AI startup with $80 billion ARR, 90% taken by 2 companies

Author | Huálín Wǔwáng

Editor | Jìng Yǔ

AI, as the hottest track in the past two years, has attracted countless entrepreneurs to enter and realize the "AGI" dream. However, even in such a crowded track, the concentration of investment and revenue is even more pronounced than during the internet boom.

According to the latest analysis from The Information, the annualized revenue of 34 leading AI startups has already reached about $80 billion, a 112% increase compared to six months ago.

This number sounds prosperous, and the entire track is racing forward. But upon closer inspection, you'll find a chilling data point:

OpenAI and Anthropic together took 89% of this $80 billion.

The remaining 32 companies share the other 11%.

Let's first look at the real scale behind these two figures.

Anthropic's annualized revenue has exceeded $30 billion. OpenAI's self-reported figures are between $24 billion and $25 billion. Combined, they amount to roughly $55 billion in annual scale.

These are two "startups" founded less than ten years ago, and this is "annualized revenue," not a valuation bubble—it's real cash flowing into accounts at this speed.

Even more noteworthy are the growth logic of these two companies.

OpenAI's revenue engine mainly relies on ChatGPT's consumer subscription users. From free to Plus, Team, Enterprise, climbing step by step. This path is fast, but also hits a ceiling—consumer subscription willingness and paying capacity are limited, and this market heavily depends on product experience perception. Once competitors launch better products, user migration costs are almost zero.

Anthropic is taking a different route. From day one, Dario Amodei has focused on enterprise clients and API access as the core battleground. Claude isn't meant to be a chatbot that users like, but to become a foundational component in enterprise software stacks. This strategy has much stronger stickiness—once a company deeply integrates Claude's API into its products and workflows, migration costs become extremely high.

In April this year, a number officially confirmed the effectiveness of this strategy: Anthropic's market share in the US enterprise sector surpassed OpenAI for the first time, reaching 34.4%. Meanwhile, mid-2023, this figure was less than 1%.

From 1% to 34%, Anthropic took less than two years.

01 Other AI companies live in the gaps

Of course, the AI startup market isn't just OpenAI and Anthropic. Mistral, Cohere, AI21 Labs, Perplexity, Character.AI... and many other companies that have raised large amounts of funding and recruited top talent, each with their own stories and tactics.

But 11% market share is to be divided among 32 companies, averaging about 0.34% each.

This doesn't mean these companies lack value. Perplexity has built a real user base in the AI search niche; Mistral has established a unique moat in Europe through open-source strategies; Cohere focuses on enterprise-level private deployment, serving high-security financial and medical institutions. These are real businesses with real revenue.

But a harsh reality is emerging: as resources, talent, and computing power procurement increasingly concentrate among the top players, the survival space for mid-tier companies will be systematically squeezed.

Top engineers will prioritize joining OpenAI or Anthropic; cloud giants will offer more favorable computing agreements to leading companies; enterprise procurement decisions increasingly default to "using ChatGPT" or "using Claude," requiring more time to explain and persuade for other options.

This is a self-reinforcing flywheel: higher revenue → greater compute investment → stronger models → higher revenue.

A Silicon Valley AI entrepreneur once said that "building foundational large models is essentially a capital-consuming war—you need enough money to survive to the next funding round, then the next, until the market structure stabilizes." Based on today's data, this consumption war is nearing its end.

02 "Oligarchs" are not at ease either

Of course, 89% ARR share doesn't mean OpenAI and Anthropic are worry-free.

In the past two weeks, OpenAI has faced several dizzying situations simultaneously.

Sam Altman testified in court, revealing that Musk once demanded 90% ownership of OpenAI. The outcome of this lawsuit will directly impact OpenAI's governance structure and its transition from a nonprofit to a for-profit entity.

Meanwhile, negotiations between OpenAI and Apple over the Siri partnership have hit serious disagreements, with reports suggesting OpenAI is preparing legal action. This is a subtle signal—collaboration with Apple was once a crucial channel for OpenAI to reach hundreds of millions of iPhone users. If this partnership breaks down, the impact could be significant.

On the product front, OpenAI's pace remains rapid. On May 11, it launched OpenAI Deployment Company to help enterprises build around AI; on May 15, it released a limited preview of GPT-5.5-Cyber aimed at cybersecurity professionals; now, free users can also see inline images in conversations.

The density of product releases and business disputes is almost rising in tandem.

This is a typical sign of a company entering the "ruler's anxiety" stage. When you're already the market leader, you must deal with technical pressure from followers, business friction with partners, commercialization expectations from investors, and scrutiny from regulators and courts—all at once, draining your attention.

In contrast, Anthropic's external image is much "quieter." No sensational lawsuits, no dramatic court appearances by CEOs. Led by Dario Amodei and Daniela Amodei, the team focuses on expanding enterprise clients and iterating model capabilities, gradually eroding OpenAI's enterprise market share.

Of course, "quiet" doesn't mean no pressure. Behind Anthropic is Amazon's hundreds-of-billions-dollar-level investment, and such capital support comes with equally large expectations for commercialization.

03 After 89%, where is the industry headed?

Such concentration—89%—is not uncommon historically.

Smartphone operating systems, Android and iOS, have long exceeded 99%.

Search engines, Google alone takes over 90%.

Cloud computing, AWS, Azure, and GCP combined account for over 65%.

These precedents show that the infrastructure industry tends to form oligopolies naturally. The reasons are simple: scale effects, network effects, switching costs—these three forces together create an almost insurmountable moat.

Large AI models, especially general-purpose large models, also possess these three features. Therefore, today’s 89% concentration might not be the end, but a middle state—ultimately, the pattern could become even more concentrated.

But there is a variable not seen in historical precedents—the speed of AI capability advancement is much faster than the iteration of operating systems, search engines, or cloud computing technologies.

Anthropic's growth from 1% in 2023 to 34% today is fundamentally due to the qualitative leap in the capabilities of the Claude series models. If a currently obscure team trains a model tomorrow that significantly surpasses GPT-5 and Claude in a key dimension, the market share balance could tilt again at any moment.

For the 32 companies living in the 11% slice, the most realistic strategy might not be direct confrontation, but rather finding those vertical scenarios where "general large models are not enough, specialized models are more effective," and digging deep. Legal documents, medical imaging, code security audits, industrial quality inspection—these fields have strong professional barriers that can't be solved simply by fine-tuning GPT-5.

Industry concentration does not mean opportunity disappearance. It only means that the form of opportunity has shifted from "building a better general AI" to "creating an irreplaceable specialized AI in a specific field."

Two mountains are already standing there. Smart people are not thinking about how to move them, but about finding that fertile land at the foot of the mountains that others haven't discovered yet.