Morgan Stanley is positive on a 25%-50% ROIC path for generative AI investments

Overview

This report focuses on the return on investment question for North American internet and cloud leaders amid rapidly rising generative AI capex, compute expansion, and model training investment. Morgan Stanley believes that although the market worries AI investment by large cloud vendors and model developers could weigh on sentiment and valuation multiples, as the industry enters the inference revenue phase, GPU IaaS, model APIs, and model APIs based on third-party compute all have meaningful paths to incremental margins and ROIC.

Core views

The core views of the report are: first, tight compute supply and rising data center capacity value will allow hyperscalers such as AWS, GCP, and Azure to earn scalable and profitable revenue from inference infrastructure; second, model labs need continued investment and innovation to maintain token pricing power and improve token throughput efficiency; third, healthy returns on AI investment will also attract more competitors and open-source solutions, so product differentiation, chip efficiency, and software efficiency will become the long-term deciding factors; fourth, the report remains optimistic on AMZN, GOOGL, MSFT, and META, with META’s AI, advertising, Reels, and efficiency improvements seen as offering multiple upside options.

Analysis framework

The report uses three standardized bottom-up GenAI ROIC frameworks: 1) Hyperscaler GPU Rental, where cloud vendors rent out GPU inference capacity; 2) Model Enabled API on owned infrastructure; 3) Model Enabled API running on rented third-party infrastructure. Each framework separately breaks down revenue-side variables such as GPU pricing, utilization, token pricing, and token throughput, as well as cost-side items including IT/server/network depreciation, powered shell depreciation, energy, operations and maintenance, labor, and third-party compute procurement costs.

Methodology notes

• ROICHyperscaler GPU Rental

  Estimate cloud vendors' GPU rental business based on 1GW of NVDA GB300 inference capacity.

  The model derives revenue from GPU hourly pricing and utilization, and includes IT and non-IT depreciation, energy costs, and other operating expenses in costs. The report estimates that this model can achieve about 60%-70% incremental EBIT margins and 25%-40% ROIC.

• ROICModel Enabled API

  Model developers provide API access on owned data centers.

  Key revenue variables include the share of compute used for inference, tokens/second/GPU, and token pricing; costs include depreciation of data center assets, energy, and operating expenses. The report estimates that this model can achieve incremental EBIT margins above 70% and ROIC above 40%.

• ROICModel Enabled API on 3P Infrastructure

  Model developers rent compute from hyperscalers or other third parties to provide APIs.

  This model avoids building part of the infrastructure in-house but requires paying third parties for compute. The report estimates it can still achieve about 30% incremental EBIT margins and ROIC above 25%, with token pricing, token throughput, and compute procurement costs being the most sensitive variables.

• Assumption1GW NVDA GB300 Standardization

  Use NVDA GB300 GPUs as the standard compute device across the three frameworks.

  The report uses 1GW of GB300 capacity as a common benchmark and applies an all-in capital cost of about $39bn/GW, corresponding to about $6bn of annual depreciation for owned infrastructure; this assumption affects GPU count, hourly pricing, token throughput, and depreciation expense.

• AssumptionToken Throughput Range

  Token throughput is not a fixed hardware attribute, but depends on the model, software stack, and input/output mix.

  The report uses an illustrative range of 2,000-3,500 tokens/second/GPU and notes that while larger-parameter models may command higher token pricing, they usually come with lower throughput, so higher pricing does not necessarily translate linearly into higher revenue/GW.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• META PLATFORMS INC (US.META)

  Top Pick; beneficiary of model APIs, AI advertising, Reels, data centers, and efficiency improvements.

  Strengths

  The report highlights META’s structural shift toward efficiency, improved revenue and engagement, progress in Reels monetization, and potential upside options from AI, subscriptions, and click-to-message; target price $775.00, bull-case scenario $1,000.00.

  Weaknesses

  AI and data center investment raises capital intensity; if Reels monetization is slower than expected or Reality Labs losses widen, operating profit and free cash flow could be pressured.

  Comparison

  Compared with the more infrastructure-oriented AWS, GCP, and Azure, META’s opportunity is more tied to models, advertising products, user engagement, and monetization of owned data centers.

  Risks

  Macro pressure, weak consumer demand, regulatory limits on targeted advertising, wider Reality Labs losses, poor data center buildout execution, and rising long-term capital intensity.

• Amazon.com Inc (AMZN.O)

  AWS and high-margin businesses benefit from cloud adoption and AI inference compute demand.

  Strengths

  The report notes that Amazon’s high-margin businesses support profitability improvement, while AWS is in a multi-year cloud adoption cycle, and advertising is also an important source of growth and profit; target price $330.00.

  Weaknesses

  Retail fulfillment, last-mile delivery, Prime, Fresh, Alexa/Echo, India, and AWS still require continued investment.

  Comparison

  AMZN’s AI investment returns are more concentrated in AWS compute rental and the cloud infrastructure layer, while retail and advertising provide diversified profit pools.

  Risks

  Investment spending remains higher than expected and lasts longer, merchandise margins come in below expectations, or AWS revenue growth slows or margins decline.

• Microsoft (MSFT.O)

  Azure, Microsoft AI Copilots, and M365 commercial SKUs benefit from cloud and AI adoption.

  Strengths

  The report emphasizes that Azure growth, Copilot monetization, adoption of premium M365 SKUs, and seat growth can drive high-teens to low-twenties revenue CAGR and support EPS growth above 20%; target price $600.00.

  Weaknesses

  Recently reported margins may be below the theoretical incremental margins of the GB300 inference scenario, because the current chip generation, traditional cloud workloads, and additional R&D/labor costs still affect reported results.

  Comparison

  MSFT has both cloud infrastructure and the software application layer, with AI monetization spanning Azure and Copilot.

  Risks

  Macro weakness affecting IT spending, on-premise being displaced by cloud, increased investment weighing on margins, and limited AI adoption.

• Alphabet Inc. (GOOGL.O)

  Beneficiary of AI innovation across GCP, Google Gemini, Search, and YouTube.

  Strengths

  The report believes AI-driven platform-level innovation across Search, YouTube, Cloud, and other products can improve the sustainability of long-term growth, while expense discipline may also drive upward revisions to EPS and FCF; target price $400.00.

  Weaknesses

  Model development and infrastructure investment are highly intensive, requiring continued product innovation and efficiency gains to preserve token pricing and throughput advantages.

  Comparison

  GOOGL combines cloud infrastructure, model APIs, and consumer distribution channels, placing it at both the hyperscaler and model developer ends.

  Risks

  AI competition, changes in the search business model, cloud growth below expectations, and costs running above expectations.

Key data

• Capex of major hyperscalers$1.4trln+Citing prior research, the report expects the capex path of major hyperscalers to exceed $1.4trln by 2028.
• Growth in compute capacityabout 4x, to about 120GWCompute capacity is expected to grow about fourfold from 2025 to 2028.
• Target ROIC for the three GenAI models25%-50%Both the report title and core argument point to a 25%-50% return on capital path for generative AI investments.
• Hyperscaler GPU Rental incremental EBIT marginabout 60%-70%The report believes cloud rental of GB300 inference capacity can generate high incremental margins.
• Hyperscaler GPU Rental ROIC25%-40%This model is affected by GPU pricing, utilization, depreciation, and energy costs.
• Owned-infrastructure model API incremental EBIT margin70%+The report believes model APIs on owned data centers can generate higher incremental margins.
• Owned-infrastructure model API ROIC40%+Token pricing, token throughput, and product innovation are key drivers.
• Third-party infrastructure model API incremental EBIT marginabout 30%Third-party compute costs depress margins, but unit economics remain attractive.
• Third-party infrastructure model API ROIC25%+The report describes returns in this scenario on a NOPAT basis.
• GB300 all-in capital cost assumptionabout $39bn/GWThis corresponds to about $6bn of annual depreciation expense for owned infrastructure.
• Owned GPU utilization assumption75%The report says this assumption is close to the industry average for well-optimized AI data centers.
• META target price$775.00The report’s META risk/reward page shows a target price of $775.00 and labels it Top Pick.
• META current price$595.19The regulatory disclosure table lists Meta Platforms Inc at $595.19.

Impact & implications

The report’s investment implication is positive: AI capex will continue to rise, but if inference demand scales and generates API and cloud infrastructure revenue, capex can be converted into high incremental profits and ROIC. For cloud vendors, compute scarcity reinforces the value of data center capacity and GPU supply; for model companies, sustained model innovation and software efficiency improvement determine token pricing power and the cost curve; for internet giants, companies with capital, data, distribution, and compute scale are better positioned to absorb training costs and earn returns from inference monetization.

Risks

• If AI capex and model training investment cannot be effectively converted into inference revenue, ROIC may fall short of the report framework assumptions.
• GPU pricing, utilization, token pricing, token throughput, and third-party compute procurement costs are highly sensitive variables, and deterioration in any one of them would compress unit economics.
• New entrants and open-source models could intensify competition, potentially lowering token pricing and weakening model API margins.
• The report uses standardized assumptions based on GB300 and 1GW, while actual business outcomes will be affected by chip generation, workload mix, software configuration, and data center operating differences.
• Incremental margins disclosed by cloud vendors in the short term may be below theoretical scenarios because traditional cloud workloads, older chip capacity, and additional labor and R&D expenses still remain.
• The main downside risks for META include macro pressure, advertising regulation, weaker-than-expected Reels monetization, wider Reality Labs losses, and rising capital intensity caused by data center execution mistakes.

What to watch

• Whether AI capex, data center capacity, and compute capacity at major hyperscalers continue to expand as expected.
• GPU hourly pricing, supply timing, utilization, and depreciation assumptions for GB300 and subsequent chips.
• Token pricing, tokens/second/GPU, input/output mix, and changes in parameter size for model APIs.
• The allocation ratio between training and inference compute, especially whether training/model maintenance costs crowd out monetizable inference capacity.
• Changes in AI inference revenue, cloud growth, and disclosed margins at AWS, GCP, and Azure.
• META’s Reels engagement and monetization, AI advertising effectiveness, subscriptions, and click-to-message progress.
• Whether GOOGL’s Gemini, Search, YouTube, and GCP AI innovations translate into revenue and profit upgrades.
• MSFT’s Azure AI services, Copilot adoption, and premium M365 SKU penetration.