Microsoft and Amazon have surged by more than 20% in three days, is the AI sector starting to trade on the "profit rotation" logic?

The underlying logic of the capital market has always been straightforward: where profits flow, capital will surge there.

At different stages of industrial development, profits will always be realized in the most scarce and most bargaining-powerful links, and the profit-seeking nature of capital is destined to drive it to migrate across all links of the industrial chain.

In the past year and a half, the right to profit realization in the AI industrial chain has been in the hands of the "shovel sellers": NVIDIA, TSMC, ASML, SK Hynix. Whoever sells chips, equipment, or HBM sees their stock price rise.

However, hidden worries lie beneath this prosperity: except for a few links with high barriers such as NVIDIA's GPUs and TSMC's advanced processes, most upstream hardware will eventually face the dilemma of unsustainable profits due to capacity expansion and intensifying competition. The maximum drawdown of SK Hynix's stock price in this round has directly been cut by half, which is exactly caused by similar concerns.

Amid the fluctuations of the AI market, cloud vendors, the "shovel users" closest to profits, have become the best-performing AI sector recently. Microsoft and Amazon rose by 25% and 18% respectively in the three trading days after their earnings releases.

The rise in cloud vendors' stock prices is firstly because they have begun to restrain their investment in large models that are fiercely competitive and have not yet generated profit returns, which meets the current cautious preference of the capital market.

The more core logic is that the revenue, profit margin and Backlog of cloud vendors are rising simultaneously, indicating that early capital expenditures are accelerating to be converted into current revenue, operating profits and visibility of future revenue.

The rhythm of cloud vendors' profit realization also directly determines the trading rhythm of the "profit redistribution" of AI investment. Although the specific time point of this inflection point is still uncertain, under the constraint of doubts about the sustainability of upstream hardware profits, cloud vendors, as the unavoidable infrastructure charging layer of the AI industry, will become the key direction for the market to focus on in the future.

/ 01 / Rising Revenue, Profit and Backlog Ease Monetization Anxiety

Since May, the market has raised doubts about cloud vendors' capital expenditures, and Goldman Sachs and Morgan Stanley have released multiple research reports directly pointing out that cloud vendors' capital expenditures are unsustainable.

This doubt extended to Google's earnings report. On July 23, Google's stock price plummeted by 7% after releasing its earnings report. The core negative factor is that against the background that Google's quarterly free cash flow turned negative for the first time, it still chose to raise its full-year capital expenditure guidance, which further amplified the market's concerns about cash burning.

However, the anxiety was relieved after just one week. Microsoft and Amazon successively released their earnings reports on the 29th and 30th, with their stock prices rising by 15.5% and 9.5% in a single day respectively. There are two key supports behind this:

First, the managements of the two companies took the initiative to downplay the unrestricted involution investment in large models at the earnings conference call, easing the market's concerns about cash burning; the more core signal is that the commercial return of AI investment has begun to accelerate: the earnings reports of the "three major clouds" cross-verify that cloud vendors are gradually running through the business closed loop of "capital expenditure - computing power supply - cloud revenue - profit growth".

In the second quarter, AWS's cloud business revenue increased by 37% year-on-year, hitting a new high in 18 quarters. Azure's cloud computing business revenue increased by 43% year-on-year, higher than analysts' expectation of 39.98%; Google Cloud's cloud service revenue soared by 82%, also far exceeding market expectations.

The improvement of profit margin is also obvious. AWS's cloud operating profit margin is 39.4%, a year-on-year increase of 6.3 percentage points, and Google Cloud's profit margin has increased from 20.7% to 35.6%.

What can better illustrate the long-term trend is the growth of backlog orders.

Microsoft's commercial RPO, AWS Backlog and Google Cloud Backlog (RPO/Backlog = the total amount of all signed contracts where customers have committed to pay and not yet amortized and recognized) reached year-on-year growth rates of 84%, 154% and 385% respectively, significantly higher than their respective cloud revenue growth rates. This means that the three major clouds will accelerate their performance growth in the future, which is the core leading indicator for this round of sharp rise in cloud stocks.

The simultaneous rise of comprehensive revenue, profit margin and backlog orders indicates that early capital expenditures are accelerating to be converted into current revenue, operating profits and visibility of future revenue.

Two major factors, namely demand structure and supply upgrading, are driving cloud vendors to accelerate into the AI return period.

From the perspective of demand structure, AI demand has spread from the training demand of a few cutting-edge model companies such as OpenAI and Anthropic to a wider range of enterprise customers, reasoning and application scenarios, and simultaneously drives the consumption of traditional cloud resources such as computing, storage and network.

This has been verified by Microsoft at the conference call: all the new RPO reserve orders in this quarter come from industrial customers other than OpenAI and Anthropic.

While the demand scenarios are increasing, the computing power supply of cloud vendors is also being optimized. Taking AWS as an example, its AI-related revenue is divided into two major categories: AI IaaS (underlying computing power leasing) and Bedrock (TaaS/distribution mode, Model as a Service). The proportion of Bedrock in AWS's AI revenue has increased from 9% last year to 37% this year.

This structural change shows that cloud vendors' AI monetization mode is extending from underlying computing power leasing to model distribution, application monetization and other fields. This stock price growth driven by performance growth also indicates that the investment logic of cloud vendors has changed.

/ 02 / The Realization Rhythm of ROIC Determines the Medium-term Trend

In the past two years, the market has rewarded "who spends the most on AI". Under the initial FOMO sentiment, every dollar spent by cloud vendors can increase their market value by 2 dollars. The market can ignore temporary returns and profits, and as long as you invest, it means you have potential.

However, with hundreds of billions of dollars in capital expenditures invested and the free cash flow of giants turning negative, the investment logic of cloud vendors has begun to shift from "buying grand narratives" to "buying realization evidence", that is, the current market's focus is accelerating to shift to the realization stage of ROIC.

ROIC refers to Return on Invested Capital. Its calculation formula is Net Operating Profit After Tax divided by Total Invested Capital, which is used to measure: for every 1 dollar invested in computing power infrastructure, how much after-tax profit can be stably earned every year. It is a core indicator to judge whether a heavy-asset business is worth continuous investment.

Following the market's doubts about the realization of AI investment, Amazon, whose free cash flow has turned negative, also explained the ROIC doubts at the conference call: the payback period of its computing power investment is less than 3 years. According to the current 5-year depreciation period of core hardware such as servers, the investment can be recovered in just a little more than half of the time, and huge profits can be brought in the remaining more than two years.

Amazon's statement is not much different from Morgan Stanley's measured ROIC data: the stable ROIC of cloud vendors' pure GPU leasing is about 31%, which is the return of cloud vendors' basic computing power business. The ROIC of cloud vendors' self-owned computing power + model platform is 46%-50%.

Compared with the 10%~18% ROIC of mature businesses in traditional clouds and data centers, the return on investment of giants in AI is high enough.

A high ROIC means a good business model, but the question is whether the investment is a little excessive? Amazon also explained this doubt: the company calculates the input-output ratio based on the confirmed orders in hand, and once the production capacity is released, the contract amount can be converted into recognized revenue.

Following Amazon's logic, the total unfulfilled orders of Amazon, Google, Microsoft and Oracle amount to about 2.23 trillion US dollars, while the total capital expenditure of the four major clouds in 2026 is about 725 billion US dollars.

At first glance, the capital expenditure seems to be fully covered by the potential revenue of the orders in hand, but it should be noted that unfulfilled orders are essentially service commitments that have been signed but not yet fully counted as revenue. The "rigidity" of these contracts is different:

Some are hard minimum purchase volumes, where customers must purchase a specified quantity; some are cancellable contracts, but customers need to pay compensation if they exit in advance; some have prices and configurations that can still be renegotiated, and the amount is still under continuous negotiation; others are where customers bring their own hardware and cloud vendors help operate, and the revenue is only the hosting service fee. The four types of contracts are mixed in the same total, but their value is completely different.

The key caliber to judge the value of orders is how much revenue can be recognized in the near future.

Among them, Amazon and Microsoft have the fastest revenue recognition, which can recognize 40% and 30% of the unfulfilled order revenue respectively within the next 12 months. More than half of Google's backlog orders will be realized in the next 24 months. Only more than 10% of Oracle's backlog orders can be converted into revenue within 12 months.

The difference in the value of unfulfilled orders and ROIC may lead to differentiation in the subsequent performance of cloud vendors in the capital market.

/ 03 / AI Starts Trading "Profit Redistribution"

Since July, the US stock AI market has fluctuated back and forth among upstream hard technology, midstream computing power and downstream applications. To judge whether the main line of AI has changed, we need to track the profit realization rhythm of the industrial chain.

At different stages of industrial development, profits will continue to concentrate in the most scarce and most bargaining-powerful links, and capital in the capital market chases profits all the way, flowing from one link to another, and will always flow into the industry with the highest profit in the AI industrial chain.

In previous years, the AI profit transmission chain was mainly concentrated in the upstream AI hardware and AI infrastructure industries. From NVIDIA GPUs to core hardware such as storage and optical modules in the cabinet, driven by the contradiction between supply and demand, the profit margins of relevant links rose sharply, forming a strong market trend of "standing in the light, walking in the chip". At the same time, upstream supporting tracks such as AI power, computing power land and liquid cooling also ushered in simultaneous outbreaks.

However, the problem is that except for NVIDIA's AI chips and TSMC's advanced processes which have long-term high barriers and their profits will not be diverted, most of the upstream hardware including storage will experience cyclical profit decline due to capacity expansion and intensifying competition. Therefore, the maximum drawdown of SK Hynix, Samsung and Micron in this round reached 40%—58%.

Under the circumstance that the sustainability of upstream hardware profits is in doubt, the market has begun to trade the "profit redistribution of the AI industrial chain".

Michael Wilson, Chief US Strategist at Morgan Stanley, reminded investors in his weekly report in July to reduce their overweight position in the semiconductor sector and turn their attention to hyper-scale cloud service providers. He believes that AI investors are shifting from chasing upstream hardware to focusing on the commercial returns of downstream cloud services. In other words, institutions believe that cloud vendors are relatively certain in the profit realization of the middle and lower reaches of the industrial chain.

The logic is not difficult to understand: when upstream hardware is no longer scarce and application scenarios are gradually implemented, more profits in the AI industry will be transferred to midstream cloud vendors, model companies and downstream application links.

Compared with the uncertainty of the diversified downstream applications, midstream cloud vendors and model companies belong to the unavoidable infrastructure charging layer, and different from the upstream hardware mode of "one-time sale", they can achieve continuous charging.

In the comparison between cloud vendors and model companies, the model industry is still in the stage of "high growth but negative profit", while cloud vendors, with a clear competitive pattern and mature business model, have already entered a benign stage of "high growth and rising profit margin".

From the current trend, cloud vendors can even squeeze the profits of the model layer upward. AWS's Bedrock model is a typical example: it does not independently develop large models, but earns higher profits than model vendors by distributing models such as Claude. Data shows that the EBIT profit margin of AWS's TaaS/distribution mode can reach up to 55%, while the average gross profit margin of most model vendors is only 30%.

Although the specific time point for the switch of the AI investment main line is not yet clear, the profit-seeking nature of capital will not change, and the next link of profit realization in the AI industrial chain will eventually become the inevitable choice of the market in the next stage.

Disclaimer: This (report) is written based on publicly available information or information provided by interviewees, but Dongde Finance and the author of the article do not guarantee the completeness and accuracy of such information. Under no circumstances shall the information or opinions expressed in this (report) constitute investment advice to anyone.

This article is from the WeChat official account "Dongde Finance", author: Dongdejun, authorized for release by 36Kr.