Anthropic's Silicon Valley Story: From Leaving OpenAI to Filing for an IPO

Text | Beyond the Page, Author | Huahua, Banjun

The story of Silicon Valley often begins with a parting of ways.

Five years ago, a group of core researchers at OpenAI chose to leave. They had different judgments about the direction of AI development, and decided to start over elsewhere because of it.

That departure later gave rise to an AI company whose valuation and revenue both climbed rapidly: Anthropic.

In 2021, in an old factory building at Pioneer Square in San Francisco, Dario Amodei and his sister Daniela Amodei were packing their belongings into cardboard boxes.

Those who left OpenAI with them included core researchers who later took part in GPT-4 development, as well as key members of OpenAI's safety team.

The reason for leaving OpenAI was not complicated: CEO Sam Altman was pushing OpenAI to transform from a nonprofit research institution into a company pursuing high valuations and commercialization. Dario believed this rapid sprint was letting safety slip out of control.

They named the new company Anthropic (meaning human-centered), and drew up a charter, hoping to build a laboratory devoted to restraint and safety amid Silicon Valley's AI frenzy.

English text is being prepared.

It was precisely this determination to be independent of Sam Altman that pushed it onto a capital game from which there was no turning back.

I. Seed Funding and the Chips That Were Chosen

To train frontier models capable of rivaling OpenAI, compute is an unavoidable physical threshold, and compute requires enormous funding to support.

To preserve the independence of its research, Dario deliberately avoided bringing in traditional Silicon Valley venture capital in the early days.

But fate, here, produced a dramatically interlocking turn. In 2022, the one who provided Anthropic with $580 million in early funding was Sam Bankman-Fried (SBF), founder of the cryptocurrency exchange FTX.

Yet just a few months later, FTX collapsed, and SBF was brought to court.

As part of assets that were nearly deemed to be liquidated, Anthropic's equity faced the risk of judicial auction, and the company was once on the edge of survival.

The market waits for no one, and AI competition iterates by the day, even by the hour. This sudden funding rupture meant Dario needed to find, in an extremely short time, a rescuer large enough and able to provide stable compute.

Cloud giants, sensing the opportunity, stepped in.

In the fall of 2023, Amazon announced an investment of up to $4 billion in Anthropic, and Google followed with $2 billion. The funds eased the immediate crisis, but also fixed the real price of this rescue.

To implement the agreement, AWS even built a dedicated large cluster of Trainium2 chips for it. Anthropic also agreed to give Amazon early priority tuning rights over its architecture improvements, deeply binding chip design and algorithm optimization together.

II. Refusing the Throne

Just as Anthropic was struggling forward under the giants' infrastructure framework, a turning point that seemed capable of changing the landscape suddenly appeared.

On the evening of November 17, 2023, OpenAI's board abruptly fired CEO Sam Altman. In the following 48 hours, the chaotic OpenAI board secretly called Dario Amodei and proposed an unexpected plan: inviting Dario to take over as OpenAI CEO and merge the two companies.

This was the closest moment for Dario to retake command of his old team and integrate resources.

But he chose to refuse.

He saw clearly the conflicts of interest behind that governance crisis, and ultimately returned to Anthropic's laboratory to continue advancing Claude's development.

This refusal also led Sam Altman, after returning to OpenAI, to accelerate a compute expansion plan on the scale of hundreds of billions of dollars.

Sam Altman, skilled at the capital game, kept squeezing Anthropic's market space with offensive moves. To avoid being left behind, Anthropic had to multiply the scale of its training clusters in developing the Claude 3 and 3.5 series.

By the time of the Claude 3.5 Sonnet iteration, the compute cost consumed by a single complete training and alignment run alone had already reached several times that of the previous generation's flagship model.

To avoid being opened up a generational gap by OpenAI in model iteration, Dario had no way back and could only follow by pushing Claude's training scale and compute consumption to ever higher levels.

And to support this unprecedented compute expenditure, Anthropic needed to find a business model that could quickly monetize.

III. A Business Model Forced into Being

Under the financial pressure of compute costs doubling month by month, Anthropic could not, like OpenAI, rely on a general-purpose chat box aimed at the masses to slowly cultivate consumer payment habits; it went directly to customers willing to pay big money for efficiency.

So Anthropic made a key strategic pivot: it gave up competing with ChatGPT for consumer traffic and turned to the enterprise battlefield.

In 2025, Anthropic launched Claude Code, a programming tool that runs directly embedded in the terminal command line.

This breakthrough came from the team's restructuring of the underlying Prompt and computation graph architecture.

In the past, developers had to manually package and paste the code files of an entire project into a web box, whereas Claude Code can directly mount the local file system and precisely invoke indexes in the background of the command line, automatically completing cross-file refactoring at the scale of tens of thousands of lines without leaking the enterprise's core assets.

It did not spend resources on generic AI conversation; instead, it let the model enter the software development process directly, handling code refactoring, bug hunting, and architecture debugging for enterprises. Engineers no longer treat Claude as simple Q&A software, but as engineering productivity whose input-output ratio can be quantified.

Revenue paid by actual compute usage (API) then exploded: in 2025 Anthropic's revenue reached $4.585 billion; by the first half of 2026, this figure had surged directly to $16.286 billion, and enterprise customers spending over $100,000 annually surged to more than 11,900.

However, the commercial explosion did not resolve Dario's predicament; instead, it pushed him into an even more urgent situation.

The details in the prospectus reveal the coldness of this business model. In the third quarter of 2026, more than 80% of revenue came from new model versions released less than six months earlier.

This means that rapid iteration of model capabilities will cause the commercial value of old versions to decay within a short period.

The more enterprise customers rely on Claude to handle complex tasks, the more Anthropic needs to continuously invest funds to purchase higher-performance compute to train new models.

The product developed to earn money to pay compute bills ultimately requires the company to invest even greater compute.

IV. The Hidden Clauses in the Prospectus

On October 26, Anthropic officially submitted its prospectus to the U.S. Securities and Exchange Commission (SEC), planning to list on Nasdaq under the ticker ANT.

The filing disclosed the interlocking operating results of recent years.

In the first half of 2026, the book loss was as high as $80.748 billion. Of this, about $72.4 billion came from fair value changes in convertible notes caused by rising valuations. Under Non-GAAP measures, its operating profit in the first half had already reached $1.618 billion, and operating cash flow also turned positive to $3.929 billion.

But what truly put an end to this five-year story was the hundreds of pages of infrastructure procurement contracts in the prospectus notes.

To fulfill its earlier compute commitments to the giants and maintain model iteration, Anthropic signed long-term arrangements totaling about $637 billion, including:

If other compute contracts such as those with Microsoft and Lambda are added, the overall long-term commitment scale is even higher.

Behind Broadcom's hardware partnership worth up to $161.2 billion lies customized binding for custom ASIC chips and next-generation optical interconnect buses. This means that Anthropic's future compute foundation is already tightly bound to Broadcom's supply chain at the physical hardware level.

The data also shows that in 2025, 47% of Anthropic's sales (about $2.16 billion) were completed through Amazon AWS and Google Cloud marketplaces, for which it paid $351 million in distribution channel fees.

Amazon and Google are both shareholders of Anthropic and its largest compute suppliers and distribution channels, while also developing their own Gemini and hardware chips.

V. Silicon Valley's Fate

Anthropic's IPO is not simply a company going to the public market; it is also a rare collective bet in Silicon Valley in more than two decades.

Since Google's listing in 2004 and Meta's (Facebook's) bell-ringing in 2012, Silicon Valley has not witnessed such a massive listing event with equity and infrastructure so highly integrated for a long time.

But this is no longer the Silicon Valley story people once knew.

For the past half century, the myth Silicon Valley prided itself on was the combination of garage startups and venture capital.

Two young people wrote code in a garage, got a few million dollars in venture funding, and quickly overturned huge industry giants through product experience.

That was the era of Fairchild Semiconductor, the era of Apple and Microsoft, and also the era of early Google and Facebook.

And Anthropic's listing declares the end of this go-it-alone model.

No startup can run this marathon independently without the giants' cloud infrastructure.

The essence of this listing is that Silicon Valley giants, using capital, chips, and data centers, have co-opted a rebel that originally tried to challenge the old system back into the existing Silicon Valley industrial order.

Amazon, Google, Microsoft, and Broadcom are no longer just financial investors sitting in the audience; they are funders, lessors, channel distributors, and the biggest winners of this technological revolution.

Even top university laboratories and research institutions, once independent and neutral forces, cannot afford computing clusters made up of tens of thousands of GPUs and have to attach their research projects to the giants' compute clouds. This concentration of resources in a few giants is reshaping Silicon Valley's entire innovation model.

Dario Amodei, the idealist who once walked out of an old factory building in San Francisco, spent five years completing a long circular journey.

He escaped Sam Altman's battlefield, but ultimately handed in his answer sheet to the new order of all Silicon Valley.

[Beyond the Page]'s note:

Every generation of technological revolution changes the threshold for innovation.

As it becomes harder for an era in which code alone could launch a company, how much compute one has and how much capital one can mobilize also begin to determine who is qualified to compete.

Technology is still moving forward, but the threshold for innovation is rising.

Silicon Valley's next round of reshuffling may start right here.