Recently, during DevDay in October, OpenAI leaked a list of Top 100+customers, which includes over 30 companies with token consumption exceeding 1 trillion.
Since the release of ChatGPT, the industry has been discussing the AI business model. If we calculate based on the cost of GPT-5, with an input to output ratio of 10:1, OpenAI’s top 100 clients have already provided the company with over $100 million in revenue.
Among them, 30 companies with a consumption of 1 trillion tokens each will reach over 60 million US dollars. What is the concept of 1 trillion tokens? If an ordinary person lives for 80 years and writes 1000 words per day, 1 trillion tokens is equivalent to writing 60 million Chinese characters, which is the total amount written by 200000 people in their lifetime; If used in a company scenario, it is equivalent to using GPT year-round in customer service chat, document summarization, and code scenarios, with a continuous 9-year rotation.
The usage of tokens does not necessarily mean a company’s technological level, but it indicates that AI has been deeply integrated into various businesses of the company, and a new technological paradigm is happening. For these 100 companies, big models have become an indispensable infrastructure, just like water and electricity.
This list reveals who is currently making money with AI? How much do you earn? How to earn it?
Among these 30 clients with a consumption of over 1 trillion tokens, we have observed some interesting phenomena and companies, such as AI programming companies and infrastructure providers being major consumers of tokens; We also see some vertical businesses with commercial prospects, such as Harvey for legal AI and Decagon for medical AI communication; Search, conferences, education, video, sales, and graphic design business scenarios are areas with greater potential for AI commercialization.
According to publicly available information, organize
AI native companies are surpassing mature companies
Firstly, among the top 30 companies, the proportion of AI native startups exceeds that of traditional mature companies.
Model as product “allows these AI native startups to deeply embed large models and AI into their business and workflow from the first day of establishment.
For example, Cognition, a unicorn in the field of AI programming, has an AI for coding product called Devin. This company was founded in 2023, and in March of this year, Cognition completed a financing round with a valuation of $4 billion. This year, they also acquired another AI programming company Windsurf, which of course gave the AI programming company enough funds to purchase OpenAI’s API. CEO Scott Wu stated in a podcast that in order to reshape the future of software engineering, Cognition’s interview process takes 6 to 8 hours to build its own Devin – a very AI native interview method.
Another notable company is Genspark, a general agent company co founded by former Baidu “Xiaodu” CEO Eric Jing and former CTO Kay Zhu.
According to The Information on October 31st, Genspark has completed a new round of financing, with investors including Tencent. After the transaction, Genspark’s valuation exceeded $1 billion. According to the statistics of third-party data company Xsignal (Singular Factor), the maximum number of Genspark users does not exceed 5 million, which is not high. Manus’ website and app have over 7 million monthly active users, but Genspark has been able to enter OpenAI’s top 30 customers, indicating that it not only achieves user growth on the C-end, but more importantly, deeply uses AI in its own business. Because Genspark uses models not only OpenAI, but also Claud, Gemini, and others.
In contrast, mature companies tend to use AI in their workflows through modular transformation.
However, we can also see from this list companies that have undergone thorough AI transformation and achieved significant results, such as Notion. After Notion’s transformation, its latest Notion 3.0 version deeply embeds AI into workflows. As a writer, the most practical feature is a natural and smooth triggering method that reduces the threshold for use. For example, by selecting any text, the “Ask AI” and “Refine” functions can pop up to process the content; Tap the space to summon the AI menu.
As of the end of 2024, over 50% of Fortune 500 large enterprises have used Notion to varying degrees internally. More than half of the startups in the YC ecosystem, the largest incubator in the United States, are using Notion to manage information.
To B company ‘eat’ a large number of tokens, AI penetrates into vertical industries
Among OpenAI’s 100 major clients, over 75% of the companies are concentrated in ToB and developers, with no more than 15 companies in ToC.
If further subdivided, some of ToB’s enterprises are developer oriented, such as Cognition and JetBrains; Part of them are enterprise SaaS service providers, such as Salesforce and Zendesk; New generation enterprise tools, such as Ramp and Glean, the enterprise search company.
Among them, Glean is an AI driven enterprise search and knowledge discovery platform, and its Work AI platform indexes and understands document contexts from dozens of products through over 100 SaaS connectors, providing knowledge workers with a Google like search experience. Therefore, behind every API call of Glean, there is a huge enterprise customer network, which makes Glean one of OpenAI’s major customers. The company achieved 1.7 million visits in July this year, with a growth rate of 10.7%.
Among them, there are also some companies that are AI agents targeting vertical industries, such as the legal AI agent – Harvey, which indicates that these AI agents in vertical industries have already run a data flywheel through deep penetration and data accumulation in vertical industries. Harvey achieved $30 million in revenue within two years of its establishment, while the best B2B SaaS product in the past 10 years achieved an average revenue of $1 million in two years.
The characteristics of these vertical AI agents are that the industry has high-value, high complexity scenarios, long task chains, and short feedback loops. This also provides some inspiration for AI entrepreneurship nowadays, which requires finding vertical fields with high value, complexity, and professionalism to directly address industry pain points.
The C-end scenarios are concentrated in high-frequency, large user base, and clear payment models
Among the top 30 customers, only Genspark, a general AI agent company, Perplexity, an AI search company, Duolingo, a language learning platform, WHOOP, a health wear enterprise, and Read AI, an AI conference tool, provide products for the C-end. These C-end scenarios have a common feature of high-frequency usage, large scene space, and clear paid subscription scenarios.
Duolingo can be said to be one of the education companies with the fastest and most thorough AI transformation. They have long integrated their business with large models, such as using AI (such as GPT-4) to generate exercises and course content on a large scale; In terms of personalized learning, Duolingo dynamically adjusts the learning difficulty and content recommendations for each user through its self-developed AI algorithm “Birdbrain”; In 2024, Duolingo released Duolingo Max, which features over 200 scenes and dialogues for AI role-playing; AI graphic and text interaction functions. These multimodal features have also made Duolingo a major consumer of tokens.
Another high-frequency scenario is conferences. Read AI is an AI conference content summary platform that provides conference summaries, text recordings, and conference replays to help users efficiently view meetings, emails, and messages. Use multimodal AI to automatically merge meetings, emails, and chats into a searchable and executable knowledge base. As of 2025, the number of Read AI users has exceeded one million, with an increase of 720% in active users over the past 12 months.
Model dependency anxiety in start-up companies
Of course, behind the top customer list, these companies spend a lot of money on OpenAI every year. For example, the AI search startup Perplexity paid $8 million to Anthropic and OpenAI in 2024, which accounted for almost a quarter of its revenue.
What’s even more frightening is that many companies here not only access OpenAI’s API, but also include Google’s Gemini and Claude, which will cause these companies to continuously raise large amounts of funding or pass on costs to users.
It’s no wonder that during a conversation with Paddy Kosgrave, CEO and founder of Web Summit, I mentioned that when Deep Seek first came out, almost all companies in San Francisco were using it throughout February and March because of its lower cost.
Another layer of anxiety lies in the degree of overlap with OpenAI’s business – if OpenAI earns money and the business is swallowed up, it will put the startup in a dangerous situation. In theory, Open AI will do everything. For example, recently this “behemoth” launched an AI browser, which is a deep crisis for the general agent platform. Liu Zhen, founder and CEO of Xsignal (Singular Factor), told Huxian that OpenAI will incorporate the “general agent” function into its ecosystem, which means that startups that independently build general agent platforms will face uncertain but likely tragic outcomes, similar to the outcome of independent search AI applications two years ago, which will be harmed by the siphon effect of ChatGPT.
Liu Zhen gave several judgments on which fields OpenAI will enter, providing some reference for startups. Firstly, whether it can significantly increase user stickiness or business revenue; Secondly, can the company control more user touchpoints (browsers, agents, OS); Thirdly, whether the market is large enough and high-frequency; Fourth, whether the ROI of computing power and revenue is reasonable; Fifth, whether external dependencies (partners, regulators, copyrights) are controllable.
No comment content available at the moment