Five years, seven people, staged a stunning counterattack against their former employer, sprinting for the largest IPO in history.

Dario Amodei is one of the most high-profile figures in the current AI industry.

On October 26, Anthropic, the company he founded, officially submitted its S-1 filing, planning to list on NASDAQ under the ticker symbol "ANT" with a valuation ranging from 1.5 trillion to 2 trillion US dollars. Estimated by this scale, this IPO will surpass the record set by SpaceX in June this year and become the largest IPO in history. It took him less than 5 years to grow the company from its founding to a market capitalization of nearly 2 trillion US dollars.

At the end of 2020, due to various disagreements with the OpenAI management, Dario left the company and founded Anthropic in early 2021. In March 2023, Anthropic released its first-generation Claude chatbot; in July of the same year, it launched Claude 2, which received extensive market attention and wide acclaim.

However, Dario did not focus on developing consumer applications, but was mainly committed to building AI for enterprise application scenarios. With the explosion of demand for AI code generation, Anthropic launched Claude Code, an AI programming tool, in February 2025, and the company's revenue grew rapidly accordingly. The prospectus discloses that in the first half of 2026, Anthropic's revenue was 162.86 billion US dollars, a year-on-year increase of nearly 13 times, and the net dollar retention rate of customers reached 900%.

In just 5 years, Dario overtook his former employer OpenAI and will be the first to complete the listing. But reality is not a rags-to-riches fantasy story, behind the rapid growth, the company needs to spend heavily on R&D, computing power, infrastructure and other aspects. According to the prospectus, the company has signed future computing power commitments exceeding 6.5 trillion US dollars. Previously, due to the pressure on the AI sector and the slowdown of the company's Q3 revenue quarter-on-quarter growth rate, some investors have discussed lowering Anthropic's valuation expectation from 2 trillion US dollars to about 1.5 trillion US dollars.

At a time when the profit distribution in the AI industry chain is extremely unbalanced and tech giants are taking on massive debt for infrastructure, can the capital and patience of the public market really support this "largest IPO in history"?

The "Seven Rebels" from OpenAI Built Anthropic

The opening chapter of *The History of Venture Capital* tells the story of the "Traitorous Eight" who collectively left Shockley Semiconductor Laboratory, and eventually gave birth to Fairchild Semiconductor and the entire Silicon Valley. More than half a century later, the same scenario played out in San Francisco. The birth of Anthropic also started with a Silicon Valley-style collective departure.

As the former Vice President of Research at OpenAI, Dario made great contributions to the R&D of GPT-2 and GPT-3, and was deeply involved in the early commercialization process.

However, amid the rapid iteration of models, Dario began to question the direction of AGI. He repeatedly emphasized to the outside world that this technology must be advanced in an extremely cautious, direct and honest manner. The increasingly radical commercialization pace of OpenAI, as well as the fundamental value disagreements between him and the management including Sam Altman, led to the complete breakdown of mutual trust.

In December 2020, Dario decided to leave the company. His sister Daniela Amodei, then Vice President of Safety and Policy Operations, Tom Brown, the first author of the GPT-3 paper, Jared Kaplan, the physicist who proposed the Scaling Laws, and six other core key members of OpenAI at that time left with him.

In February 2021, Dario and others, who were determined to start their own business, established Anthropic.

Dario was born in San Francisco in 1983. According to his description, he has a very close relationship with his parents. Therefore, in 2006, affected by his father's death from a rare disease, he, who was then pursuing a graduate degree in theoretical physics, transferred to biophysics. In 2011, Dario entered Stanford University School of Medicine to pursue postdoctoral research, focusing on topics related to neural circuits and tumors, but he soon realized that a single researcher could not achieve research breakthroughs in this field within a limited time.

In order to find technical solutions, Dario paid attention to AI. At that time, deep learning had made major technical progress. In his view, AI has the potential to "take us beyond the human scale".

In November 2014, Dario joined Baidu Silicon Valley Artificial Intelligence Laboratory, and led the DeepSpeech 2 project in the team of Andrew Ng, Baidu's former chief scientist, realizing a system solution that directly converts raw audio into text. This experience laid the groundwork for his later intuition about "Scaling Laws".

In 2015, he briefly joined Google Brain as a senior research scientist, but he could not adapt to the pace and atmosphere of large companies, and chose to leave less than a year later.

In July 2016, Dario joined OpenAI, which was initiated by Elon Musk, Altman and others. He successively served as the head of the AI Safety Research Group, Director of Research, and eventually rose to Vice President of Research.

Dario's attention to AI safety began during his tenure at Google. He also co-authored papers with colleagues to explore the potential undesirable behaviors of AI. After the release of GPT-3, on the one hand, he was excited about the technological progress, on the other hand, he became more and more convinced that many things cannot be controlled just by training models.

This concept began to cause divisions within OpenAI. People who agreed with Dario quickly gathered around him, forming a tight circle; but on the opposite side, the management and some colleagues believed that he was trying to "control" the company with safety, and the conceptual differences became wider and wider.

Michael Berry, Dario's mentor, once commented on him: "He is one of the most talented students I have ever taught, and he is also a rather proud person." This "pride" may be the reason why he later completely broke with Altman and others. On the development (safety) of AGI, he refused to back down and compromise.

Very dramatically, Musk, who had fallen out with Altman, directly fired on the social platform X on February 13 this year after Anthropic announced a $30 billion financing, calling Claude "anti-human and evil". But less than three months later, on May 6, Anthropic officially announced that it would exclusively lease all computing power of the Colossus 1 data center from SpaceX, including more than 220,000 NVIDIA GPUs and 300 megawatts of computing power. Musk also changed his attitude on X: "I talked to the executive team of Anthropic, and I was deeply impressed. No one triggered my evil detector."

On September 12 this year, Dario published a 3800-word long article "We Must Pace the Frontier", advocating that the industry must "slow down the speed of improving model capabilities" in exchange for 1 to 2 years to deploy safety guarantees.

Anthropic's prospectus also reflects Dario's emphasis on the safety narrative. In the 261-page main text, the risk factors section takes up more than 80 pages, while the part describing business operations only accounts for 48 pages.

Under the "arms race" of industry models and the performance pressure of listing, although Dario has repeatedly emphasized slowing down the improvement of capabilities, Anthropic has actually officially released at least 9 models this year, and has successively released 5 models since September.

Click on the cover to subscribe to the magazine for the whole year

The AI Bubble Under Trillion-Dollar Valuation

In fact, since 2025, Anthropic's model release pace has been accelerating, updating about 10 versions of sub-models throughout the year, and its performance has also surged.

The prospectus shows that Anthropic's revenue is growing at a rate of about 10 times per year. From 2023 to 2025, the company's revenue was 41 million US dollars, 386 million US dollars and 45.85 billion US dollars respectively; the revenue in the first half of 2026 was 162.86 billion US dollars.

Anthropic's chosen enterprise-level (to B) business route has been verified by the market. The prospectus shows that in the first half of 2026, pay-as-you-go enterprise API calls contributed about 80% of the revenue, and the number of enterprise customers with annualized consumption of more than 100,000 US dollars skyrocketed from about 950 a year ago to nearly 12,000.

Operating losses have narrowed significantly. In the first half of 2026, Anthropic's operating loss was 693 million US dollars, compared with 2.135 billion US dollars in the same period of the previous year; the operating loss rate dropped from about 184% to about 4%. The company even achieved operating profit in a single quarter in the second quarter of 2026, with an operating profit of 974 million US dollars in the quarter and an operating profit margin of about 8%. After excluding items such as equity incentives and non-cash charitable donations, the adjusted operating profit for the quarter was 2.235 billion US dollars, and the adjusted operating profit margin was 19%.

The technological progress and performance of Anthropic since its establishment have continued to attract capital attention in the primary market. The prospectus shows that as of the end of June 2026, the company has raised a total of 1.411 trillion US dollars in financing. Tech giants such as Amazon, Google, Microsoft, NVIDIA and many leading VCs are on its shareholder list.

However, from the disclosure of public information and the prospectus, after 2025, the investment logic of large manufacturers has undergone substantial changes: capital investment is more and more linked to computing power procurement or chip supply, rather than pure financial investment. The price Anthropic pays for receiving funds is to return funds to cloud vendors and chip vendors in the form of long-term procurement commitments.

The prospectus shows that the company's "computing and infrastructure" expenditure is expanding rapidly: this expenditure was 433 million US dollars in 2023, rose to 2.522 billion US dollars in 2024, and reached 73.26 billion US dollars for the whole year of 2025. In 2026, the computing power expenditure in the first half of the year alone was as high as 89.45 billion US dollars, exceeding the total amount for the whole year of 2025.

Not only that, Anthropic has also signed long-term computing power and infrastructure contracts of more than 6.5 trillion US dollars with 8 companies including Amazon, Microsoft, and NVIDIA, covering the next ten years, about 80% of which are non-cancelable or must be paid regardless of actual usage.

High infrastructure expenditure is a dilemma faced by the entire AI industry. OpenAI estimates that from 2026 to 2030, the investment in computing power and infrastructure will be about 856 billion US dollars; Google's capital expenditure in the second quarter of 2026 rose to 44.9 billion US dollars, and its free cash flow in the quarter fell to negative 5.9 billion US dollars, the first time it turned negative since its listing. In addition, Meta, Amazon and other companies have also experienced negative free cash flow and raised capital expenditure guidance.

This also reflects that at the current node, the value distribution of the AI industry chain is still extremely unbalanced. For every 1 dollar of token-related revenue, the hardware layer accounts for about 70% of the total industry revenue, the cloud service layer accounts for about 20%, and the revenue of the model and application layer accounts for less than 10%. From the perspective of operating profit, the data is more extreme. JPMorgan Chase counted in *China AI Value Chain: The Expanding Cake and Value Competition* that the hardware link accounted for 93% of the industry's operating profit in 2026, and the profit of the model and application layer was close to zero; this situation is expected to improve by 2030: the hardware profit share will shrink to 30%, cloud services 15%, and models and applications rise to 54%.

The latest data shows that Anthropic's Q3 single-quarter revenue is about 169 billion US dollars, and the quarter-on-quarter growth rate has slowed down from 144% in Q2 to about 47%. The prospectus also gives a reminder: The recent growth may reflect the characteristics of early adoption and the initial stage of the market, and may not be sustainable; the company expects the growth rate to decline as the revenue base increases, and may fall faster and more than expected. In the risk factors, the company directly states that it may not be able to achieve or maintain profitability in the future.

At present, model layer companies like Anthropic are in the middle of the entire industry chain with the thinnest profit but the heaviest capital consumption, which also confirms the core logic of the analogy that "AI will become utilities like water, electricity and gas". As infrastructure, they are indispensable, but their profit margin is constrained by factors such as regulation, competition and capital intensity, and the maximum economic value eventually flows to the people who use them, not the people who provide them. In the future, it is more likely that the application layer will hold larger profit margins.

It is still unknown whether model companies including Anthropic can carve out a sufficiently large share of the cake in the application field.

Dario, who is in the industry, has a clear understanding of the judgment that the profit of the model layer is thin and the value eventually flows to the application layer. He once admitted in an interview with Z Finance: "I think there are still a lot of opportunities in the application layer."

But he also put forward a different opinion: model companies release a new version every two or three months, and each upgrade breaks the original capability boundary, making products that were impossible to make before suddenly feasible. In Dario's view, this is exactly where the vitality of the API model lies. It does not statically move towards commoditization, but is dynamic and constantly evolving. As long as the boundary of achievable AI capabilities continues to expand, API is the tool for entrepreneurs to seize these new boundaries at the first time.

This also explains why Dario is willing to bear a computing power bill of more than 6.5 trillion US dollars, but still maintain high-frequency model iteration. Because he knows very well that only when the underlying model capabilities continue to iterate can the application layer thrive. Entrepreneurs should think more about: "What new things that others cannot make can I build based on these models and APIs? What truly new ideas can I come up with?"

This article is from the WeChat official account "China Entrepreneur Magazine" (ID: iceo-com-cn), author: KONG Yue