AI investment enters a period of bull-bear divergence: demand, CapEx, and unit economics become the core conflicts

Overview

This report summarizes an AI bull-bear debate hosted by Jefferies, focusing on whether AI infrastructure buildout remains in the early stages of a long-term growth cycle or has evolved into a bubble supported by hyperscale CapEx and demand from a small number of frontier labs. The report does not provide a rating for any single stock; instead, it discusses AI computing, data centers, power bottlenecks, cloud-provider capital expenditures, software revenue conversion, valuation, and potential risk catalysts from a strategy perspective.

Core views

The bull case, presented by Brent Thill, argues that the AI cycle remains in its early stages. Insufficient computing supply, data-center power and labor bottlenecks, and the gap between approximately $2 trillion of contracted backlog and approximately $784 billion of deployed CapEx support continued expansion in AI infrastructure demand. AI revenue still accounts for less than 5% of software companies’ total revenue, suggesting that enterprise adoption is only beginning; through 2026, capital is more likely to continue flowing to the “shovel sellers” across semiconductors, networking, hardware, and the data-center supply chain. The bear case, presented by Ed Zitron, argues that AI capital expenditures depend heavily on a small number of loss-making frontier labs, with two labs consuming 70% to 85% of AI computing capacity. Unit economics are difficult to establish, and enterprise ROI still lacks quantifiable evidence. If leading cloud providers cut CapEx, the impact could spread across chips, memory, and Asian assembly chains.

Analysis framework

The report uses a bull-bear debate rather than a single-directional assessment. The bull case analyzes AI industry-chain opportunities through supply-demand gaps, contract backlogs, the early stage of enterprise adoption, valuation multiples, and capital flows; the bear case analyzes systemic risks through computing-demand concentration, training and inference costs, mismatches between subscription prices and token costs, open-source model price declines, circular financing, and potential cloud-provider CapEx reductions.

Methodology notes

• Scenario analysisBull-bear scenario comparison

  Assess the returns and risks of the AI industry chain simultaneously through optimistic and pessimistic scenarios.

  The bull case emphasizes supply shortages and long-term adoption potential, while the bear case emphasizes overheated CapEx and flawed unit economics. This framework is suitable for assessing the risk-return distribution of thematic investments.

• Industry-chain analysisShovel-seller framework

  Prioritize suppliers providing infrastructure, chips, networking, and hardware for AI buildout.

  The report argues that capital may continue flowing to semiconductor and infrastructure suppliers against a backdrop in which approximately 70% of software-covered companies may slow in 2026.

• Capital-expenditure cycleHyperscaler CapEx transmission

  Changes in cloud-provider capital expenditures will transmit to chips, memory, data centers, and assembly chains.

  The bear case argues that if leading hyperscalers cut CapEx, a chain reaction could trigger adjustments across the AI hardware and data-center supply chain.

• Valuation analysisRelative valuation multiples

  Use multiples such as P/E and EBITDA to assess whether AI-related assets are excessively priced.

  The bull case notes that some semiconductor companies trade at multiples in the teens, MSFT at approximately 20x earnings, and AMZN at approximately 11x EBITDA, arguing that these do not indicate extreme overvaluation given the current growth trajectory.

Asset mapping & comparison

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

• Semiconductors and AI hardware

  Core upstream beneficiaries of AI computing buildout

  Strengths

  Benefit from tight computing supply, GPU demand, data-center construction, and capital flows to the “shovel sellers.”

  Weaknesses

  If actual demand is primarily generated by a small number of loss-making labs, hardware orders could be highly sensitive to a CapEx pullback.

  Comparison

  Compared with software, the bull case believes capital is more likely to continue favoring semiconductors and infrastructure through 2026.

  Risks

  Hyperscaler CapEx cuts, overly high GPU sales expectations, and simultaneous downward revisions across memory and Asian assembly chains.

• Data centers and power infrastructure

  Bottleneck and investment-transmission segment in AI buildout

  Strengths

  Approximately 40% of data centers are delayed by power interconnection and labor shortages, indicating that near-term supply bottlenecks support construction demand.

  Weaknesses

  Construction cycles lasting 18 to 36 months or longer increase project payback-period and utilization risks if demand expectations change.

  Comparison

  Compared with pure software, data centers have more direct exposure to the CapEx cycle and power constraints.

  Risks

  Local opposition, power-interconnection delays, supply-chain constraints, and slower cloud-provider investment.

• Cloud computing and hyperscalers

  Core platforms for AI-computing purchases, resale, and capital expenditures

  Strengths

  The bull case believes major cloud providers have no sellable AI computing capacity over the next six months and that AI margins may be higher than those of traditional cloud businesses.

  Weaknesses

  The bear case argues that a portion of cloud providers’ future contracted revenue comes from labs they themselves fund, creating concerns over circular financing and the authenticity of demand.

  Comparison

  Cloud providers are simultaneously beneficiaries of AI infrastructure and bearers of CapEx risk.

  Risks

  Leading cloud providers cutting CapEx, insufficient customer ROI, open-source models lowering prices, and on-premise migration weakening recurring cloud revenue.

• Enterprise software

  Downstream segment for converting AI applications into revenue

  Strengths

  AI revenue accounts for less than 5% of total revenue at key software companies, leaving substantial long-term upside if enterprise adoption expands.

  Weaknesses

  Approximately 70% of software-covered companies may slow in 2026, while enterprise deployment is constrained by data governance and access to private data.

  Comparison

  Compared with hardware and infrastructure, software depends more heavily on AI-feature commercialization and enterprise ROI validation.

  Risks

  Insufficient AI revenue contribution, inability to quantify ROI, and disruption to incumbent software vendors from frontier labs or AI-native products.

Key data

• Contracted backlog demand at hyperscalersApproximately $2 trillionUsed by the bull case to demonstrate that contracted demand is significantly higher than deployed CapEx.
• Deployed CapExApproximately $784 billionCompared with approximately $2 trillion of contracted backlog demand.
• AI share of total revenue at key software companiesBelow 5%The bull case believes enterprise AI adoption remains in its early stages.
• Data-center delay rateApproximately 40%Delays include power interconnection and labor shortages.
• Share of computing capacity consumed by leading AI labs70% to 85%The bear case believes AI computing demand is overly concentrated.
• Recent financing by the two largest computing-capacity consumersApproximately $200 billion, within six monthsThe bear case points out that they remain deeply loss-making.
• Mismatch between monthly user fees and token consumption costsA $200-per-month plan may consume $8,000 to $14,000 in monthly token costsUsed by the bear case to illustrate unit-economic pressure.
• Forward computing commitmentsApproximately $1.1 trillionThe bear case believes open-source model price declines will pressure labs carrying large computing commitments.
• Portion of the top three cloud providers’ future contracted revenue related to labs they fundApproximately $748 billionThe bear case argues that circular financing could amplify risks.
• Potential GPU sales and annualized computing-revenue requirementApproximately $1 trillion in GPU sales and approximately $400 billion in annualized computing revenueThe bear case believes chip guidance implies extremely high requirements for actual demand.

Impact & implications

For portfolios, the report suggests that the AI theme is not simply a one-sided bullish or bearish trade. If the bull scenario prevails, semiconductors, networking, data-center equipment, power-related infrastructure, and cloud-infrastructure suppliers may continue to benefit from supply shortages and contract backlogs. If the bear scenario prevails, the most vulnerable areas may be the GPU, memory, cloud-computing, and data-center chains that depend heavily on demand from a small number of frontier labs, circular financing, and continued CapEx expansion. The key variable for software is whether enterprise AI revenue can convert from pilots and narratives into measurable ROI.

Risks

• AI computing demand is overly concentrated among a small number of deeply loss-making frontier labs.
• Mismatches between subscription prices and token consumption costs make unit economics difficult to establish.
• Enterprise AI investment may fail to link to measurable business outcomes, leaving ROI insufficiently validated.
• Lower-priced open-source or open-weight models may compress pricing for frontier models and cloud computing.
• Hyperscaler CapEx cuts could spread across chips, memory, data centers, and Asian assembly chains.
• Circular financing, neoclouds, and private-credit structures could amplify downside risks in AI infrastructure.
• Power interconnection, labor shortages, construction cycles, and local opposition at data centers could delay supply delivery.

What to watch

• Potential IPOs by Anthropic and OpenAI in the second half of 2026 through the first half of 2027, along with their financial disclosures.
• Whether major hyperscalers’ AI CapEx guidance continues to rise or begins to slow.
• AI computing utilization, customer mix, and whether demand spreads beyond frontier labs.
• Whether AI revenue as a share of enterprise software companies’ revenue rises materially from below 5%.
• Comparability between cloud providers’ AI gross margins and traditional cloud-business margins.
• The impact of declining open-source model costs on frontier-model pricing and cloud-computing demand.
• Data-center power interconnection, construction delays, and local regulatory resistance.