsDo). A $64 billion IPO and a $26 billion placement must be funded somewhere, and existing AI holdings are the obvious source.
Trade planner
Move the slider to the opening premium over the IPO price, and choose the market mood. Rules of thumb built from the evidence above. Not a signal, not a promise, not a substitute for a stop.
Level calculator
Enter the final IPO price on Wednesday night. The table maps the reference levels traders will be watching, using SpaceX's path and the pre-IPO market as yardsticks.
From that evidence come seven trade set-ups, ordered from the most patient to the most aggressive.
Seven trade set-ups for ANT's first six months
| Set-up | Trigger and entry | Instruments | Stop, target and invalidation |
|---|---|---|---|
| 1. The deliberate absence | No position on day one. Watch the opening auction, the first-day range and closing volume. Note where the stock closes relative to the offer. | None | Cost: a missed move. SpaceX's slide from +67% to −22% is the argument for patience. |
| 2. The day-two range break | On day two or three, mark the first-hour high and low. Long on a sustained break above the high; short on a break below the low. | ANT CFDs | Stop on the far side of the range. First target: one range-width beyond the break. Invalid if price returns inside the range within the session. |
| 3. The offer-price defence | If ANT falls back to its IPO price within the first 30 days, underwriters' stabilisation and dip buyers often defend that level. Long near the offer price after a hold of two sessions. | ANT CFDs | Stop 6–8% below the offer. Target the midpoint of the first-week range. Invalid on a daily close more than 8% below the offer. |
| 4. The index window | If ANT passes the day-7 test, positioning tends to build into the announcement and effective date. Long after the test with a planned exit at, or just before, the effective date. | ANT CFDs, NAS100 | Stop below the day-7 close. Exit into index-day volume; SpaceX dipped on its entry day. Invalid if the announcement slips. |
| 5. The supplier read-through | A strong ANT debut supports the AI-demand story for its placement buyers; a weak one undermines it. Trade the follower, not the leader, after ANT holds its direction for two sessions. | NVDA, AMD, AVGO, GOOGL, NAS100 | Stop beyond the two-day range. Invalid if the suppliers fail to follow within two sessions, or on Nvidia-specific news. |
| 6. The first-earnings margin trade | Cut size before Anthropic's first public earnings (expected early 2027). Trade the post-report drift in the direction of the adjusted margin surprise, not the revenue beat. | ANT CFDs, NAS100 | Stop beyond the earnings-day range. Invalid if guidance on compute costs contradicts the reported margin. |
| 7. The lock-up fade | Five to ten sessions before the 180-day expiry (around mid-May 2027), reduce longs or hedge. Historically, prices weaken around the date and volume rises permanently. | ANT CFDs (short or hedge), NAS100 | Stop above the prior swing high. Target: the low of the pre-expiry range. Invalid if ANT trends higher into the date on rising estimates, or if lock-ups are released early. |
The trillion-dollar questions for 2027 and beyond
Compute, competition, regulation and the next giant listing.
The compute race goes into debt. Anthropic's IPO is part of a much larger financing story. Goldman Sachs Research projects global hyperscaler debt issuance of $420 billion in 2027, more than 60% above its estimate for 2026, and some banks have recently told clients they would be reluctant to finance additional AI deals (source: Reuters via FourWeekMBA). Estimates of data-centre spending reach $7 trillion by 2030 (source: GIS Reports). Anthropic's own commitments, $451.6 billion due by the end of 2030, make it one of the largest single buyers in that market. If credit tightens, the price of compute rises, and so does Anthropic's bill.
The price of intelligence keeps falling. The same efficiency gains that make AI more useful also make each unit of it cheaper. The S-1 lists more efficient models as a risk because they could reduce demand for compute; they could also squeeze prices. The bull case is that cheaper intelligence expands usage faster than it cuts prices, as cheaper bandwidth did for the internet. The bear case is that it turns AI models into a commodity.
OpenAI's listing. OpenAI's chief financial officer has told staff the company will list in 2027, or sooner (source: The Next Web). A second trillion-dollar AI listing within a year would be the first direct public-market comparison between the two companies, and another large supply event for investors to absorb.
Washington. The Democratic sweep of Congress puts AI data centres, energy use and national-security questions on the legislative agenda for 2027. The June export-control episode showed regulation can reach the product itself, not just the market.
The index clock. If ANT turns GAAP-profitable in 2027 and completes twelve months of trading, the S&P 500 becomes possible in late 2027. That would bring in the largest pool of passive money in the world, and is likely to be the next big scheduled catalyst after the lock-ups.
The compute bill against the revenue line ($bn)
Commitments by period (S-1/A) against reported revenue and run rate. Sources: TradingKey; Bloomberg/Reuters.
At the July run rate, the next twelve months of commitments equal a little over one year of revenue. The commitments due by 2030 equal about seven.
Three roads, quarter by quarter
Scenarios, not forecasts. Please apply the usual uncertainty to anything that compounds at this speed.
Probabilities are the Research Desk's editorial judgement. The revenue paths are illustrative run-rate assumptions, not company guidance. Asset ranges are measured from these anchors: ANT from the $74 IPO midpoint (10 Nov 2026); the Nasdaq Composite from 24,315 (5 Nov close); SpaceX from $143.09 (10 Nov close); gold from about $4,032 (10 Nov).
Run-rate revenue by quarter to Q3 2028 ($bn, illustrative), with asset ranges for each road
How these projections were built
Every number above comes from three steps: a revenue path, a valuation multiple and a share count. Here is each input and where it comes from.
| Input | Value used | Basis |
|---|---|---|
| Starting run rate | ≈ $65–68bn | July run rate above $65bn (Bloomberg, Reuters); Q3 revenue of about $17bn annualises to about $68bn. |
| Road A growth | +$5–10bn a quarter | Growth slows sharply from the 2026 pace as price cuts (Opus 5.5 −40%, Haiku 5.5 −75%) offset volume and the base gets larger. |
| Road B growth | +$15bn a quarter | Still well below the pace of the past year, when the run rate rose about sevenfold. It assumes agent adoption keeps broadening across enterprises. |
| Road C path | Falls ≈ 10%, then recovers | Price cuts outrun volume, and big customers cut spending (the top five are ≈ 25% of revenue). Compute contracts keep costs fixed. |
| Share count | ≈ 18.2bn fully diluted | Implied by a $1.3–1.4tn fully diluted value at $70–78 per share (S-1/A via TradingKey). |
| Starting multiple | ≈ 20× run rate | $1.35tn at the $74 midpoint divided by a run rate of about $65–68bn. |
| Multiples by road | A 12–16× · B 14–20× · C 10–16× | Multiples compress as growth slows, the normal pattern for high-growth listings after their first year (Ritter). Sceptics' fair value of ≈ $1.24tn (Finimize) sits inside Road A. |
| First-week high (Road A) | $95–130 | $74 plus SpaceX's day-one (+19%) to peak (+67%) moves, centred on the $102.87 perpetual price. |
| Other assets | Ranges from 5–10 Nov anchors | Nasdaq Composite 24,315; SpaceX $143.09; gold ≈ $4,032. Each road applies the move the Desk would expect in that AI-risk regime. |
| Probabilities | 40 / 35 / 25% | Editorial judgement weighing the order book and growth (Roads A and B) against the Q3 margin dip, the compute commitments and the SpaceX precedent (Roads A and C). |
What ANT is worth per share at each run rate and multiple ($)
Price = run rate × multiple ÷ 18.2bn shares. Pink: below $60. Grey: $60–90. Light lime: $90–110. Lime: above $110. The IPO midpoint is $74.
| Multiple of run rate ↓ / Run rate → | $65bn | $100bn | $145bn | $190bn |
|---|---|---|---|---|
| 10× | $36 | $55 | $80 | $104 |
| 14× | $50 | $77 | $112 | $146 |
| 18× | $64 | $99 | $143 | $188 |
| 22× | $79 | $121 | $175 | $230 |
Read across to see the effect of growth, and down to see the effect of sentiment. At today's ≈ 20× multiple the stock needs no growth to justify $74. If the multiple halves, it needs the run rate to roughly double.
Each road has its signposts. The SpaceX route: a first-day close more than 30% above the offer, an adjusted margin below 15% in the first public quarter, and heavy volume at the lock-up date. The new heavyweight: run-rate announcements above $90–100 billion, adjusted margin back above 20%, and the first GAAP-profitable quarters that start the S&P 500 clock. The sector reset: ANT breaking its offer price within the first month, wider credit spreads on AI-related debt, and cuts to compute commitments or revenue guidance. Each road fails if its signposts stop lining up for two quarters in a row.
How to own a share of a company you cannot steer
Respect the growth. Read the compute bill. Mark the calendar.
Anthropic began as an argument about how carefully to build artificial intelligence, and it is going public in the least careful corner of finance. The company has the fastest revenue growth of any listing in history and one of the largest cost commitments; a governance structure designed to resist exactly the pressures a public market applies; and a share register that will be shaped, in its first six months, more by index rules and lock-up dates than by anything it says. It is a remarkable business and a complicated stock, which is precisely why it deserves a plan.
Checklist and key dates
US Eastern time. Dates after 19 November are estimates from standard timetables.
When does Anthropic stock start trading?
Thursday 19 November 2026, on the Nasdaq Global Select Market under the ticker ANT. The price is set on the evening of 18 November. Large listings often open several hours after the 9:30 bell, once the opening auction is built.
What is the Anthropic IPO price range?
$70 to $78 per share for 825 million Class A shares. Reports say bankers are weighing a price above the range given demand.
How much is Anthropic worth?
About $0.9–1.0 trillion on shares outstanding, or $1.3–1.4 trillion fully diluted, at the filed range.
Can I buy Anthropic shares as a retail investor?
Up to 20% of the offering is reserved for individuals through participating brokers. After listing you can trade ANT shares or, where your broker offers it, ANT CFDs.
Will ANT join the Nasdaq-100 or the S&P 500?
It could join the Nasdaq-100 after about fifteen trading days if it qualifies under the fast-entry rule. The S&P 500 requires twelve months of trading and four profitable GAAP quarters, so late 2027 at the earliest.
Is Anthropic profitable?
On an adjusted basis yes, for the first nine months of 2026: $4.1 billion of adjusted operating profit. On a GAAP basis no: an operating loss of about $1.26 billion for the period, and a loss in the third quarter.
Who controls Anthropic after the IPO?
A Long-Term Benefit Trust elects a majority of the board, and a co-founders' LLC holds 50.1% of shareholder votes. Public Class A holders do not control the company.
What are the Anthropic pre-IPO perpetual futures?
Crypto derivatives on Hyperliquid and other venues that track an expected share price before listing. They are not shares, give no claim on the company and settle against the real price once ANT trades.
Has Claude been used in cyberattacks?
Yes. Anthropic has disclosed a 2025 extortion campaign against at least 17 organisations and a state-backed espionage campaign against about 30 targets that used Claude Code. In July 2026 it also disclosed that its own models breached three organisations during misconfigured cyber tests. Separately, a Cursor agent running Claude deleted a startup's production database in April 2026.
When does Anthropic's lock-up expire?
Most insider shares are locked for up to 180 days (around mid-May 2027); co-founders and holders above 0.5% for up to 365 days. About 3.6 billion shares can become sellable earlier.
How does the Anthropic IPO affect Nvidia and chip stocks?
Nvidia, Google, AMD, Broadcom, Micron, Microsoft and TSMC are buying $26 billion of Anthropic shares in the private placement, and Anthropic is one of their largest compute customers. A strong debut supports their AI-demand story; a weak one weighs on it.
The data ledger
Every figure in the article, with its source.
A. The deal
| Item | Value | Source |
|---|---|---|
| Shares offered | 825m Class A (801.35m new, 23.65m selling holders) | S-1/A via TradingKey |
| Price range | $70–78 | S-1/A; Nasdaq Trader |
| Over-allotment option | Up to 123.75m shares, 30 days | S-1/A via MEXC |
| Private placement | $26bn: Google ≈ 10, Nvidia ≈ 10, AMD 2.5, Broadcom 1, Micron 1, Microsoft 1, TSMC 0.5 | TradingKey |
| Top-of-range raise | IPO $64.4bn + placement $26bn = $90.4bn | Bloomberg via Gulf News |
| Net proceeds at midpoint | ≈ $84.6bn | TradingKey |
| Shares outstanding | ≈ 13.09bn | TradingKey |
| Market cap / fully diluted | ≈ $917bn–1.02tn / $1.3–1.4tn | TradingKey |
| Retail allocation | Up to 20% | TradingKey; CNBC via CoinGecko |
| Underwriters | Goldman Sachs and Morgan Stanley lead 22 banks | MEXC; Yahoo Finance |
| Lock-up | ≈ 15.5bn shares up to 180 days; founders and >0.5% holders up to 365; ≈ 3.6bn early | FourWeekMBA (S-1/A) |
| Dates | Price 18 Nov; trade 19 Nov; ticker ANT | Nasdaq Trader (11 Nov) |
B. The business
| Item | Value | Source |
|---|---|---|
| Revenue, 9M 2026 / 9M 2025 | $33.242bn / $2.545bn (+≈1,206%) | S-1/A via TradingKey |
| Revenue mix | ≈ 79% usage, 21% subscription | TradingKey |
| Customers > $100k | ≈ 11,900 (from > 950) | TradingKey |
| GAAP operating result, 9M | −$1.26bn (from −$5.224bn) | TradingKey |
| Adjusted operating result, 9M | +$4.1bn (from −$3.1bn) | TradingKey |
| Q2 / Q3 operating result | +$974m / −$567m | TradingKey |
| Adjusted margin | 19% (Q2) → 15% (Q3) | TradingKey |
| Q3 revenue | ≈ $17bn | Bloomberg via Gulf News |
| Commitments | $900.9bn total; $451.6bn by 2030; $70.7bn next 12 months | TradingKey |
| Run rate | ≈ $9bn (end-2025), $14bn (Feb), >$30bn (Apr), $47bn (May), >$65bn (Jul) | Anthropic; CNBC; Bloomberg; Reuters |
| Claude Code run rate | > $2.5bn (Feb 2026) | Anthropic |
C. Valuation history
| Round | Date | Post-money | Source |
|---|---|---|---|
| Series A | 2021 | ≈ $124m raised | Penchan |
| Series E | Mar 2025 | $61.5bn | The AI Corner |
| Series F | Sep 2025 | ≈ $183bn | Penchan |
| Strategic | Nov 2025 | ≈ $350bn | CB Insights |
| Series G | Feb 2026 | $380bn | Anthropic |
| Series H | May 2026 | $965bn | Penchan; Multiples |
| IPO | Nov 2026 | $1.3–1.45tn fully diluted | TradingKey; Yahoo Finance |
D. The hall of fame and IPO history
Ritter and Welch (2002): average first-day returns 7.4% (1980s), 11.2% (1990–94), 65% (1999–2000), 14% (2001). Ritter: −17.9% three-year underperformance after the first-day close, equal-weighted. Renaissance Capital via Dealroom: 114 US IPOs raising $147.6bn in 2026 to mid-October.
E. Index rules and lock-up evidence
| Item | Value | Source |
|---|---|---|
| Nasdaq-100 fast entry | From 1 May 2026; tested day 7, announced after day 10, added after ≈ 15 trading days | Leverage Shares |
| SpaceX Nasdaq-100 entry | 7 July 2026 | Nasdaq.com |
| S&P 500 rules | 12-month seasoning, four profitable GAAP quarters, 10% float | AI Weekly (Ars Technica) |
| Lock-up expiry effect | +40% permanent volume; ≈ −1.5% three-day abnormal return (1,948 lock-ups) | Field and Hanka (2001) |
F. Market set-up and pre-IPO prices
| Item | Value | Source |
|---|---|---|
| Order book | ≈ 2× day 1; 6× day 4; ≈ 7× by 16 Nov; SpaceX ≈ 2× on day 2 | Reuters; Bloomberg |
| xyz:ANT perpetual | $111.10 (1 Nov), $116.95 (5 Nov), $101.68 (6 Nov), $102.87 (10 Nov) | FourWeekMBA; CoinGecko |
| Nasdaq Composite | −5.52% to 24,315 (5 Nov); ≈ 12% below 6 Oct record | CNBC via FourWeekMBA; Reuters |
| SpaceX | $135 IPO; $160.95 day one; $225.64 high; $104.83 low; $143.09 (10 Nov) | Gotrade; SpaceNexus |
| Nvidia | Huang resigned 12 Nov; Colette Kress interim CEO; earnings 17 Nov | CNBC; Nvidia 8-K; Yahoo Finance |
| Hyperscaler debt | $420bn projected for 2027 | Goldman Sachs via Reuters |