On May 23, 2026, an event occurred in the AI industry significant enough to reshape the landscape: according to reports from Bloomberg, Cailian Press, and other authoritative media, Anthropic is about to close a new $30 billion funding round, pushing its post-money valuation above $900 billion and officially surpassing OpenAI's latest valuation of about $852 billion to become the world's highest-valued AI startup. Even more striking is the contrast in financial data: Anthropic is expected to post $10.9 billion in revenue in Q2 2026, up 127% quarter over quarter, and to achieve its first-ever single-quarter operating profit of about $559 million — turning a profit two years ahead of its original plan. Over the same period, although OpenAI posted $5.7 billion in Q1 revenue, its adjusted operating margin was -122%, meaning it loses $1.22 for every $1 it earns. This result overturns most people's understanding. For the past three years, the entire industry took it for granted that 'the company with the strongest technology wins,' and OpenAI was seen as the undisputed leader. Why is it that Anthropic, whose technology is not absolutely leading, achieved profitability first and completed the valuation overtaking? To understand this reversal, we cannot look at a single dimension of technology or product; we must return to the essence of business. Every successful business model is, in essence, a complete closed loop driven by a value proposition: whom you choose to solve what unique problem for fundamentally determines what product you build, what customers you win, and ultimately how far you can go.
I. The Value-Proposition-Driven Business Loop
Many people treat PSF, PMF, and GTM as four separate business tools, but in fact they are an organically integrated whole, interlocking and building on one another, and the core that connects all of them is the value proposition.
Value Proposition: This is the starting point and soul of the entire loop. It answers the most fundamental question: for whom do you solve what unique problem?
PSF (Problem-Solution Fit): Derived from the value proposition. The kind of value proposition you choose determines which pain points you need to validate and what solution you provide.
PMF (Product-Market Fit): Derived from PSF. Whether your solution truly solves that pain point, and whether the market is willing to keep paying for it, is the core PMF must validate.
GTM (Go-To-Market Fit): Derived from PMF. What kind of customers your product suits determines how you should reach and convert them.
The failure of many companies is not due to poor technology or weak products, but because this loop breaks. Either the value proposition is unclear, or products are built blindly without validating PSF, or scaling is pursued blindly without achieving PMF, or the GTM approach does not match the product and customers. In the large-model industry, once the technical gap between top models narrows to the point where users barely perceive it, the health of this business loop becomes the single decisive factor in a company's success or failure.
II. Two Completely Different Business Paths: Anthropic and OpenAI
2.1 Anthropic: 'Slow Is Fast,' Determined by Its Value Proposition
From its very first day, Anthropic established a very clear and never-wavering value proposition: to provide the most controllable, compliant, and secure large-model services to large enterprises in heavily regulated industries such as finance, law, and healthcare. The core reason founder Dario Amodei left OpenAI was his belief that OpenAI was deviating from the 'safety-first' course, over-pursuing model-capability improvements while neglecting risk. This divergence ultimately led the two companies down completely different trajectories.
PSF Validation: Precisely Hitting Enterprises' Most Painful Compliance Need
From its founding, Anthropic targeted a pain point ignored by most companies: general-purpose large models carry risks such as data leakage, hallucination, and uncontrollability, and cannot be used in enterprises' core businesses. This pain point is real and urgent: financial institutions such as JPMorgan Chase and Goldman Sachs once completely banned employees from using ChatGPT, fearing leakage of core trading data and customer information; the legal industry demands extremely high accuracy from AI output, where a single hallucination can cause losses of millions of dollars; the healthcare industry is strictly bound by regulations such as HIPAA, which mandate that data must not leave its domain. In response to these pain points, instead of blindly piling on parameters or chasing speed, Anthropic invested heavily in developing the 'Constitutional AI' technical approach. Unlike OpenAI's RLHF (Reinforcement Learning from Human Feedback), Constitutional AI trains models through a clear set of moral and legal principles, making it more transparent, interpretable, and controllable than RLHF. At the same time, Anthropic launched a complete enterprise-grade security and compliance suite supporting private deployment, data isolation, audit logs, and more, fully meeting the compliance requirements of heavily regulated industries.
PMF Validation: High Retention, High Price per Customer, High Gross Margin
Its precise value proposition and solution quickly won Anthropic recognition from enterprise customers. According to the latest data from The Wall Street Journal and CNBC:
Customer retention exceeds 90%, and renewal rates among Fortune 500 clients approach 100%. Million-dollar-level paying customers grew from 500 to 1,000 in just six months; about 85% of revenue comes from enterprise and developer customers, with the Claude Code product alone already generating $2.5 billion in annual revenue; customers are willing to pay 3–5 times the price of general-purpose models, because for them safety and compliance are not a bonus but an entry ticket. More importantly, Anthropic's business model is extremely healthy. Its gross margin exceeds 70%, far above the industry average. This is because enterprise customers have strong paying power, and once their core business data is integrated with Claude, switching costs are extremely high, so they almost never churn.
GTM Fit: Not Chasing Traffic, Only Precise Customers
Unlike OpenAI, Anthropic abandoned the consumer (C-end) market from the start. It does not use free products for traffic, does not wage price wars, and does not pursue user-count growth. Its customer acquisition method is simple but extremely effective:
Word-of-mouth from flagship customers: leveraging the industry influence of early flagship clients such as JPMorgan Chase and Goldman Sachs to win more customers in the same industries; embedded partnerships with cloud providers: selling through the cloud marketplaces of Amazon AWS, Google Cloud, and Microsoft Azure to reach enterprise customers worldwide.
Dedicated direct-sales teams: dedicated direct-sales teams for Fortune 500 enterprises, providing one-on-one customized service. Although this approach looks slow, its customer acquisition cost is extremely low. According to disclosures by investors, Anthropic's customer acquisition cost (CAC) is less than one-fifth of customer lifetime value (LTV), far below the industry average.
2.2 OpenAI: The 'Curse of Scale' Brought by Its Value Proposition
In sharp contrast to Anthropic stands OpenAI. Its value proposition has been grand from the very beginning: to provide the most powerful, most general artificial intelligence tools to everyone, and to bring AI to every person. This value proposition brought OpenAI unprecedented success. When ChatGPT launched in November 2022, it surpassed 100 million users in two months, creating a growth miracle in internet history. But it is precisely this value proposition that has brought OpenAI an inescapable 'curse of scale.'
PSF Validation: Universal Demand, but Insufficient Willingness to Pay
Demand for general-purpose AI is indeed universal. Ordinary people need a simple, easy-to-use AI assistant for daily tasks; developers need a powerful API platform to build AI applications; enterprises need a general AI tool to boost employee productivity. ChatGPT's explosive growth fully validates this universality of demand. But this value proposition also harbors an inherent contradiction: trying to satisfy everyone's needs results in no single need being met to the extreme. For enterprises that need security, it is not secure enough; for creators who need efficiency, it is not professional enough; for industries that need customization, it is not vertical enough.
PMF Validation: Huge Scale, but Hard to Profit
OpenAI has indeed achieved PMF, and it is one of the largest PMFs in history. According to data from The Information:
Globally, weekly active users exceed 900 million, making it the most-used AI product in the world; ChatGPT Plus paid users exceed 55 million; the API platform has over 10 million developers. With $5.7 billion in Q1 2026 revenue, it remains the highest-revenue large-model company in the world
But the problem is that the quality of this PMF is not high. Consumer-end users have a payment rate below 6%, and their loyalty is extremely low, making them easily replaced by cheaper, better products. More seriously, 70% of computing power is consumed by free users, who generate no revenue at all. This has driven OpenAI's overall gross margin below 20%, trapping it in the vicious cycle of 'the more users, the more losses.'
GTM Fit: The Double-Edged Sword of Free Customer Acquisition
OpenAI's GTM strategy was very successful: through free-product traffic acquisition plus viral spread, it gained a massive user base in an extremely short time. This approach's acquisition cost is nearly zero, the classic growth model of the internet era. But in the large-model era, this model hits a fatal problem: computing costs are too high. Internet products have near-zero marginal cost — the more users, the lower the allocated fixed cost and the higher the profit. But large-model products have high marginal cost — every user query consumes computing power. This means the more users OpenAI has, the more it loses. To solve its profitability problem, OpenAI is being forced to pivot to the B-end, raising the share of enterprise revenue. But compared with Anthropic, its enterprise products have no advantage in safety and compliance, so the pivot is not easy.
III. Track Divergence: Two Other Possibilities in the Large-Model Industry
The competition between Anthropic and OpenAI is just a microcosm of the large-model industry. In fact, the industry has already diverged into four completely independent tracks, each with its own clear value proposition and business loop.
3.1 Kimi (Moonshot AI): The Value Proposition of the Tooling Track
Kimi's value proposition is very clear: to provide the most user-friendly long-text AI processing tool for individual creators and small and medium enterprises. It precisely hits a shortcoming of general-purpose large models: poor long-text processing capability. Groups such as self-media creators, students, lawyers, and consultants often need to process large files and long documents, but general-purpose models can only handle a few thousand characters, failing to meet their needs. Riding the advantage of an ultra-long context window, Kimi quickly won user recognition. According to the '2026 Q1 AI Application Value Ranking' released by QuestMobile, Kimi's domestic monthly active users are about 8.338 million, leading the long-text segment. In March 2026, Moonshot AI's overall ARR surpassed $100 million, with overseas revenue already exceeding domestic — validating the commercial viability of the tooling model. Kimi's GTM strategy also fits its value proposition well: free long-text features for traffic plus user word-of-mouth, with low acquisition cost, a lightweight product, and no need for a complex sales team.
3.2 Zhipu AI: The Value Proposition of the Localization Track
Zhipu AI's value proposition is: to provide the safest, most compliant, and most domestic-chip-compatible large-model services to governments and central and state-owned enterprises. It targets a pain point unique to the Chinese market: the need for local substitution. Governments and central/state-owned enterprises cannot use foreign large models, posing data-security risks; at the same time, most large models cannot run efficiently on domestic chips. Zhipu AI is the earliest domestic large-model company to comprehensively adapt to domestic chips such as Phytium, Kunpeng, and Hygon, and it possesses deep government resources. According to its 2025 annual report, the company achieved full-year revenue of ¥724 million, a year-on-year increase of 131.9%; its MaaS API platform ARR reached ¥1.7 billion, up 60-fold year on year. It has served over 100 central enterprises and 500 local government departments, with a project renewal rate above 85%. Zhipu AI's GTM strategy is: expansion through government relationships plus cooperation with system integrators — although the sales cycle is long and delivery is heavy, customer stability is extremely high.
IV. Core Observation: After Seeing These Four Companies' Development Paths
After seeing these four companies' development paths, one can clearly perceive that the large-model industry has already formed distinctly different business logics. Today, there is no longer such a thing as a 'single correct' business model within the industry. Anthropic went deep on the pure B-end to achieve profitability; OpenAI leveraged a general platform to build scale advantages; Kimi focused on the tooling track for rapid growth; Zhipu rooted itself in the localization market for stable revenue. Each path has its rationale for existing, and each faces its own challenges and limitations. In the early days of the large-model industry, the core competitive dimension in the market centered on technology — parameter scale, computing power, and inference speed were the key metrics for measuring a company's strength. At the time, the industry widely believed that technology was an insurmountable barrier, and whoever could first build the most powerful general model would achieve winner-takes-all. But with the rapid iteration of technology, the gap in base capabilities among top models has significantly narrowed, and most users can hardly perceive obvious differences between models in daily scenarios. When technology is no longer the decisive factor, the center of competition naturally shifts — from competing on a model's 'smartness' to competing on the depth of understanding customer needs. A common misconception in business practice is trying to build a perfect product to satisfy all users' needs. But this pursuit of comprehensiveness often causes a product to lose its core competitiveness. No product can cover all user groups, and no company can occupy the entire market share. A healthy business model is never the pursuit of flawless perfection, but the achievement of precise fit at every link: value proposition matches the customer's core pain point, product capability matches the value proposition, and acquisition path matches product attributes. If any one of these three links is misaligned, no matter how advanced the technology, it is hard to convert into business success. The industry widely pursues growth speed to the extreme — faster user growth, larger market scale, higher valuation — becoming the core goal of many companies. But Anthropic's trajectory offers another possibility. It took three years to accumulate 1,000 enterprise customers, a growth rate far below OpenAI's explosive expansion, yet its business model of high price per customer, high retention, and high gross margin let it achieve single-quarter profitability first, while the industry as a whole was still in its cash-burning phase. Those seemingly fast growth paths often come with unsustainable cost pressure; those seemingly gentle development rhythms may instead build a more solid business foundation. The business world has no ultimate law that applies everywhere. All successes that withstand the test of time are, in essence, the result of choosing a direction that matches one's own capabilities at a specific industry stage and then cultivating it continuously.
