©Understand Original/Production

The underlying logic of the capital market has always been straightforward: capital flows to where profits go.

At different stages of industrial development, profits tend to concentrate in the most scarce and price-setting segments. The profit-seeking nature of capital ensures its migration across various segments of the industry chain.

Over the past year and a half, the profit realization in the AI industry chain has been dominated by the 'shovel sellers'—NVIDIA, TSMC, ASML, and SK Hynix. Companies selling chips, equipment, or HBM have seen their stock prices rise.

However, this prosperity hides concerns: apart from a few high-barrier segments like NVIDIA's GPUs and TSMC's advanced processes, most upstream hardware will eventually face unsustainable profits due to capacity expansion and increased competition. SK Hynix's stock price halved in this round, reflecting similar concerns.

Amid AI market fluctuations, cloud providers, the closest to profit among 'shovel users,' have become the best-performing AI sector recently. Microsoft and Amazon rose by 25% and 18%, respectively, over three trading days after their earnings releases.

The rise in cloud providers' stock prices is partly due to their restrained investments in large models, which are highly competitive and not yet profitable, aligning with the current cautious bias of the capital market.

A more fundamental reason is the simultaneous increase in cloud providers' revenue, profit margins, and backlogs, indicating that earlier capital expenditures are quickly converting into current revenue, operating profits, and future revenue visibility.

The pace of profit realization by cloud providers directly determines the trading rhythm of 'profit redistribution' in AI investments. Although the exact timing of this inflection point remains uncertain, given the doubts about the sustainability of upstream hardware profits, cloud providers, as an indispensable infrastructure layer in the AI industry, will become a key focus for the market.

/ 01 / Simultaneous Growth in Revenue, Profit, and Backlog Eases Monetization Anxiety

Since May, the market has intensified scrutiny of cloud providers' capital expenditures. Goldman Sachs and Morgan Stanley have released multiple research reports pointing out the unsustainability of cloud providers' capital spending.

This skepticism continued with Google's earnings report. On July 23, Google's stock price plummeted by 7% after its earnings release. The primary concern was that Google, despite its quarterly free cash flow turning negative for the first time, still chose to raise its full-year capital expenditure guidance, further amplifying market concerns about money-burning.

However, anxiety eased a week later. Microsoft and Amazon released their earnings on the 29th and 30th, with their stock prices surging by 15.5% and 9.5% in a single day. Two key factors supported this:

First, management from both companies actively downplayed unrestrained investments in large models during their earnings calls, alleviating market concerns about money-burning. A more critical signal was the accelerated commercialization of AI investments: the earnings reports of the 'three clouds' (AWS, Azure, Google Cloud) cross-validated that cloud providers are gradually completing the commercial closed loop of 'capital expenditure—computing power supply—cloud revenue—profit growth.'

In the second quarter, AWS's cloud business revenue grew by 37% year-over-year, a new 18-quarter high. Azure's cloud computing revenue increased by 43%, exceeding analyst expectations of 39.98%. Google Cloud's revenue surged by 82%, far surpassing market expectations.

Profit margin improvements were also significant. AWS's operating profit margin reached 39.4%, up 6.3 percentage points year-over-year. Google Cloud's profit margin rose from 20.7% to 35.6%.

More indicative of long-term trends was the growth in backlog orders.

Microsoft's commercial RPO, AWS Backlog, and Google Cloud Backlog (RPO/Backlog = total contracted, customer-committed, unamortized amounts) grew by 84%, 154%, and 385% year-over-year, respectively, significantly higher than their respective cloud revenue growth rates. This suggests accelerated future performance growth for the three clouds and serves as a core leading indicator for the recent surge in cloud stocks.

The simultaneous increase in revenue, profit margins, and backlog orders indicates that earlier capital expenditures are quickly converting into current revenue, operating profits, and future revenue visibility.

Two major factors—demand structure and supply upgrades—are driving cloud providers to accelerate into the AI return phase.

From a demand structure perspective, AI demand has shifted from training needs of a few frontier model companies like OpenAI and Anthropic to a broader range of enterprise customers, inference, and application scenarios, simultaneously driving consumption of traditional cloud resources like computing, storage, and networking.

This has been validated by Microsoft in its earnings call: all new RPO orders this quarter came from industrial customers outside OpenAI and Anthropic.

While demand scenarios are increasing, cloud providers' computing power supply is also optimizing. Taking AWS as an example, its AI-related revenue falls into two categories: AI IaaS (bottom-layer computing power leasing) and Bedrock (TaaS/distribution model, Model as a Service). Bedrock's share of AWS's AI revenue has grown from 9% last year to 37% this year.

This structural change indicates that cloud providers' AI monetization models are extending from bottom-layer computing power leasing to model distribution and application monetization. This stock price growth driven by performance growth also reflects a change in the investment logic for cloud providers.

/ 02 / The Pace of ROIC Realization Determines Mid-Term Trends

Over the past two years, the market has rewarded those who spent the most on AI. Initially driven by FOMO (fear of missing out), for every $1 cloud providers spent on AI, their market value rose by $2. The market could overlook temporary returns and profits as long as investments were made, signaling potential.

However, with hundreds of billions of dollars in capital expenditures and negative free cash flow for giants, the investment logic for cloud providers has shifted from 'buying grand narratives' to 'buying evidence of realization.' The market's focus is now accelerating toward the ROIC (Return on Invested Capital) realization phase.

ROIC measures the return on invested capital, calculated as after-tax net operating profit divided by total invested capital. It gauges how much after-tax profit can be stably earned annually for every $1 invested in computing infrastructure, serving as a core indicator for judging whether capital-intensive businesses are worth continued investment.

Addressing market doubts about AI investment realization, Amazon, whose free cash flow turned negative, explained during its earnings call that its computing power investment payback period is less than three years. Given the current five-year depreciation period for core hardware like servers, Amazon can recover its investment in just over half the time and generate substantial profits in the remaining two-plus years.

Amazon's claim aligns with Morgan Stanley's ROIC calculations: stable ROIC for pure GPU leasing by cloud providers is around 31%, representing the return on basic computing power businesses. ROIC for cloud providers' proprietary computing power + model platforms ranges from 46% to 50%.

Compared to the 10% to 18% ROIC of traditional cloud and data center mature businesses, giants' ROIC on AI investments is sufficiently high.

A high ROIC indicates a good business model, but the question remains: Are investments excessive? Amazon also addressed this, explaining that it calculates ROI based on 'firm' orders in hand, ensuring that capacity can convert contracted amounts into recognized revenue once available.

Following Amazon's logic, Amazon, Google, Microsoft, and Oracle have a combined ~$2.23 trillion in outstanding orders, with ~$725 billion in capital expenditures planned for the four clouds by 2026.

At first glance, capital expenditures seem fully covered by potential revenue from outstanding orders. However, unfulfilled orders represent service commitments already contracted but not yet fully recognized as revenue. The 'firmness' of these contracts varies:

Some involve mandatory minimum purchase quantities, requiring customers to meet targets. Others are cancellable contracts with penalties for early termination. Some have negotiable prices and configurations, with amounts in ongoing negotiations. Others involve customer-provided hardware, with cloud providers earning only management fees. Mixed within the same total, these contract types vary significantly in value.

The key metric for judging contract value is how much revenue can be confirmed in the near term.

Among them, Amazon and Microsoft have the fastest revenue recognition, confirming 40% and 30% of unfulfilled order revenue, respectively, within the next 12 months. Over half of Google's backlog will materialize in the next 24 months. Only about 10% of Oracle's backlog will become revenue within 12 months.

Differences in unfulfilled order value and ROIC may lead to divergent performances among cloud providers in the capital markets.

/ 03 / AI Begins Trading 'Profit Redistribution'

Since July, US stock AI sentiment has oscillated among upstream hardware, midstream computing power, and downstream applications. Judging whether the AI theme has shifted depends on tracking profit realization rhythms across the industry chain.

At different stages of industrial development, profits continuously concentrate in the most scarce and price-setting segments. Capital market funds chase profits, flowing from one segment to another, always targeting the most profitable sector within the AI industry chain.

In previous years, the AI profit transmission chain focused on upstream AI hardware and infrastructure, from NVIDIA's GPUs to core hardware like storage and optical modules in cabinets. Driven by supply-demand imbalances, profit margins in these segments surged, creating strong market performance. Meanwhile, upstream support sectors like AI power, computing land, and liquid cooling also experienced explosive growth.

However, apart from NVIDIA's AI chips and TSMC's advanced processes, which have long-term high barriers and undiverted profits, most upstream hardware, including storage, will face cyclical profit declines due to capacity expansion and increased competition. Hence, SK Hynix, Samsung, and Micron saw maximum drawdowns of 40% to 58% in this round.

With doubts about the sustainability of upstream hardware profits, the market has begun trading 'profit redistribution' across the AI industry chain.

Michael Wilson, Morgan Stanley's Chief US Strategist, reminded investors in his July weekly report to reduce overweight positions in semiconductors and shift focus to hyperscale cloud providers. He believes AI investors are moving from chasing upstream hardware to focusing on downstream cloud services' commercial benefits. In other words, institutions see cloud providers as more certain in realizing profits in the mid-to-lower reaches of the industry chain.

The logic is straightforward: when upstream hardware is no longer scarce and application scenarios gradually materialize, more AI industry profits will shift to midstream cloud providers, model companies, and downstream application segments.

Compared to the uncertainty of diverse downstream applications, midstream cloud providers and model companies represent an indispensable infrastructure layer. Unlike the 'sell-and-leave' model of upstream hardware, they can achieve sustained revenue.

Among cloud providers and model companies, the model industry remains in a 'high-growth but negative-profit' phase, while cloud providers, with a clear competitive landscape and mature business models, have entered a virtuous cycle of 'high-growth and rising profit margins.'

From current trends, cloud providers can even squeeze profits from the model layer. AWS's Bedrock model is a typical example: without self-developed (independently developing) large models, it earns higher profits than model vendors by distributing models like Claude. Data shows that AWS's TaaS/distribution model EBIT margin can reach up to 55%, while most model vendors' average gross margin is only 30%.

Although the exact timing of the AI investment theme switch remains unclear, the profit-seeking nature of capital will not change. The next segment for profit realization in the AI industry chain will inevitably become the market's focus in the next phase.