2026-07-29 18:01

Introduction: The key isn’t how many GPUs you buy, but capturing pricing power during scarcity cycles—self-built infrastructure captures upstream margins, while renting models leaves only 25% return

Hyper-scale players are investing hundreds of billions of dollars annually into data center construction. The market treats this capital expenditure as a cost. Morgan Stanley’s July 27 report presents a different calculation.

Building in-house GPU clusters for compute leasing can achieve up to a 40% ROIC. Using self-built compute for API calls also reaches ~40%. However, when compute is rented, ROIC drops to just 25%.

The gap stems from pricing power during compute scarcity. Whoever controls self-built compute holds the upper hand in profit distribution. Morgan Stanley maintains Overweight ratings on Microsoft, Amazon, Meta, and Google, with target prices of $600, $330, $775, and $400 respectively.

Markets see these firms spending heavily. Morgan Stanley sees that capital being converted into profit. If 40% ROIC proves sustainable, the AI businesses of these four companies will consistently outperform expectations—and the market has yet to fully price this in.

Significant ROIC Divergence Across Compute Models

Morgan Stanley breaks down generative AI’s commercialization into three distinct models, with ROIC ranging from 25% to 40%. The differences stem fundamentally from how compute is acquired and thus how profits are allocated.

· First model: Hyper-scale players build their own GPU clusters for compute leasing. Based on NVIDIA GB300, Morgan Stanley assumes 410,000 GPUs, 3.6 billion total GPU-hours, 75% utilization, and $8.50/hour rental rate. Incremental EBIT margin ranges between 60% and 70%, yielding ROIC between 25% and 40%.

Cost structure includes depreciation of IT equipment (servers + networking) and non-IT infrastructure (power, cooling, racks), plus energy and operational expenses. In a persistently scarce compute environment, rental prices are determined by supply-demand dynamics—not cost-plus markup. This enables EBIT margins reaching 60%-70%, the core driver of high returns.

· Second model: Model providers build their own compute for API inference. Assumptions: 65% compute dedicated to inference, 2,750 tokens per second per GPU, $1.75 per million tokens. Incremental margin ~70%, ROIC ~40%.

Morgan Stanley views this pricing level as positive for model providers like Google Gemini, Meta’s API platform, and xAI Grok. However, a persistent constraint remains: allocating compute between revenue-generating inference and cost-heavy training. Training compute generates no immediate revenue but bears full depreciation and OPEX. This trade-off directly impacts the sustainability of ROIC.

· Third model: Model providers rent third-party compute for APIs. No hardware ownership—compute is leased hourly from hyper-scale providers. With an additional layer of “intermediary margin,” incremental margin falls to 30%, and ROIC settles at ~25%. Here, the hyper-scale provider’s rental income becomes a cost item for the model vendor, with full pricing power concentrated upstream.

Comparing all three models, self-built compute delivers nearly double the ROIC of the rental model. The core difference lies in pricing power under compute scarcity. Those with owned infrastructure control the profit allocation. Morgan Stanley’s ROIC estimates already exclude training costs, using a more conservative assumption than the market. Efficiency improvements in chip and software token throughput continue to progress, suggesting the ROIC ceiling may exceed current assumptions.

Four Cloud Giants: Profitability & Valuation Scenario Breakdown

Morgan Stanley maintains Overweight ratings on all four firms, each with clear drivers and risk-reward structures.

· Microsoft: Target price $600, implying 25x P/E on expected 2028 fiscal year EPS of $23.86. Current stock price reflects less than 16x P/E on 2028 GAAP EPS—Morgan Stanley considers valuation undervalued.

Key drivers: The inflection point in Azure growth combined with Copilot’s monetization potential. Adoption of Azure AI services is accelerating, M365 Copilot penetration in enterprise clients continues rising, and upgrades to higher-priced SKUs are lifting average revenue per user.

Bull case target: $795, based on ~29x P/E on $27.39 EPS. Bear case not specified, but implied ~12x P/E on $21.64 EPS. Under bear scenario, Azure growth slows further due to base effects, Copilot adoption remains limited, and macro weakness constrains corporate IT spending.

· Amazon: Target price $330, based on 25x P/E on average projected EPS of ~$13 for FY2027–2028. Morgan Stanley sees profit improvement driven by three engines simultaneously gaining momentum: AWS cloud growth accelerating, advertising continuing to deliver high-margin revenue, and retail fulfillment efficiency improving.

Recurring revenue from Prime memberships and favorable business mix shift are central to sustaining valuation premium. Bull case: $400; Bear case: $210.

· Meta: Target price $775, derived from a DCF model implying ~23x P/E on expected 2027 EPS. Morgan Stanley forecasts Meta’s 2026 ad revenue growth of ~27%, with AI investment as the primary driver.

AI is enhancing Reels engagement and monetization efficiency. Ad measurement and attribution capabilities continue recovering post-privacy changes. New ad products like click-to-message are unlocking incremental growth. Bull case: $1,000; Bear case not quantified.

· Google: Target price $400, based on ~24x P/E on average projected EPS of $15–$18 for FY2027–2028, equivalent to a 1.6x PEG ratio—~35% above sector median. Morgan Stanley believes AI-driven innovation in search, YouTube, and cloud platforms is improving long-term growth visibility. Products like AI Overviews enhance UX while maintaining strong ad monetization efficiency. Bull case: $450; Bear case: $225.

Valuation Framework May Face Reassessment

If 40% ROIC is real, then these hundreds of billions in capex are generating value. Market perception of this capital will shift—and once it does, the valuation frameworks for these four companies must evolve accordingly.

Disclaimer: This article is a synthesis and interpretation of a third-party brokerage research report (Morgan Stanley, July 27, 2026), compiled using publicly available market information. All ratings, target prices, earnings forecasts, and related judgments cited herein reflect the views of the analysts at the originating institution and represent their institutional positions only—not those of ChaoXiang Research, nor constitute any investment advice. Markets involve risk; decisions must be made independently. This content should not be used as a basis for buying or selling any securities.

Source: DeepTide TechFlow