Summary of Key Points
A recent report from Barclays breaks down the financial dynamics within the AI industry: For every $100 earned by AI model companies, approximately $35 to $40 is paid in fees to the three major cloud providers—Amazon AWS, Microsoft Azure, and Google Cloud—for computing resources. These cloud companies generate profits of $10 to $20 on this amount, with profit margins ranging from 35% to 45%. The profit margins of AI research laboratories are soaring, expected to reach 50% to 65% by 2026, but are expected to decline due to increased competition and greater availability of computing power. Different business models (API vs subscription) and accounting methods result in significant differences in laboratory profits. After 2028, AI laboratories' own computing capabilities are likely to gradually erode the market share of the major cloud providers.
Detailed Analysis
1. How much do cloud providers earn from AI companies' revenue?
Nearly 40% of AI model companies' revenue goes towards paying for computing resources from cloud providers. For example, if an AI company earns $100, $35 to $40 is spent on computing fees. After deducting costs such as servers and electricity, the cloud provider makes a profit of $10 to $20, with a profit margin of 35% to 45%—which is higher than in many traditional industries.
The amount paid by different laboratories varies:
- Laboratory A (70% revenue from APIs, 30% from subscriptions): Pays the cloud $35, and the cloud makes a profit of $11.8, with a profit margin of 34%.
- Laboratory B (80% from subscriptions, 20% from APIs): Pays the cloud $41, and the cloud makes a profit of $19.1, with a profit margin of 47%.
The difference is due to Laboratory B's partnership agreement, which accounts for 20% of its revenue; however, this represents a higher nominal profit. Excluding this partnership, the actual profit per AI task processed by the cloud provider is similar to that of Laboratory A, and this partnership will be discontinued after 2028.
2. Why have AI laboratory profit margins soared?
By 2026, AI laboratory profit margins are expected to reach 50% to 65%, a significant increase from the low double-digit figures of 2025 (10% to 20%). The main reasons are:
- Compulsory customer demand: Enterprise customers have made AI models and intelligent agents an essential part of their operations (similar to how companies must use Office software), so laboratories no longer need to compete on price to gain market share, allowing them to increase profits.
- Cost reductions and revenue increases: Improvements in model efficiency (using fewer tokens for the same task), higher API prices, and advancements in computing technology (e.g., more efficient servers) have expanded profit margins.
However, these margins are expected to stabilize as more companies enter the AI market, increasing competition and reducing the cost of computing power.
3. Which business model is more profitable: API or subscription?
Direct API sales are the most profitable, while subscription models generate lower margins:
- Direct API sales: Profit margins exceed 80%. Developers pay based on the number of tokens used, allowing laboratories to earn more through efficiency improvements and price increases.
- Subscription models: Profit margins are around 70%. For example, Anthropic’s Claude Code offers a fixed monthly fee, but laboratories may subsidize token costs to retain customers, reducing their profit margins.
- Indirect API sales: The experience is similar to direct API sales, but the payment is first made to the cloud provider before being distributed to the laboratory. Different accounting methods (e.g., gross vs net revenue) can lead to significant differences in reported profits.
4. Are cloud providers' profits truly high, or are they just high on paper?
Although Laboratory B pays more to the cloud provider, this figure is inflated due to a partnership agreement. Excluding the partnership, the actual profit per token processed by the cloud provider is the same as that of Laboratory A. Additionally, subscriptions from intelligent agents generate additional revenue for cloud providers, as these services (e.g., storage and database access) generate additional profits.
5. Will the market share of the major cloud providers in AI computing power shrink?
Yes! After 2028, the market share of the major cloud providers in AI computing power is expected to decline:
- AI laboratories are building their own computing facilities, which will reduce their reliance on cloud providers.
- Training costs are also decreasing: In 2024, it cost $0.96 to train an AI model for every $1 earned; by 2028, this cost will drop to $0.3, indicating that laboratories will become more profitable and capable of building their own computing capabilities.
Both the training and inference markets will see the major cloud providers lose some of their market share.
In Summary
Currently, most of the profits in the AI industry go to cloud providers. However, as AI laboratories improve their profitability and develop their own computing capabilities, the cloud providers' share of the market will gradually shrink, while laboratory profits will first increase and then stabilize. Different business models and accounting methods can make profit figures appear more or less accurate.