GenAI investment is expected to move toward 25%-50% ROIC, benefiting hyperscalers and model developers
AI summary card
GenAI investment is expected to move toward 25%-50% ROIC, benefiting hyperscalers and model developers
Through three bottom-up GenAI ROIC frameworks, Morgan Stanley argues that in the AI inference era, GPU leasing, model APIs, and third-party infrastructure API businesses can all generate attractive incremental returns.
- The report believes investor concerns about returns on AI capex are overstated, and that major cloud providers and model developers still have strong return potential during the inference monetization phase.
- Hyperscaler GPU leasing businesses can achieve roughly 60%-70% incremental EBIT margins and 25%-40% ROIC under GB300 inference capacity assumptions.
- Model API businesses running on owned infrastructure offer the highest returns, with the report estimating over 70% incremental EBIT margins and over 40% ROIC.
- Model API businesses renting third-party infrastructure can still achieve roughly 30% incremental EBIT margins and about 25% ROIC, though economics are more dependent on compute procurement costs.
- The report is positive on AMZN, GOOGL, MSFT, and META, with META's risk-reward page showing a target price of $775.00 and designating it as the Top Pick.
Report interpretation
Overview
This report focuses on returns from North American internet and AI infrastructure investment. Morgan Stanley notes that AI capex, compute capacity, and model training investment by major hyperscalers are still rising rapidly through 2028, and the market's core debate is whether these investments can generate sufficient ROIC. Using bottom-up frameworks across three GenAI business models, the report concludes that commercialization paths in the AI inference era can support roughly 25%-50% incremental ROIC, and therefore it remains positive on AMZN, GOOGL, MSFT, and META.
Core views
The core view is that AI capex is not simply a cost burden, but an upfront investment for cloud computing and model API monetization. GPU IaaS can achieve high utilization and pricing power in a compute-constrained environment; model APIs running on owned infrastructure can generate the highest returns through token pricing, throughput efficiency, and scaled inference; even when model developers rent third-party infrastructure, there is still a path to roughly 25% ROIC after deducting intermediate compute costs. The report also emphasizes that high returns will attract new entrants and open-source competition, so sustained product innovation, chip efficiency, software stack optimization, and data center capacity are critical to maintaining returns.
Analysis framework
The report uses a bottom-up scenario modeling approach centered on 1GW NVDA GB300 compute deployments, standardizing three frameworks: Hyperscaler GPU Rental, Model Enabled API, and Model Enabled API Running on 3P Infrastructure. Key revenue-side sensitivities include GPU hourly pricing, capacity utilization, the share of inference compute, tokens/second/GPU, and token pricing; cost items include IT equipment depreciation, powered shell depreciation, energy, operations, labor maintenance, and third-party compute rental costs. These assumptions are then mapped to incremental EBIT margins, NOPAT, and ROIC.
Methodology notes
ROIC frameworks for three AI inference business models
The report breaks AI inference monetization into three models—GPU leasing, model APIs on owned infrastructure, and model APIs on third-party infrastructure—and evaluates revenue, cost, margins, and return on capital for each.
Using 1GW of GB300 compute as a common valuation benchmark
The report uses GB300 as the standard IT/compute device, assuming an all-in capital cost of about $39bn/GW, and based on this estimates roughly $6bn of annual depreciation expense in the owned infrastructure scenario.
Token pricing and tokens/second/GPU jointly determine model API revenue elasticity
The report emphasizes that token throughput is determined not only by chips, but also by model parameters, the software stack, input-output structure, and interaction speed; larger models may command higher pricing but lower throughput, so higher pricing does not necessarily translate proportionally into higher revenue per GW.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- META.USTop Pick; a dual beneficiary as both a model developer and an advertising platform
- Strengths
- AI investment can improve Reels engagement, ad measurement and attribution, Family of Apps monetization, and create additional upside optionality in AI, subscriptions, and click-to-message.
- Weaknesses
- Capex intensity and Reality Labs losses may weigh on free cash flow and operating profit.
- Comparison
- Compared with pure-play cloud providers, META's core returns also come from improved ad efficiency and platform engagement; the report assigns a target price of $775.00.
- Risks
- Macro pressure, regulatory limits on targeted advertising, slower-than-expected Reels monetization, poor data center execution, and rising long-term capital intensity.
- AMZN.USA beneficiary through AWS and high-margin e-commerce businesses
- Strengths
- AWS is in a long-term cloud adoption cycle, and high-margin businesses support continued investment while improving profitability; the ad business also has room for growth and profit contribution.
- Weaknesses
- Retail fulfillment, last-mile delivery, Prime content, Alexa, and overseas investments may still pressure margins.
- Comparison
- The report lists AMZN with a target price of $330.00, based on roughly 25x average 2027/2028 EPS.
- Risks
- Investment duration exceeding expectations, merchandise margins below expectations, or AWS revenue slowing or margins declining.
- GOOGL.USA beneficiary through GCP, Search, and Gemini model APIs
- Strengths
- AI-driven innovation across Search, YouTube, Cloud, and other products can improve long-term growth visibility, while expense discipline can support upward revisions to EPS and FCF.
- Weaknesses
- Higher AI investment and competitive pressure may affect short-term margins, and the durability of the Search business model in the AI era still needs to be continuously proven.
- Comparison
- The report lists GOOGL with a target price of $400.00, based on roughly 24x average 2027/2028 EPS.
- Risks
- AI search competition, cloud business growth below expectations, regulatory pressure, and capex returns falling short of expectations.
- MSFT.USA beneficiary through Azure, M365, and Copilot
- Strengths
- Azure growth, M365 commercial SKU upgrades, seat growth, and Copilot adoption are driving high-single-digit to low-twenties revenue growth, with operating leverage supporting EPS growth.
- Weaknesses
- Reported margins in the near term may be below the theoretical incremental margins of the GB300 inference framework, because capacity mix, traditional cloud workloads, and additional R&D expenses dilute reported margins.
- Comparison
- The report lists MSFT with a target price of $600.00, with the base case using 25x FY28e EPS of $23.86.
- Risks
- A macro environment that suppresses IT spending, cloud cannibalization of on-premise deployments, increased investment hurting margins, and limited AI adoption.
Key data
- AI capex by major cloud providers$1.4trln+Citing prior research, the report expects capex by major hyperscalers to exceed $1.4trln by 2028.
- Compute capacity growthAbout 4x, to ~120GWCompute capacity is expected to roughly quadruple from 2025 to 2028.
- GPU IaaS incremental EBIT margin~60%-70%Based on GB300 GPU hourly pricing, utilization, depreciation, energy, and other cost assumptions.
- GPU IaaS ROIC25%-40%The report believes AWS, GCP, and Azure can generate scalable returns from next-generation inference compute.
- Owned infrastructure model API incremental EBIT margin~70%+Applicable to model API business frameworks such as GOOGL Gemini, META API, and Grok.
- Owned infrastructure model API ROIC~40%+Model quality, token pricing, throughput efficiency, and the allocation of training and maintenance costs are key.
- Third-party infrastructure model API incremental EBIT margin~30%Renting third-party compute requires paying intermediary profits, reducing unit economics.
- Third-party infrastructure model API ROIC~25%Still viewed by the report as attractive, but more sensitive to compute procurement costs.
- Owned GPU utilization assumption75%Used for the owned infrastructure API scenario, reflecting average industry utilization levels for optimized AI data centers.
- META target price$775.00The risk-reward page shows META as the Top Pick, with the target price based on DCF and a long-term EBITDA multiple, implying about 23x 2027 P/E.
Impact & implications
The investment implication is that if the market focuses only on AI capex pressure, it may underestimate the high incremental margins and high ROIC that inference monetization can generate. The report is broadly positive on cloud infrastructure at hyperscalers, API monetization by model developers, and AI infrastructure enablers. At the single-stock level, AMZN, GOOGL, and MSFT benefit from cloud compute leasing and AI service growth, while META also benefits from ad business efficiency, improved Reels monetization, AI-driven engagement gains, and potential monetization of APIs and data center capacity.
Risks
- AI capex and model training investment continue to rise, but the pace of inference revenue monetization may be slower than expected.
- Declining token prices, insufficient token throughput efficiency, or weak model product differentiation could weaken model API ROIC.
- New entrants and open-source models increase competition, compressing pricing and the long-term profit pool.
- Data center construction costs, energy costs, equipment depreciation, and third-party compute rental costs may exceed assumptions.
- Short-term reported margins at cloud providers may be weighed down by generational capacity mix, traditional cloud workloads, and additional R&D expenses.
- Macro weakness, slower enterprise IT spending, softer advertising demand, or tighter regulation could affect internet platform revenue.
What to watch
- AI capex growth, data center capacity ramp schedules, and compute utilization at major hyperscalers.
- Actual per-GPU token throughput, energy efficiency, and hourly pricing for GB300 and subsequent chips.
- Token pricing for model APIs, developer adoption, enterprise deployment, and the pace of revenue scaling.
- The allocation between training and inference compute, and the extent to which model maintenance costs erode ROIC.
- Revenue growth and cloud margin trends for AWS, GCP, and Azure in AI inference capacity.
- META's Reels monetization, ad measurement improvements, AI product/API progress, and data center capital intensity.