Anthropic: The $1.5 Trillion Company That Let ChatGPT Go First

Last Updated:
Seven former OpenAI colleagues kept Claude out of public hands because they feared starting an AI race. Their old company started it anyway. Four years later, Anthropic has priced its Nasdaq IPO at $83 a share. The founders have told investors there’s one decision they’re prepared to make all over again
Dario Amodei (in pic) and six colleagues celebrated their last day at OpenAI on a video call. Some of Anthropic’s earliest meetings took place in his sister Daniela’s San Francisco backyard
Dario Amodei (in pic) and six colleagues celebrated their last day at OpenAI on a video call. Some of Anthropic’s earliest meetings took place in his sister Daniela’s San Francisco backyard Credits: Getty Images

The model works. That’s the problem.

It’s the summer of 2022. Anthropic has about 50 people, a working chatbot and a question going round on Slack: should we release it?

Some employees want to wait. What worries them is what happens if Claude succeeds. Once the public gets a look at it, other labs will have every reason to hurry their own chatbots out the door. How much time will anyone spend testing the second or third model once the first is loose? Dario Amodei, Anthropic’s chief executive, shares enough of that fear to hold Claude back.

Sign up for Open Magazine's ad-free experience
Enjoy uninterrupted access to premium content and insights.

November brings two blows, nineteen days apart.

First comes the collapse of FTX. Only that April, Sam Bankman-Fried had led a $580 million funding round for Anthropic. Now his crypto exchange is bankrupt. Among the assets its creditors will eventually sell is a stake in a small AI lab whose chatbot hardly anyone outside the business has heard of.

open magazine cover
Open Magazine Latest Edition is Out Now!

The Wealth Issue 2026

06 Nov 2026 - Vol 05 | Issue 45

Portraits of Dreamers and Doers under 35

Read Now The Wealth Issue 2026

Then, on November 30, Anthropic’s old colleagues at OpenAI release ChatGPT.

If you worked at Anthropic that afternoon, you could open a browser and try the competition. Most of the world had no idea you were the competition.

Two years later, a Time reporter asked Amodei about that summer. By then he knew what going first had done for his former employer. He also knew that the race his colleagues had worried about had begun anyway.

“I suspect it was the right thing to do,” he said.

HOW CLAUDE FOUND A WAY IN

Once ChatGPT was out, keeping Claude private could do little to slow OpenAI down. Anthropic could still choose who saw its own model and learn something from watching them use it. By March 2023, when the company formally introduced Claude, a handful of partners had been testing it for months. Among them were the workspace company Notion, the question-and-answer site Quora and the search engine DuckDuckGo.

It was a small opening, but it offered the seven founders a way around a formidable problem. OpenAI had given the public a good reason to go to ChatGPT. Anthropic could take Claude to places people already went to work. A Notion user didn’t have to make a grand decision about which chatbot to adopt. There was a job to do and a tool at hand.

Claude Code began with a song. Engineer Boris Cherny asked a prototype what was playing on his Mac. Claude wrote a few lines in AppleScript to find out. Cherny didn’t know the language
Claude Code began with a song. Engineer Boris Cherny asked a prototype what was playing on his Mac. Claude wrote a few lines in AppleScript to find out. Cherny didn’t know the language Credits: Getty

The seven had started out in much the same fashion themselves, finding places to work during the pandemic. On their last day at OpenAI, they got on a video call to celebrate the decision to leave. In the early months they met mostly on Zoom. When they could get together, the Financial Times reported, it was often in Daniela Amodei’s small backyard in San Francisco. Sometimes they met at Sam McCandlish’s house or at Dario’s.

Between them, they understood both the appeal of building a more powerful model and the difficulty of explaining what it had learnt. Tom Brown had led engineering on GPT-3. Jared Kaplan had studied how a model’s capabilities changed as researchers fed it more data and computing power. Chris Olah worked on understanding what went on inside it. Jack Clark came from policy, McCandlish from safety research. Daniela had led safety and policy work at OpenAI. Together with Dario, they had helped build the advances that now gave them pause.

A backyard was ample space to discuss the next generation of Claude. Training it was another matter. In September 2023, Amazon announced an investment of up to $4 billion. Anthropic chose Amazon Web Services as its primary cloud provider and agreed to use Amazon chips to train and run future models. The arrangement also put Claude within reach of companies already shopping for AI tools through Amazon’s Bedrock service. Anthropic had found an investor, a supplier and a route to customers in one deal. Amazon would complete the $4 billion investment the following March.

Getting Claude in front of a company was one problem solved. The harder one was finding a job people would give it often enough, and value highly enough, to keep paying for it.

One answer started with a song.

THE SONG ON BORIS CHERNY’S COMPUTER

Boris Cherny joined Anthropic in September 2024. One of his early experiments put Claude in the terminal, the plain text window where programmers type commands. He connected the prototype to AppleScript, a language that lets applications on a Mac talk to one another. Then he asked Claude what music he was listening to.

Claude wrote a little piece of code, asked the music player and told him. Cherny himself didn’t know AppleScript.

Dario Amodei held Claude back in 2022, worried that releasing it would start an AI race. OpenAI launched ChatGPT anyway. Asked later whether he’d made the right call, Amodei told Time: “I suspect it was the right thing to do.”
Dario Amodei held Claude back in 2022, worried that releasing it would start an AI race. OpenAI launched ChatGPT anyway. Asked later whether he’d made the right call, Amodei told Time: “I suspect it was the right thing to do.”  Credits: Getty

As a way of identifying a song, it was absurdly elaborate. He could have glanced at the screen. What interested Cherny was how little instruction the machine had needed. He had told it what he wanted to know. Claude had worked out where to find the answer and written code in a language its user didn’t speak. What else could it do if he let it use more of the computer?

Cherny and his colleague Cat Wu discussed giving the prototype more room. Once it could write and run code in the terminal, Cherny found himself using it every day. Other engineers could hand it a task in a project they already knew, inspect the changes and run the tests. Coding had an advantage over some of the grander promises being made for AI. The customer had a reasonably quick way to discover whether the thing worked.

SEVEN HOURS AT RAKUTEN

In Japan, a machine-learning engineer named Kenta Naruse gave Claude Code a task in an open-source software library containing some 12.5 million lines. He wanted a way to extract particular numerical signals from an AI model. He already had a reference method against which he could check the result.

Claude Code ran for seven hours. In a case study published by Anthropic, Naruse recalled what he did during that time: “I didn’t write any code during those seven hours. I just provided occasional guidance.” Rakuten reported 99.9% numerical agreement with the reference.

Word got round. Other Rakuten engineers started asking about the experiment. A colleague had used the tool on a difficult job, checked its work and could tell them exactly how much help it had needed. Anthropic didn’t have to make the next introduction. Naruse made it for them.

At Japan’s Rakuten, Kenta Naruse gave Claude Code a job that ran for seven hours. He offered occasional guidance without writing a line of code himself. When word got round, his colleagues wanted to try it
At Japan’s Rakuten, Kenta Naruse gave Claude Code a job that ran for seven hours. He offered occasional guidance without writing a line of code himself. When word got round, his colleagues wanted to try it Credits: Getty

Claude Code became generally available in May 2025. Six months later, Anthropic said it had reached $1 billion in annualised revenue. That was the pace of sales at the time, not a billion dollars already earned. By September 2026, the IPO filings offered a firmer measure for the business as a whole. About 250 customers had each accounted for at least $10 million in revenue over the preceding twelve months. More than 11,900 had generated at least $100,000 apiece. Those figures cover all of Anthropic’s products, not just Claude Code.

The seven had given up their chance to introduce the public to a Claude-like chatbot first. They had found their way onto the purchasing budgets of companies willing to spend eight figures a year. Then came a quarter in which sales jumped by nearly half and the operating profit disappeared.

WHO PAYS FOR THOSE SEVEN HOURS?

In the three months to September 30, 2026, Anthropic’s revenue rose 47% from the previous quarter to $16.96 billion. In that earlier quarter it had earned an operating profit of $974 million. In the bigger one, it recorded an operating loss of $567 million.

Go back to Naruse’s afternoon. For Rakuten, part of the attraction was that an engineer could let a task run for seven hours and do something else. At Anthropic’s end, the computers were still at work. Give an engineer the ability to set several jobs running at once and the saving on one side creates demand for more machines on the other.

The price of securing those machines was one pressure on margins. Anthropic disclosed that it took delivery of $5 billion worth of computing capacity under short-term commitments in the third quarter. It said this capacity cost significantly more than comparable computing power booked years ahead.

Buying early gets Anthropic better terms. It also requires somebody to decide, years in advance, how many engineers like Naruse there will be and how much work they will hand over. Book too little and a customer ready to roll Claude out to an entire team may find a rival with capacity to spare. Book too much and the supplier still expects payment.

At $83 a share, Anthropic’s founders gave buyers fair warning: they might keep a product off the market even if a rival won the business. They had already delayed a model over cybersecurity concerns and told investors the decision had cost them sales
At $83 a share, Anthropic’s founders gave buyers fair warning: they might keep a product off the market even if a rival won the business. They had already delayed a model over cybersecurity concerns and told investors the decision had cost them sales Credits: Getty

In its prospectus, Anthropic put its purchase and lease commitments at roughly $900 billion, overwhelmingly for computing and infrastructure. About $70.7 billion was due in the twelve months after September 30, 2026. Roughly $451.6 billion was expected to be incurred before the end of 2030. Most of the rest stretched from 2031 to 2045. Anthropic has time to grow into those purchases. Its suppliers have contracts if it doesn’t.

The Amazon deal looks different in that light. What began as a young lab’s solution to three problems became a feature of its finances. Consider Google: it supplies infrastructure to Anthropic and is one of seven technology companies that agreed to buy a combined $26 billion of its shares in a private placement alongside the IPO. Nvidia and Microsoft are in that group, too. Their investment signals confidence in Claude. Several of them also have a commercial interest in Anthropic remaining rich enough to buy what they sell. Both can be true of the same cheque.

WHAT $83 A SHARE BOUGHT

On November 10, Anthropic began marketing 825 million shares at an expected $70 to $78 each. Eight days later, it priced 980 million at $83. The enlarged public offering was worth $81.34 billion, of which about $79.38 billion was attributable to shares sold by Anthropic before fees. Existing shareholders accounted for the balance; the $26 billion private placement was separate. The Nasdaq ticker was ANT.

The $1.5 trillion is the fully diluted valuation at the offer price, which takes potential additional shares into account. Reuters put the non-diluted figure at about $1.1 trillion. The underlying bet was the same: that enough companies would put enough work through Claude to justify the infrastructure Anthropic had already ordered.

One early backer had other obligations. In April 2024, the FTX bankruptcy estate closed the sale of nearly 29.5 million Anthropic shares for about $884 million. The seller in bankruptcy court had creditors to repay. The buyer at $83 had a different document to read: a letter from seven founders explaining the circumstances in which they might choose to forgo revenue.

They had a recent example.

WOULD THEY HOLD CLAUDE BACK AGAIN?

While testing a model called Claude Mythos Preview, Anthropic’s researchers concluded that its cybersecurity capabilities had advanced beyond what they felt comfortable releasing widely. In their letter to prospective shareholders, the founders said they restricted access until they could build additional safeguards. Those capabilities subsequently reached the public through Claude Fable 5. The founders said the delay cost Anthropic revenue and gave rivals an advantage.

Sam Bankman-Fried led a $580 million funding round for Anthropic in 2022. Two years later, the bankrupt FTX estate sold nearly 29.5 million Anthropic shares for $884 million. The creditors had a payday long before the IPO
Sam Bankman-Fried led a $580 million funding round for Anthropic in 2022. Two years later, the bankrupt FTX estate sold nearly 29.5 million Anthropic shares for $884 million. The creditors had a payday long before the IPO 

Caution can make commercial sense. An engineer trying Claude on a single task can inspect its work herself. An executive giving thousands of employees access has to answer for what happens on machines she will never see. A supplier willing to explain why it withheld a powerful model may be easier to trust with a large contract. The founders acknowledge that possibility in their letter, although they say it is not the reason they make such decisions.

Sometimes, of course, the rival wins the account and the customer never comes back. The founders asked shareholders to make room for that outcome, too.

In the summer of 2022, the question went round a Slack channel at a company of about 50 people. Should we release it? Some said wait. Their chief executive agreed. For most people outside that channel, the first sign of what a chatbot could do came when somebody else’s appeared on their screen.

The next time the question goes round Anthropic, an engineer in Japan may be waiting for the model. A cloud provider may already have reserved the machines. A shareholder will be able to put a price on another quarter of delay.

The seven have told all three that their answer could still be wait.