Morgan Stanley recommends treating the AI infrastructure pullback as a buying opportunity
AI summary card
Morgan Stanley recommends treating the AI infrastructure pullback as a buying opportunity
The report is bullish on the long-term value of the "intelligence superhighway," arguing that compute demand will likely materially exceed supply for years, though data center construction, power access, policy divergence, and model competition will create periodic speed bumps.
- The highest-conviction view is that compute demand will materially exceed supply over multiple years, and the nonlinear improvement in AI capabilities will continue to lift the value of intelligence and compute.
- The report rebuts concerns that enterprise "tokenmaxxing" will suppress revenue, arguing that the cost-benefit of enterprise AI use cases remains highly attractive.
- Progress in Chinese LLMs poses a real competitive threat to U.S. frontier model developers, but it may also reinforce total compute demand through Jevons Paradox.
- Data center expansion faces labor, power, and political resistance; the U.S. could face a data center capacity shortfall of more than 10% by 2028.
- Core positioning areas include AI infrastructure bottlenecks, the compute manufacturing ecosystem, Chinese AI solution providers, energy security assets, and hyperscale cloud providers such as META, GOOGL, MSFT, and AMZN.
Report interpretation
Overview
This is a thematic Morgan Stanley report on global AI infrastructure investment opportunities and risks after the pullback. The report argues that after sharp volatility in AI infrastructure stocks, market concerns have focused mainly on Chinese model competition, constraints on enterprise token spending, policy and data center construction bottlenecks, model lock-in, and physical supply constraints. The authors are broadly bullish on the AI infrastructure cycle, believing the "intelligence superhighway" will generate significant net benefits for the global economy and that compute demand will likely materially exceed supply over multiple years.
Core views
The core views are: first, AI capabilities are improving at a nonlinear pace, and the value of intelligence and compute will continue to rise rapidly; second, the economics of enterprise AI use are strong, with token costs representing only a small portion of potential labor-cost savings; third, returns on AI capex remain attractive, and both large and efficient models can support strong ROI for underlying infrastructure; fourth, Jevons Paradox will continue to apply, meaning more efficient models and chips do not necessarily reduce total compute consumption, but may instead broaden adoption and increase usage frequency and complexity; fifth, labor, power, and political constraints in data center construction are real issues, but are more like speed bumps than long-term barriers.
Analysis framework
The report starts from the debates investors care about most, decomposing the AI infrastructure pullback into fundamental and technical factors, and uses token economics models, the Intelligence Factory model, U.S. power shortage analysis, data center resistance tracking, enterprise and consumer AI adoption surveys, global AI stock mapping, and Future of Work analysis to assess AI demand, supply bottlenecks, profitability, and asset mapping.
Methodology notes
Efficiency improvements may expand total demand
The report applies Jevons Paradox to AI compute: more efficient chips and models reduce the unit cost of generating intelligence, which may stimulate more users, more applications, and more complex workloads, ultimately increasing total compute and power demand.
Enterprise AI adoption and token sales margin estimation
The report compares cost savings from enterprise AI tasks with token costs, and estimates the margins from selling tokens out of data centers across different GPU generations to assess the economics for both AI adopters and compute infrastructure owners.
ROIC and margin analysis for data center and model combinations
The report uses the Intelligence Factory model to evaluate returns from different models and data center configurations. For example, Kimi K3 versus existing frontier models could moderately improve the operating margin of a 200MW data center by about 3%.
The impact of power access speed on AI infrastructure value
The report emphasizes that "time to power" is an important source of alpha, focusing on solutions such as fuel cells, natural gas turbines, nuclear sites, energy storage, power developers, and bitcoin site conversions that can shorten time to electricity supply.
Divergence in China-U.S. AI policy and market access
The report believes the probability of AI policy intervention is rising in both the U.S. and China, which could further fragment how AI is accessed, adopted, and brought to market globally.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- AI infrastructure bottleneck assetsDirectly benefit from data center expansion and supply shortages
- Strengths
- Areas such as labor, time to power, fuel cells, turbines, energy storage, power developers, and data center REITs can ease the most critical bottlenecks.
- Weaknesses
- Construction cycles, grid-connection approvals, shortages of skilled workers, and local political resistance may slow realization.
- Comparison
- Compared with pure AI application assets, bottleneck assets have more direct exposure to the compute supply-demand gap.
- Risks
- Failure to secure power access, project delays, local opposition to data center expansion, and excessive capex.
- Compute manufacturing ecosystemBenefits from rising intelligence value and undersupplied compute
- Strengths
- GPUs, TPUs, and related manufacturing chains benefit as AI capabilities improve and demand for training and inference grows.
- Weaknesses
- Valuations and cycle expectations may already reflect high growth, and the sector is affected by supply chains and technology iteration.
- Comparison
- Compared with downstream applications, the manufacturing ecosystem is more directly tied to AI infrastructure capex.
- Risks
- Improved efficiency of Chinese models may trigger market concerns about training compute demand, along with chip-generation competition and supply bottlenecks.
- Chinese AI solution providersBenefit from improving model capabilities, cost competitiveness, and adoption in the domestic market
- Strengths
- The report argues that Chinese LLM capabilities and global competitiveness are strengthening, and that many related stocks have not yet fully priced in the upside.
- Weaknesses
- Policy, market access, and China-U.S. technology divergence may constrain global expansion.
- Comparison
- Compared with U.S. frontier model developers, Chinese solutions may be more competitive in cost and in the open-weight ecosystem.
- Risks
- Regulatory restrictions, geopolitics, uncertainty around model commercialization, and further fragmentation from the U.S. AI ecosystem.
- Energy security and energy storage assetsProvide power security for AI infrastructure and are important enablers of data center expansion
- Strengths
- Energy storage is highlighted by the report as a standout asset class; power developers and energy companies can benefit from rising AI loads.
- Weaknesses
- Project returns depend on electricity prices, approvals, grid connection, and capital investment.
- Comparison
- Compared with traditional energy demand, AI data center loads have higher growth visibility but are also more concentrated.
- Risks
- Grid access constraints, policy opposition, construction delays, and energy price volatility.
- META, GOOGL, MSFT, AMZNHave scale advantages and may achieve attractive ROI from AI capex
- Strengths
- The report explicitly states Overweight on META, GOOGL, MSFT, and AMZN, and is positive on their capacity expansion and AI monetization capabilities.
- Weaknesses
- Capex is large in scale, and investors will continue to focus on the pace of AI revenue realization.
- Comparison
- Compared with smaller AI companies, hyperscale cloud providers have scale, customer bases, and compute deployment capabilities.
- Risks
- AI monetization falling short of expectations, compressed returns on capex, and regulatory and competitive pressures.
Key data
- Median current monthly token spending by enterprise users< $11/monthThe report cites Ramp estimates to show that the current base of enterprise token spending is very low, leaving substantial room for future growth.
- Average cost savings per enterprise AI use case$55The report states that a set of enterprise AI use cases can save about $55 per execution on average.
- Estimated token cost to execute the same economic task$2-5Based on assumptions of $3-10 per million tokens, 5-7 AI agents, and 375,000-525,000 tokens.
- Blackwell data center token sales marginAbout 60%Includes data center costs, but excludes LLM development costs.
- Rubin and Feynman data center token sales marginsAbout 80% and 90%The report argues that more advanced GPU generations can significantly improve the economics of token sales.
- Potential token price reduction from subsequent NVIDIA GPU generationsAbout 75%The report says that from Blackwell to Feynman, hyperscale cloud providers could reduce token prices by 75% while achieving the same margin in leased data centers.
- U.S. data center capacity shortfall riskCould exceed 10% by 2028Relative to the data center capacity required under global semiconductor analysts' forecasts for AI chip sales.
- Expected AI capex of five major companies$1.2tr/$1.4trThe report cites U.S. internet analysts' views that capex will continue to rise.
- Available compute capacity of hyperscale cloud providersClose to 120GW by 2028A sharp increase versus about 30GW in 2025.
- Available compute capacity of AWS and GOOGL in 2028AWS about 35GW, GOOGL about 31GWThe report expects GOOGL to add the most compute capacity in 2027/2028, but AWS to have the highest available compute capacity in 2028.
- META total capacity forecast14GW in 2027, 21GW in 2028A significant increase versus about 3.5GW at the end of 2025.
Impact & implications
The investment implication is that the recent pullback in AI infrastructure stocks can be viewed as a thematic entry window, but investors should prioritize assets that genuinely own bottleneck resources or benefit from easing bottlenecks. The report's preferred areas include labor and power bottlenecks, the compute manufacturing ecosystem, leading Chinese AI solution providers, energy security and storage assets, and hyperscale cloud providers with scale advantages that can earn attractive returns on AI capex.
Risks
- Caps on token usage by enterprises or governments could suppress revenue for some LLM developers, although the report believes current data do not support this concern.
- Progress in Chinese LLMs poses a real competitive threat to U.S. frontier model developers.
- Data center construction faces shortages of skilled labor, especially electricians, welders, and pipefitters.
- Grid access is becoming increasingly difficult and may limit the pace at which data centers come online.
- Political resistance to data center growth is rising across multiple U.S. states and jurisdictions.
- The probability of AI policy intervention is rising in both the U.S. and China, which could further fragment the global AI market.
- If recursive self-improvement is achieved, it could bring enormous benefits but also intensify risks of AI disruption, misuse, and weaponization.
What to watch
- AI monetization data points from AMZN, META, and GOOGL in the upcoming earnings season.
- Whether capex by the five major hyperscale cloud providers continues to rise, and whether new compute capacity can come online on schedule.
- The progress of GW-scale capacity expansion by AWS, GOOGL, and META into 2027/2028.
- Changes in the performance, pricing, and global competitiveness of Chinese models such as Kimi K3, GLM, Qwen, and MiniMax.
- Trends in U.S. data center power shortages, grid-interconnection queues, and local moratoriums.
- Whether token usage intensity, employee substitution, and ROI evidence continue to improve in enterprise AI adoption surveys.
- Whether AI policy moves toward "two worlds," especially U.S. restrictions and China's market access policies.