Summary of Key Points
This article, based on a Barclays report, dissectes the profit-making mechanisms of leading AI companies such as Anthropic and OpenAI. The profit margins from their reasoning activities (i.e., the profits generated by using models to answer questions and complete tasks) are expected to soar from 10%-20% in 2025 to 50%-65% in 2026. However, for every 100 yuan earned, 35-40 yuan is taken by cloud service providers like AWS and Azure. Additionally, the revenue figures of AI companies are not entirely reliable due to different accounting methods. Around 2028, as AI companies start building their own infrastructure, their reliance on cloud providers will decrease, and the focus of competition will shift from model capabilities to the right to distribute those profits.
1. AI Company Revenue Figures May Be Inflated: Don’t Be Misled by the Numbers
Many people think AI companies are making huge profits based on their revenue figures (such as Annual Recurring Revenue, ARR). But Barclays argues that simply comparing numbers is meaningless because of the different accounting methods used. For example, if a user spends 100 yuan on an AI service through Azure, an AI company using the “gross method” would count the entire 100 yuan as revenue and deduct the 30 yuan paid to Azure as a cost. In contrast, using the “net method,” only the 70 yuan received by the AI company would be recorded as revenue. Therefore, just because Company A has ARR of 20 billion and Company B has 15 billion, it doesn’t necessarily mean Company A is more profitable—Company A might just be using the gross method, including all the transactions that pass through its accounts. What really matters is whether the profits from reasoning activities are sufficient to cover the costs of training new models, which can be extremely expensive. Only if they can do so can AI companies truly break away from their “money-burning” phase.
2. Rapid Growth in AI Company Profit Margins, but Most Goes to Cloud Providers
By 2026, AI companies’ profit margins from reasoning activities are expected to reach 50%-65% (with Anthropic at 65% and OpenAI at 48%), a significant increase from 2025. There are two main reasons for this:
1. More Efficient Models: AI models require fewer computing resources to complete the same tasks.
2. Increasing Number of Corporate Customers: Companies are willing to pay more for API services or enterprise versions, and their usage is stable.
However, not all of this profit stays with the AI companies. For every 100 yuan earned, 35-40 yuan goes to cloud providers for computing power. Moreover, cloud providers earn additional revenue from other services such as databases, storage, and networking, essentially taking a triple profit. As a result, AI companies are essentially “working for” these cloud providers.
3. 2028: A Turning Point for AI Companies to Reduce Dependence on Cloud Providers
AI companies will not always remain dependent on cloud providers. Barclays predicts that the market share of traditional cloud providers will decline starting in 2028 for two reasons:
1. Sufficient Profit Margins: Profit margins from reasoning activities will grow enough to cover the costs of model training, allowing AI companies to fund the construction of their own infrastructure.
2. Economical Benefits of Scale: When AI companies generate hundreds of billions in revenue and have stable computing needs, building their own data centers, purchasing GPUs, and securing power supply becomes more cost-effective than renting from cloud providers. For instance, Anthropic is building its Stargate infrastructure to reclaim the profits that were previously taken by cloud providers.
4. The Next Phase of AI Competition: The Battle for Profit Distribution
In the past, AI companies competed on who had the smarter models. Now, as the gap in model capabilities narrows, the competition focuses on who can keep more of the profits for themselves. The two key turning points in this competition are:
1. When profit margins from reasoning activities will be able to consistently cover training costs.
2. When AI companies will move most of their computing needs from cloud providers to their own infrastructure.
The company that achieves these goals first will gain a dominant position in profit distribution. After all, the key to making money is not how much you earn, but how much you can keep for yourself.
This report essentially suggests that the money in the AI industry hasn’t disappeared; it’s just temporarily in the hands of cloud providers. In the coming years, the “profit battle” between AI companies and cloud providers will intensify.