The pullback in AI infrastructure provides a positioning window, but supply bottlenecks and policy divergence are the main risks

Overview

Following the recent sharp volatility in AI infrastructure stocks, this report reassesses investors' core concerns about Chinese models, token spending constraints, political resistance to data centers, model blockades, and physical bottlenecks. The conclusion is that the long-term fundamentals of AI infrastructure remain strong, and the recent decline reflects technical pressure more than deterioration in demand or return logic.

Core views

The core views include: first, improvements in AI capabilities are nonlinear, and the value of computing power and "intelligence" will continue to rise; second, the cost-benefit ratio of enterprise AI applications remains highly attractive, with token costs representing only a small portion of labor-saving value; third, the ROI of AI capital expenditure remains attractive, and advanced GPU iterations are expected to reduce the per-token price while maintaining high margins; fourth, Jevon's Paradox implies that improvements in model and chip efficiency do not necessarily reduce total computing consumption, and may instead expand use cases, complexity, and frequency, driving higher total computing and power demand.

Analysis framework

The report uses a thematic investment framework, dividing the AI infrastructure value chain into investable directions such as bottleneck resources, computing-power manufacturing, China AI solutions, energy security, and Hyperscalers, while combining a token economics model, U.S. power gap analysis, data center resistance tracking, enterprise adoption surveys, consumer surveys, an AI stock-mapping database, and Future of Work analysis to assess risks and opportunities.

Methodology notes

• Demand and efficiencyJevon's Paradox

  Efficiency improvements may expand total demand

  The report applies Jevon's Paradox to AI: more efficient models and chips reduce the cost of generating each unit of intelligence, which may stimulate more users, higher frequency, and more complex tasks, thereby increasing total computing and power demand.

• Unit economics modelToken Economics Model

  Estimating returns for enterprise AI users and computing-power suppliers

  The report uses enterprise task cost savings, token consumption, token prices, and data center costs to estimate the economics for AI adopters and infrastructure owners, concluding that both enterprise AI usage and AI capex ROI are attractive.

• Supply constraintsTime to Power Analysis

  The speed of power access determines data center expansion capacity

  The report includes grid interconnection, gas turbines, fuel cells, nuclear sites, and Bitcoin site conversions in its assessment, emphasizing that "time to power" is an important source of Alpha in AI infrastructure.

• Policy and market structureTwo Worlds of AI Policy and Adoption

  U.S. and China AI policy may lead to market divergence

  The report believes the probability of AI policy intervention is rising in both the U.S. and China, which may further divide global AI access, market entry, and adoption pathways.

Asset mapping & comparison

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

• AI infrastructure bottleneck assets

  Directly benefit from shortages in computing supply and constraints on data center expansion

  Strengths

  Labor, time to power, fuel cells, turbines, energy storage, power developers, and data center REITs can gain pricing power as bottlenecks are alleviated.

  Weaknesses

  Long construction cycles, constrained by local approvals, grid interconnection, and shortages of construction workers.

  Comparison

  Compared with pure model application plays, bottleneck assets are more directly exposed to undersupply.

  Risks

  Political resistance, project delays, failure of power access, and cost overruns.

• Computing-power manufacturing ecosystem

  Benefits from the rising value of intelligence and demand for computing power exceeding supply

  Strengths

  GPUs, TPUs, and related manufacturing chains occupy a core position in the AI capex cycle.

  Weaknesses

  Valuations and order expectations are sensitive to cyclical fluctuations.

  Comparison

  Compared with downstream applications, the manufacturing ecosystem is closer to AI infrastructure capital expenditure.

  Risks

  Changes in chip generations, supply-chain constraints, and shifts in customer capex cadence.

• China AI solution providers

  Benefit from improving Chinese LLM capabilities and cost competitiveness

  Strengths

  Product capabilities and price competitiveness are improving, and some stocks may not yet fully reflect the related upside.

  Weaknesses

  Uncertainty remains around global market access, policy restrictions, and commercialization pathways.

  Comparison

  Compared with U.S. frontier LLMs, Chinese models may offer lower costs and stronger adaptation to local ecosystems.

  Risks

  U.S.-China policy divergence, export restrictions, insufficient validation of model capabilities, and intensifying competition.

• Energy security assets

  AI infrastructure expansion depends on power and energy security

  Strengths

  Energy storage is specifically highlighted by the report as a standout asset class, and the increasing scarcity of power resources raises the value of related assets.

  Weaknesses

  Project returns are affected by electricity prices, regulation, and construction progress.

  Comparison

  Compared with traditional technology assets, energy security assets provide upstream supply elasticity for AI infrastructure.

  Risks

  Regulatory changes, grid bottlenecks, equipment supply, and financing costs.

• META、GOOGL、MSFT、AMZN

  Hyperscalers with scaled AI capex and commercialization capabilities

  Strengths

  Scale advantages, expanding computing capacity, and the ability to convert AI into revenue support capex ROI; the report says Morgan Stanley is Overweight on these stocks.

  Weaknesses

  Large-scale capital expenditure may pressure short-term profits and free cash flow.

  Comparison

  Compared with smaller AI application companies, Hyperscalers are better able to bear infrastructure investment and generate scale benefits.

  Risks

  AI monetization below expectations, regulatory pressure, lagging capex returns, and intensifying competition.

Key data

• Median current enterprise token spending< $11/monthRamp estimates show that current token spending by enterprise users remains very low, and the report believes there is substantial room for future growth.
• Average cost savings per enterprise AI use case$55The report compares this with the token cost of AI Agents performing tasks to illustrate the economics of enterprise AI adoption.
• Estimated token usage for an AI Agent task375,000-525,000 tokensAssumes 5-7 AI Agents perform one economic task, with each Agent using roughly 15 times the tokens of a standard LLM query.
• Estimated token cost to execute a task$2-5Based on an average price of $3-10 per million tokens, the cost is significantly lower than the $55 benefit.
• Blackwell data center token sales marginapproximately 60%The report says this estimate includes data center costs, but excludes LLM development costs.
• Rubin and Feynman data center token sales marginsapproximately 80% and 90%More advanced GPU generations may improve margins and support lower token prices.
• Potential token price reduction in subsequent NVIDIA GPU generationsapproximately 75%The report says that from Blackwell to Feynman, Hyperscalers could reduce token prices by about 75% while maintaining the same leased data center margin.
• Expected AI capex of five major companies$1.2tr/$1.4trBrian Nowak expects capex at the five major companies to continue rising.
• Available Hyperscaler computing capacityapproximately 30GW in 2025 to 120GW by 2028The report forecasts that available Hyperscaler computing capacity will approach 120GW by 2028.
• Risk of U.S. data center capacity shortfall>10% by 2028Relative to the required capacity implied by global semiconductor analysts' AI chip sales forecasts, U.S. data center supply may face a shortfall of more than 10%.

Impact & implications

The investment implication is that the AI infrastructure pullback can be viewed as a window to position for long-term supply-demand imbalance. It is more attractive to focus on companies that can alleviate bottlenecks or possess critical resources, including power, energy storage, data center REITs, computing-power manufacturing, leading China AI solutions, and Hyperscalers with scale advantages; however, investors also need to assess volatility arising from data center approvals, grid interconnection, labor shortages, policy divergence, and model competition.

Risks

• Data center construction is constrained by labor, power, and political factors.
• Opposition to data center growth is rising in multiple U.S. states and regions.
• U.S. and China AI policies may move toward more obvious divergence, affecting market access and global adoption pathways.
• Progress in Chinese LLMs poses a real competitive threat to U.S. frontier LLM developers.
• If recursive self-improvement is achieved, it could bring enormous benefits but also intensify risks of AI disruption, misuse, and weaponization.
• If AI monetization falls short of expectations, returns on Hyperscalers' large-scale AI capex may be questioned.

What to watch

• Data points in upcoming Hyperscaler earnings reports regarding AI capex, computing capacity coming online, and AI revenue monetization.
• Changes in U.S. data center power access, local approvals, and moratoriums.
• The evolution of performance, pricing, and global competitiveness of Chinese LLMs such as Kimi K3 and the GLM series.
• The impact of advanced GPU generations on token costs, data center margins, and the pace of AI adoption.
• Changes in token spending, job substitution, and use-case ROI in enterprise AI adoption surveys.
• Further developments in U.S. and China AI policy regarding access restrictions, model blockades, market access, and industry support.