Written by: Rita

Large-scale vendors invest hundreds of billions of dollars each year to build data centers. The market sees this money as a cost. Morgan Stanley calculated a different equation in its research report on July 27.

Building in-house GPU clusters for computing power leasing can achieve a maximum return on invested capital of 40%. Creating computing power for API calls can also achieve 40%. However, if the computing power is rented, the ROIC drops to only 25%.

The gap comes from the pricing power during times of computing power scarcity. Those who hold self-built computing power control the upstream of profit distribution. Morgan Stanley maintains overweight ratings for Microsoft, Amazon, Meta, and Google, with target prices of $600, $330, $775, and $400 respectively.

The market sees these companies spending money. Morgan Stanley sees the money being spent by these companies turning into profits. If the 40% ROIC holds, the AI businesses of these four companies will continue to exceed expectations, and the market has not yet factored this in.

Different computing power models show significant ROIC return differences

Morgan Stanley breaks the commercialization path of generative AI into three models, with ROIC ranging from 25% to 40%. The differences between the models stem from the profit distribution pattern determined by the method of acquiring computing power.

· The first type is large-scale vendors building their own GPU clusters for computing power leasing. Using Nvidia's GB300 as a benchmark, Morgan Stanley assumes 410,000 GPUs, 3.6 billion total GPU hours, 75% utilization, and a rental price of $8.50 per hour. Incremental EBIT margins are between 60% and 70%, with ROIC between 25% and 40%.

In terms of cost structure, Morgan Stanley breaks costs down into IT equipment (servers + networks) and non-IT facilities (power supply, cooling, racks) depreciation, plus energy costs and operational expenditures. In an ecosystem where computing power is continually scarce, rental prices are determined by supply and demand, free from cost markup constraints. This is the fundamental reason why profit margins can reach 60% to 70%.

· The second type is model vendors building their own computing power for API calls. Morgan Stanley assumes that 65% of computing power is used for inference, each GPU processes 2,750 tokens per second, and charges $1.75 per million tokens. Incremental profit margins are 70%, with ROIC around 40%.

Morgan Stanley believes this pricing level sends a positive signal to model vendors like Google Gemini, Meta API platform, and xAI Grok. However, model vendors face a persistent constraint: computing power needs to be allocated between "income-generating inference" and "maintaining technological leadership training." The computing power invested in training does not generate current income but bears the full depreciation and operating costs. This trade-off will directly affect the sustainability of ROIC.

· The third type is renting third-party computing power for APIs. Model vendors do not own hardware and rent computing power on an hourly basis from large-scale vendors. After an additional layer of "intermediary profit," incremental profit margins drop to 30%, with ROIC about 25%. In this model, the rental income from large-scale vendors constitutes part of the costs for model vendors, and the pricing power lies entirely upstream.

When comparing the three models, the ROIC of self-built computing power is nearly twice that of the rental model. The core of the difference lies in the ownership of pricing power in an ecosystem of computing power scarcity. Those who possess self-built computing power take the initiative in profit distribution. Morgan Stanley's calculations of ROIC have already deducted training costs and are more conservative than market estimates. The efficiency of token throughput on the chip and software levels continues to improve, suggesting that the upper limit of ROIC could be higher than currently assumed.

Four major cloud vendors: Profit and valuation scenario breakdown

Morgan Stanley maintains overweight ratings for all four vendors, each with clear driving logic and risk-reward structures.

· Microsoft has a target price of $600, corresponding to a price-to-earnings ratio of 25 times the expected earnings per share of $23.86 for fiscal year 2028. The current stock price corresponds to less than 16 times GAAP earnings per share for fiscal year 2028, which Morgan Stanley views as undervalued.

In terms of driving logic, Morgan Stanley emphasizes the synergistic effect of Azure's growth inflection point and Copilot's monetization capability. The adoption of Azure AI services is accelerating, the penetration rate of M365 Copilot among commercial clients continues to rise, and the upgrades of high-priced SKUs are raising the average transaction value.

The bullish market target price is $795, corresponding to about 29 times earnings of $27.39 per share. No target price is set for the bearish market, corresponding to an earnings per share of $21.64 at approximately 12 times earnings. In the bearish scenario, Azure's growth continues to slow due to the base effect, the adoption of Copilot is limited, and macroeconomic weaknesses suppress enterprise IT spending.

· Amazon has a target price of $330, corresponding to a price-to-earnings ratio of 25 times the average expected earnings per share of about $13 for 2027 to 2028. Morgan Stanley believes that Amazon's profit improvement comes from the simultaneous efforts of three engines: AWS cloud business growth is accelerating, the advertising business continues to contribute high-margin revenue, and the fulfillment efficiency of the retail business continues to improve.

Recurring revenue from Prime memberships and positive shifts in business structure are the core logic supporting the valuation premium. The bullish target price is $400, and the bearish target price is $210.

· Meta has a target price of $775, based on a discounted cash flow model, implying an expected price-to-earnings ratio of about 23 times for 2027. Morgan Stanley expects Meta's advertising revenue to grow by about 27% in 2026, driven mainly by AI investments.

AI is enhancing engagement and monetization efficiency for Reels, ad measurement and attribution capabilities are continuously recovering post-privacy policy changes, and new advertising products like click-to-message are opening up incremental space. The bullish target price is $1,000, with no specific number provided for the bearish scenario.

· Google has a target price of $400, corresponding to an approximate price-to-earnings ratio of about 24 times the expected average earnings per share (between $15 and $18) for 2027 to 2028, which equates to a premium of about 35% relative to the industry median at 1.6 times the price-to-earnings growth rate. Morgan Stanley believes that AI-driven innovations in search, YouTube, and cloud platforms are improving the predictability of long-term growth, while new products like AI Overviews enhance user experience while maintaining advertising monetization efficiency. The bullish target price is $450, and the bearish target price is $225.

Valuation logic may face re-evaluation

If a 40% ROIC is real, then these hundreds of billions of dollars in capital expenditures are creating value. The market's perception of this money will change, and once that changes, the valuation frameworks of the four companies will also change.

Disclaimer
This article is a compilation and interpretation of third-party brokerage research reports (Morgan Stanley, July 27, 2026) by Chaoxiang Research, combined with publicly available market information. The ratings, target prices, earnings forecasts, and related judgments quoted in this article are those of the respective brokerage analysts and only represent the stance of their organizations, not the views of Chaoxiang Research, nor do they constitute any investment advice.
Markets carry risks; decisions must be independent. This article should not be used as the basis for buying or selling any securities.