Many people fail to understand where Anthropic's success truly lies.

Original Content from Xinmou · Author | Tang Ning

Silicon Valley always has a short memory.

Just a year ago, the industry had nearly concluded that the war for general-purpose AI was over. With 900 million weekly active users, Microsoft's seemingly unlimited resources, and the mythical aura surrounding Sam Altman, everyone was discussing how to divide the remaining pie. No one paid much attention to Anthropic.

Founded by 'OpenAI defectors,' the company was pinned to the pillar of shame and labeled as 'safety fanatics,' 'slow followers,' and 'perpetual runners-up living in the shadow of their former employer.'

That was until May this year. With a pre-investment valuation of $900 billion, $30 billion in new funding, projected Q2 revenue of $10.9 billion (exceeding last year's total in a single quarter), and $559 million in net profit. The irony is even more pronounced when considering OpenAI, once deemed the 'winner,' remains unprofitable at scale, with ChatGPT's growth ceiling and advertising monetization posing persistent challenges.

Eighteen months ago, Anthropic's annualized revenue was less than $1 billion. In under two years, it has grown more than tenfold. Such growth is unprecedented not just in the AI industry but in the entire history of technology.

Many attribute their success to luck, having caught the enterprise AI wave. But I believe luck is the least significant factor. What truly brought them here were three choices that seemed foolish at the time but now appear brilliantly correct.

01 The Collapse of the Traffic Myth: AI Is Not the Internet

OpenAI's predicament was sealed the moment it chose to apply internet logic to AI.

The debut of ChatGPT in late 2022 cast a powerful spell over the industry. Everyone believed AI's future lay in consumer markets, traffic, and internet-style monetization—acquiring massive users through free products and monetizing via ads and premium services. Thus, a frenzied arms race began: comparing model sizes, context window lengths, product launch timings, and user growth rates. No one cared about costs, profitability, or what problems these products actually solved.

Because in internet logic, traffic equals money.

They forgot a fundamental truth: AI is not the internet. Internet marginal costs are zero—a website costs the same whether used by 10,000 or 100 million people. But AI isn't like that. Every conversation, every generation consumes real computing power and incurs real costs. This created an absurd situation: more users meant greater losses; heavier usage meant steeper losses.

OpenAI's consumer subscription model epitomized this contradiction. At $20/month for unlimited use, 10% of heavy users consumed 90% of computing resources while paying only 10% of the fees. Estimates suggest a heavy ChatGPT Plus user costs OpenAI over $200/month. This created an unsolvable deadlock: raising prices drives users away; keeping prices low ensures ongoing losses.

From the start, Anthropic avoided consumer chatbots, idle chat users, and the pursuit of flashy monthly active user numbers. Instead, it targeted enterprise markets—people truly willing to pay for AI.

02 While Others Raced for Speed, They Competed for Safety

Anthropic's first critical choice was making AI safety its foundation rather than an afterthought.

When Dario Amodei and his team left OpenAI in 2021, nearly everyone thought they were crazy. They spurned Microsoft's $10 billion partnership and the trillion-dollar commercialization potential of GPT-3 to pursue something as abstract as AI safety.

Back then, the AI industry was flooring the accelerator. Comparisons focused on model sizes, context windows, image realism, and video length. Safety? That could wait until after making money.

Anthropic disagreed. They holed up in an unassuming office for two years. While others rushed products to market, they wrote 75 rules. While others burned cash on computing power, they taught models to self-critique and self-correct.

This became Constitutional AI.

I consider this a landmark invention in AI history. Previous alignment techniques essentially taught models to 'say what users want to hear,' even if false. Constitutional AI teaches models to 'say what's correct,' even if unpopular.

This represents a fundamental ideological split. For three years, the industry fixated on model ceilings while ignoring floors. But for enterprises, floors matter more than ceilings.

A student who scores 90 but occasionally 0 is far less desirable than one who consistently scores 80. In finance, healthcare, and law, a single error can cause billions in losses.

An unwritten industry consensus I strongly agree with: 'In consumer AI, performance is king; in enterprise AI, safety is 1, everything else is 0.'

Why did JPMorgan Chase sign a $2 billion exclusive deal with Anthropic? Not because Claude is smarter than GPT, but because Claude hasn't had a single major security incident or data leak in two years. For a bank managing trillions, this outweighs any efficiency gains.

More tellingly, Anthropic publicly rejected a $200 million U.S. Department of Defense contract, citing 'not using AI for lethal weapons.' This seemingly costly decision became their best marketing tool. Fortune 500 companies flocked to them, knowing a company that turns down the U.S. military won't leak their trade secrets.

03 The Model the Internet Disdains Won in the AI Era

Their second choice was adopting the internet industry's most derided, simplest, and dumbest business model: pay-as-you-go.

Anthropic's model is straightforward: you pay for the tokens you use. No free tier, no subscriptions, no premium services, no bundling.

Many internet veterans scoffed at this primitive approach. After two decades of free+ads and subscription+premium models, who does business like this?

Yet this primitive model made Anthropic the world's first profitable general-purpose AI company.

Why? Because AI's fundamental logic differs completely from the internet's.

Internet success relies on zero marginal costs. Software costs the same whether sold to 100 or 1 million users, allowing free usage with revenue from ads or premium features.

AI isn't like that. Every inference, every content generation consumes real computing power and incurs real costs. More usage means higher costs, fundamentally undermining traditional internet business models.

Geoffrey Parker, author of 'Platform Revolution,' made an accurate observation: 'Pay-as-you-go is AI's native business model. It perfectly matches AI's cost structure, aligning costs and revenues precisely.'

More importantly, this model creates tremendous customer stickiness. A traditional industry CIO told me they now use Claude's API in hundreds of internal applications developed painstakingly by engineers. Switching models would take at least two years with incalculable costs.

Consumer users can switch from ChatGPT to Claude instantly, but enterprises cannot. This explains Anthropic's consistently high customer retention rates.

04 Don't Be a Pawn for Giants—Sell to Them

Their third choice was avoiding dependence on any single cloud giant, instead making the giants compete for their business.

OpenAI tied itself completely to Microsoft's Azure for all computing needs. This solved early capacity issues but sacrificed independence, making them part of Microsoft's ecosystem. Microsoft could cut off their computing supply anytime or integrate their technology into its own products.

Anthropic took a different path: multi-cloud neutrality. They simultaneously partnered with Amazon, Google, Microsoft, and NVIDIA, sourcing computing power from multiple cloud providers.

Initially, everyone thought this foolish. Distributed computing increased management complexity and prevented securing maximum discounts from single providers. Why complicate things?

Subsequent events proved everyone wrong.

In April, Amazon announced increasing its investment in Anthropic to $25 billion. Three days later, Google quickly followed with $10 billion in cash and up to $40 billion total.

Two of the world's largest cloud providers competing to fund the same AI company—unprecedented in tech history.

Tech observer Ben Thompson wrote in his Stratechery column: 'Anthropic is the first AI company to truly grasp leverage over cloud giants. It didn't become anyone's pawn; instead, it made the giants bid against each other for its business.'

Why? Because in today's AI industry, the scarcest resource isn't technology or capital but computing power. And Anthropic has become the world's largest and fastest-growing consumer of computing resources.

If Amazon doesn't invest, Anthropic will give all orders to Google; if Google doesn't invest, orders go to Amazon. They'd rather mutually restrain each other than let one dominate this strategic asset.

The result: Anthropic secured not just massive computing resources but priority supply. Abundant and cheap computing power lets them maintain rapid growth while achieving profitability first.

Many say Anthropic succeeded due to luck in catching the enterprise AI wave. But I believe luck is only part of the story. More importantly, they stayed sober when others went mad; chose difficult paths when others took shortcuts; prioritized doing things right over making quick money.

Original content from Xinmou. For reprint authorization or business cooperation, please contact us.

— END —