It would be nice if companies could develop models of artificial intelligence (HI) and generate cosmic profits. But the latest research from Barclays reveals a crucial detail: the left side of the plundered sums actually settles in the guts of the dark giants – Amazon AWS, Microsoft Azure and Google GCP.

Profits of III-companies: a closer look.

For every 100 dollars paid by the company that sells the HI-models, 35 to 40 dollars go to pay for the inference fees for the services of the three main providers. From these prices, the worst services will deduct 10-20 dollars of operating income, and their margin can be as high as 35-45%.

The Barclays research, published on September 28, sheds light on where the real value of the world is going..

Increasing profitability of AI laboratories.

It is noted that the influx of paid services for AI laboratories has increased significantly: from a little more than 10% in 2025 to 50-65% and even more in 2026. Adjusted gross margin increased by 30-50 hundred hundred points.

Barclays analysts attribute this increase in profitability to two key factors: corporate clients and the so-called “agentic workflows” (working processes based on agents), which have become a major hit on the market.

At the same time, analysts admit that the real profitability of AI laboratories may be even greater, but may not show deterioration. However, it turns out that with increased competition between advanced models and the increased availability of calculations, this indicator will gradually decrease.

Business structure and margin: different approaches.

To gain a better understanding of the financial dynamics between different AI labs, Barclays modeled two hypothetical scenarios: “Lab A” and “Lab B”.

Laboratory A: 70% of income from API, 30% from subscriptions.

Laboratory B: 80% income from subscriptions, 20% from API.

The API business by its nature has more revenue (prior to renewals), lower subscriptions, and also through the division of expenses for training and recovery of partners, the difference in scoring.

The method of increasing income further enhances this “spurious” effect. Laboratory A generates indirect API income behind the scenes, while Laboratory B uses a pure method, while non-insurance indirect API income is generated by strategic partners.

Barclays is on par with Uber and Lyft: due to the similarity of the main business, different investment methods lead to similar indicators for financial information.

The investigation is ahead: if AI labs begin to publish financial information consistent with Generally Accepted Accounting Standards (GAAP), investors need to carefully understand these data..

Availability of different product lines.

Subscriptions (for example, Claude Code, Codex): Targeted profitability is about 70% – the lowest among the three main types of products. AI laboratories are ready to bear part of the cost of tokens for the sake of increasing the value of investors. These subscriptions are subject to monthly payments and may be subject to exchange rates..

Direct API: The most recent and most profitable business model. Tools for retailers, such as Cursor and Figma, charge a fee depending on the number of valid tokens. The flow rate here is 80%. Increased efficiency of models (fewer tokens for the same task), increased prices for APIs and optimization of infrastructure (quantization, speculative decoding, new generation of calculations) will continue to rise..

Indirect APIs: Provide customers with information similar to a direct API, but the client interaction with the provider of bad services is obtained directly. With the growing share of indirect APIs in the hidden income of AI laboratories, the importance of methods of generating income will further increase the differences among financial figures.

How much do lousy providers charge

Let's take a closer look at the structure of profits from bad providers:.

Pokaznik.

Laboratory A.

Laboratory B.

Income of lousy services per $100 of AI income.

$35.

$41.

Operating profit of a gloomy provider.

$11. 80.

$19. 10.

Operating margin of a bad provider.

34%.

47%.

The large margin of Laboratory B (47%) is determined by the mechanism for dividing income with strategic partners (20% of income with an accumulative limit). Barclays says that this mechanism depends on the margin performance of poor providers. Without it, the profit per token would be the same for both laboratories. It turns out that this mechanism will emerge after 2028.

The additional value for premium providers comes from subscriptions to “agentic” products, which provide access to additional software resources (for example, databases), which can be a significant benefit per unit of income..

Galuz III: transition from “beginning” to “visnovku”.

Barclays predicts a growth rate in AI laboratory income: from 7 billion dollars in 2024 to 137 billion dollars in 2026 and 690 billion dollars in 2028.

Even more disconcerting are the forecasts for river recurring revenue (ARR): close to 200 billion dollars by the end of 2026 and potentially 782 billion dollars by the end of 2028.

The amount spent on model development accounts for nearly 48% of the income of AI laboratories. However, this part is rapidly changing: from 96% in 2024 growth to the predicted 35% in 2027 growth and 30% in 2028 growth.

This means that as the bubbiness of the top grows, the extra bubbiness of the galusa will move up, and the focus will shift from “beginning” to “beginning”.

At the same time, the share of income of bad providers from AI laboratories is decreasing: from 153% in 2024 to 90% in 2026, with a forecasted 73% growth until 2028.

Barclays believes that in the near future, AWS, Azure and GCP will maintain their position in the computing power market. However, by 2028 the situation may change: the emergence of projects with the financing of IT infrastructure may lead to the loss of part of the market by disastrous “triples”, both in the area of the beginning and the renewal, which will lead to the new.

Dzherelo: https://www. ithome. com/0/996/218. htmAI companies: who really makes money on artificial intelligence? read on the HiTech website. Expert.