With a valuation of $965 billion in five years, surpassing OpenAI, what gives it the edge?
Yesterday, in an unassuming office building in San Francisco, a draft S-1 document was quietly submitted to the electronic system of the SEC (U.S. Securities and Exchange Commission).
There was no press conference, no advance notice, and even most employees of this company learned about it from the news the next day.
Signature on the document: Anthropic. Valuation: $965 billion.
A company that few people had heard of three years ago, with a valuation of only $4.1 billion a year ago, has now reached a valuation of $965 billion?
A valuation of $965 billion is equivalent to two Tencents or three Alibabas. And the company that achieved all this has fewer than 5,000 employees, with only 5% of the monthly active users of ChatGPT. It also voluntarily rejected a $200 million military contract in early 2026.
What's even more impressive is that this company, which has only been established for five years, has an annualized revenue of over $47 billion, equivalent to 340 billion RMB, which is equivalent to two Moutais. It took Moutai 73 years to reach its current status.
Many people predict that Anthropic is likely to become the first company in human history with an annual revenue of over $1 trillion.
What does this mean? The combined revenue of State Grid and PetroChina, the two largest revenue-generating companies in China, was approximately $1.04 trillion in 2024. A company that didn't exist six years ago may have more revenue than the two of them combined.
This story cannot be explained by the traditional "growth methodology" at all.
Coincidentally, there is a book that provides a framework for understanding: 10x Is Easier Than 2x (Chinese title: Why 10x Growth Is Easier Than 2x. For the sake of simplicity, it will be referred to as 10x hereinafter). The authors are Dan Sullivan and Benjamin Hardy.
Dan Sullivan spent 30 years coaching more than 25,000 entrepreneurs. His core conclusion is very counterintuitive: 10x growth is actually easier than 2x growth.
Today, we will use this framework to interpret the story of Anthropic and help you understand a very important question: When everyone is "working harder," what does "working smarter" really look like? Finally, here are four insights for Chinese entrepreneurs.
I. Starting Point: It's Better to Be Different Than to Be Better
1. How did the miracle of growing from $4.1 billion to $965 billion in one year happen?
Many people will find it incredible when they see Anthropic's valuation curve.
Early 2024: $4.1 billion; early 2025: approximately $60 billion; May 2026: $965 billion.
If this growth is understood as "running faster," the key point is completely missed. Anthropic doesn't compete with OpenAI on who can run faster; instead, it directly switches to a different track.
While OpenAI was developing hundreds of millions of C-end users with ChatGPT, Anthropic was signing enterprise contracts. OpenAI aims for full-scenario coverage, while Anthropic focuses on single-point penetration. OpenAI pursues "being able to do everything," while Anthropic pursues "being trustworthy once it does something."
The valuation logics of these two companies are also completely different: OpenAI relies on the C-end user base and general capabilities, while Anthropic relies on its irreplaceability in enterprise contracts.
How many companies are there in the world? Millions. Anthropic's products can be integrated into all their processes, so its market is not a specific track but the entire business world.
These are two completely different "growth structures."
Using the framework of 10x to understand, 2x growth is "doing more things with the same structure": more users, more scenarios, more functions. 10x growth is "doing fewer things with a different structure": fewer but deeper customers, fewer but more irreplaceable capabilities, fewer but clearer positioning.
Claude Code, Anthropic's AI programming tool, had an ARR (Annual Recurring Revenue) of $2.5 billion in February 2026. It only took nine months from its launch to reach this figure.
Where did this $2.5 billion come from? It wasn't through channel expansion, advertising, or price wars. Claude Code itself is a development tool. The more developers use it, the stronger Claude's capabilities become (through data feedback from usage), which then attracts more developers. This is a self-accelerating growth engine.
10x divides time into two types: Chronos time (linear, quantitative, 9-to-5) and Kairos time (non-linear, qualitative, critical moments). Most companies' growth is Chronos-style, investing resources step by step and linearly obtaining returns.
Anthropic's growth is Kairos-style. It seizes a few self-amplifying leverage points (enterprise market, security positioning, developer ecosystem) and then lets these leverage points reinforce each other.
This is why "10x is easier than 2x." It's not that 10x growth doesn't require effort; it's that 10x growth forces you to give up the illusion of using brute force and find the fulcrum that can move a thousand pounds with a little effort.
2. What You Don't Do Is Your Moat
The core proposition of 10x is based on a simple distinction: 2x growth is about competing in the existing stock. It involves adding more within the original framework, such as working overtime, optimizing processes, engaging in price wars, and grabbing market share. There are too many paths, and each seems feasible. Eventually, you fall into choice paralysis, trying a little of each but not fully implementing any of them.
10x growth is about qualitative reconstruction. A 10x goal directly eliminates 80% of the existing paths. You can't achieve 10x growth by working harder, so you're forced to find the leverage points for qualitative change, and the path becomes extremely simple and clear.
2x growth makes you want to try everything, while 10x growth forces you to keep only the most important things.
The establishment of Anthropic is an example of this logic.
At the end of 2020, Dario Amodei was still the vice president of research at OpenAI. He led the development of GPT-2 and GPT-3 and was a co-inventor of RLHF (Reinforcement Learning from Human Feedback). He was an absolute technical core in the AI circle.
However, he couldn't reach an agreement with Sam Altman on a fundamental issue: the development speed and safety boundaries of AI.
Altman's approach was to prioritize performance, accelerate commercialization, and accept a $10 billion investment from Microsoft. There was nothing wrong with this path, and later facts proved that it was feasible. But Dario thought that rushing full speed without building safety barriers was gambling on an irreversible future.
In February 2021, he left OpenAI with a few colleagues, including his sister, Daniela Amodei, who was the vice president of safety and policy at OpenAI.
Daniela later said something worth thinking about: "If we don't leave now, we won't have a chance to prove that another way is feasible."
Note that he didn't leave because "he couldn't continue" (this is a decision driven by "need"). Instead, he left because "we believe another way is feasible, and if we don't leave, it will be too late to prove it" (this is a decision driven by "want").
10x has a specific concept to distinguish between "need" and "want."
"Need" comes from scarcity and fear. You lack computing power, so you have to take Microsoft's money; you lack time, so you have to release quickly; you lack an advantage, so you have to be the first.
"Want" comes from freedom and real choice. You do something because you believe in it.
Most startup companies start with "need," while Anthropic starts with "want."
If we look purely at the business results, companies driven by "need" can also be very successful, as OpenAI itself proves. But companies driven by "want" have a structural advantage: on the first day of their establishment, they answer a question that companies driven by "need" may never answer - "What won't I do?"
In fact, most startup companies are pushed forward: they raise funds because they're out of money; they focus on growth because they have no users; they engage in price wars because they have no advantage. Every step is passive.
This difference may sound like a cliché, but as you'll see later, it directly determines what a company can cut and what it can reject. These "don'ts" are its deepest moat.
II. Subtraction: Turn Rejection into Pricing Power
The most core sentence in 10x is: "10X isn't about more. It's about less."
Dan Sullivan uses the metaphor of Michelangelo throughout the book. Someone asked Michelangelo how he carved the statue of David so perfectly. He replied: It's simple. I just removed the parts that didn't belong to David.
Anthropic's history is about constantly "removing the parts that don't belong to David." Three cuts were particularly precise.
First Cut: Don't Target the C-end Market
Claude App has approximately 30 million monthly active users, while ChatGPT has over 600 million. The former is only 5% of the latter. In any traditional analysis framework, this would be considered a huge failure. For a product like an AI assistant, aren't network effects and data flywheels supposed to be driven by user scale?
But this was a deliberate choice.
Anthropic has bet almost all its resources on the enterprise market. What's the result? The number of enterprise customers with an annual payment of over $1 million has increased from a dozen two years ago to over 1,000. Eight out of the top 10 Fortune companies are Claude's customers. Anthropic holds 65% of the market share in enterprise AI spending, while OpenAI only has 20%.
The logic behind this is simple: The cost for C-end users to switch AI assistants is almost zero. They may use Claude today, ChatGPT tomorrow, and Gemini the day after tomorrow. They'll use whichever model is stronger, with no loyalty at all.
But enterprise customers are different. Once they integrate AI into their core business processes, such as Bridgewater Associates using Claude to analyze economic data and Thomson Reuters using Claude to build a legal AI assistant, the switching cost is so high that it's almost impossible to replace.
For example, C-end users are like supermarket shoppers who go to the store with the best promotions. Enterprise customers are like members with annual cards. It's much easier for them to renew their membership than to switch to another store. Anthropic has chosen to focus on the membership business.
There is a corresponding case in 10x: Stream Logistics, a logistics company, gave up 95% of its regular freight customers and focused on 5% of high-risk special freight customers. The team size remained the same, but the profit quadrupled.
The principle is the same: Regular customers have a large volume but thin profit margins and high churn rates. High-risk customers have a small volume but high unit prices and strong loyalty.
Second Cut: Don't Create "People-Pleasing AI"
Anthropic has defined a 3H priority model behavior framework: Harmless > Honest > Helpful.
Most AI companies have the opposite priority: They prioritize being useful first, then being accurate, and as for safety, they do as much as they can.
Anthropic has reversed this order. If a user's request may produce harmful output, Claude will directly and actively refuse to execute it.
They'd rather offend users than take safety risks.
10x has a concept called the "fitness function." What you choose to optimize determines what you become.
Most AI companies optimize "capability" and "user growth." Anthropic's top priority for optimization is "credibility." Different fitness functions lead to completely different product choices.
Look at a set of data: In the Vectara ranking test in April 2026, the hallucination rate of Claude Opus 4.6 was approximately 4%, while that of GPT-5.4 was approximately 6%. The difference may not seem significant, but the implementation methods are completely different. When Claude is unsure about something, it will say "I'm not sure." It's not that it's smarter; it's just more willing to admit it doesn't know. It'd rather say one less thing than make something up.
Think about it. Which AI would you entrust your core business to: an AI that actively says "I can't do this" or an AI that takes on anything?
Third Cut: Reject the Pentagon
In February 2026, the U.S. Department of Defense asked Anthropic to remove Claude's safety barriers so that the military could use it. Anthropic clearly refused and then drew two red lines for itself: Don't participate in large-scale domestic surveillance and don't develop fully autonomous weapons.
The $200 million military contract was gone. A few hours later, Sam Altman announced that OpenAI had reached a cooperation agreement with the Pentagon.
10x has a principle called "Always Be the Buyer" - take the initiative in relationships and don't compromise on low-value cooperation. The judgment criterion is not "how good the conditions offered by the other party are" but "whether this cooperation deviates from your unique capabilities."
For Anthropic, its unique capability is "safe and trustworthy AI." The money from the Department of Defense is money, but taking this money would mean sacrificing its positioning, which would be a low-value cooperation.
After these three cuts, Anthropic answered a deeper question: Where does pricing power come from?
The traditional answer is: Pricing power comes from monopoly, network effects, and technological barriers.
Anthropic provides another answer: Pricing power can also come from "knowing what you definitely won't do." When customers find that only you will refuse a request for safety reasons and that your self-imposed restrictions are stricter than any regulatory requirements, trust turns into a premium.
Your "don'ts" will make others think you're reliable. And reliability is something that can be monetized.
III. Division of Labor: One for the Vision, One for Survival
The most core operating system in 1