Generative AI revenue has entered a scaled phase, but capex payback remains the key test

Overview

This report attempts to solve the demand visibility problem in the AI economy. Semiconductor companies and hyperscalers disclose the supply side relatively well, but it has remained difficult to judge which AI products customers are actually paying for, whether the revenue is real, and whether there is double counting. The report breaks down global ex-China generative AI revenue into the application, model, and infrastructure layers, emphasizing that realized revenue is a better validator of demand than supply-side indicators alone.

Core views

The core views include: first, generative AI demand is real, large in scale, and growing rapidly, with revenue growth about three times faster than previous IT waves; second, the AI economy is still in an early stage relative to GDP and corporate profits, and many social welfare gains and free digital benefits may not be fully captured in GDP; third, AI capital spending by hyperscalers and neo-cloud providers can currently cover part of depreciation through revenue, but the coverage cushion is thin; fourth, tokens may become an important unit of account for the AI economy, with falling prices driving elastic demand and higher usage; fifth, value capture is moving toward the application and model layers, but pricing power depends on competitive pressure rather than simply on position in the technology stack.

Analysis framework

The report uses company-level bottom-up revenue models, combining public filings, audited accounts, company disclosures, earnings calls, government statistics, third-party research, and traffic and capacity proxy indicators, while assigning a confidence score to each revenue line. Revenue is identified at every layer, but is ultimately deduplicated based on value added, so the same customer spending is not added repeatedly across the application, model, and hosting layers.

Methodology notes

• Revenue accountingValue-added deduplication model

  Count each dollar only once

  For example, if a customer pays $100 to an application, of which $60 flows to the model supplier and $30 then flows to the inference hosting provider, the report counts $100 as ecosystem revenue rather than summing the three layers to $190.

• Data qualityConfidence-tiering

  Score first, then model

  Reported financial data receives the highest weight, followed by other primary sources, cross-verifiable third-party estimates, and single-source claims; derived figures inherit the lowest quality tier among their input sources.

• Modeling methodCompany-level financial models and triangulation

  Extract generative AI revenue from company financials

  The report builds dedicated generative AI financial models for relevant companies and cross-checks them using independent proxy indicators such as chip revenue, construction costs, segment structure, industry research, traffic, and capacity.

• Research scopeGlobal ex-China AI ecosystem revenue definition

  Include application, model, and infrastructure revenue while excluding some adjacent items

  The report scope includes application, model, and AI infrastructure revenue; it excludes chip sales, ad revenue uplift driven by AI, traditional software feature upgrades, and financing itself.

Asset mapping & comparison

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

• COREWEAVE INC (CRWV.US) 与 NEBIUS GROUP NV (NBIS.US)

  As representatives of neo-cloud and AI hosting infrastructure, they directly benefit from external customer demand for compute and GPU leasing.

  Strengths

  Hosting-layer revenue can be fully counted within the AI revenue framework, while demand, rental pricing, and backlog are important validation signals.

  Weaknesses

  They are highly capital-intensive and heavily dependent on GPU depreciation, financing costs, utilization, and long-term customer contracts.

  Comparison

  Compared with traditional cloud giants, neo-cloud providers offer purer exposure to the AI infrastructure cycle, but their financial flexibility and financing capacity are typically more tested.

  Risks

  If compute supply becomes excessive, rental prices for GPUs such as H100 decline, or external financing tightens, the room for revenue to cover depreciation could compress rapidly.

• ALPHABET INC (GOOGL.US)、ORACLE CORP (ORCL.US)、META PLATFORMS INC (META.US)

  The report views large cloud and platform companies as important participants in AI capex, cloud revenue, ad uplift, and data center construction.

  Strengths

  They possess customer bases, cloud platforms, data center resources, and capital strength, enabling them to serve model, application, and enterprise AI demand.

  Weaknesses

  Some AI benefits may be buried within segment revenue; for companies such as Meta, AI investment may show up more as advertising uplift or indirect monetization and may not be counted as pure generative AI revenue.

  Comparison

  Compared with neo-cloud providers, large platforms have stronger financing capacity, but their AI revenue transparency is lower and the relationship between capex and actual generative AI revenue is more complex.

  Risks

  If AI-driven ad uplift, assistant products, or cloud inference services fail to convert into measurable cash flow, high capex could drag on free cash flow and depreciation coverage.

• Semiconductors and the GPU supply chain

  Chip sales are excluded from the report’s deduplicated revenue definition, but GPUs are the core input for AI compute, data center investment, and token production.

  Strengths

  Expansion in AI server compute, hardware iteration from Hopper to Blackwell to Rubin, and growth in inference workloads support long-term demand.

  Weaknesses

  Chip revenue is not the same as end AI demand; if downstream players cannot recover investment through token and application revenue, supply-side strength may become mismatched.

  Comparison

  Compared with the application and model layers, semiconductors sit further upstream, with more direct revenue recognition, but they are more vulnerable to inventory, customer capex timing, and depreciation cycles.

  Risks

  Overbuilding, overly long assumptions for effective GPU life, or rising customer financing pressure could lead to order volatility.

• Electric utilities and data center infrastructure

  AI demand is reigniting U.S. electricity demand, and data centers have become an important source of incremental load growth.

  Strengths

  The report notes that expected incremental U.S. power demand through 2030 has been revised sharply upward, with data centers contributing about 55% of the growth, benefiting power, transmission and distribution, and campus infrastructure.

  Weaknesses

  Power interconnection, grid connection, construction timelines, and regulatory approvals may become bottlenecks to AI expansion.

  Comparison

  Compared with software and model companies, power assets grow more slowly and have longer capital cycles, but they can derive steadier infrastructure demand from AI compute needs.

  Risks

  If grid expansion lags data center construction, the pace of compute capacity coming online and cloud service revenue growth may be constrained.

• AI application layer and model layer

  The report argues that value is moving up toward applications and models, but revenue must be calculated on an incremental basis after subtracting costs passed through to the model and hosting layers.

  Strengths

  The application layer is closer to customer payment, while the model layer can capture usage growth through APIs, subscriptions, and token pricing.

  Weaknesses

  Open-weight models and rapid convergence in model performance will compress pricing power, and frontier capabilities from last year may commoditize quickly.

  Comparison

  The application layer may have stronger customer relationships, while the model layer has technical capabilities, but pricing power in both depends on the competitive landscape rather than stack position alone.

  Risks

  Rapid declines in token prices, open-source substitution, and customers prioritizing cost savings may reduce unit revenue and gross margins.

Key data

• Annualized generative AI ecosystem revenue>$175bnThis is the deduplicated global ex-China figure, which the report views as validating real demand.
• Generative AI revenue over the past 12 months$110bnThe report says revenue has reached about $110 billion over the trailing 12 months.
• Equity market value supported or influenced by the AI economy$22.7tnThe report argues that without understanding real demand, it is hard to judge the health of the AI economy supporting such a large stock market value.
• Cumulative capex of hyperscalers and neo-cloud providers$2tn through 2026EIncludes cumulative capital expenditure committed by hyperscalers and neo-cloud providers for AI infrastructure.
• Token usage>30Q/monthGlobal monthly token usage exceeds 30 quadrillion and has grown about 14x year over year.
• Year-over-year growth in token usage14x YoYAgentic workloads and inference models have increased token consumption.
• Change in expected incremental U.S. power demand by 203024GW→166GWThe report says that since 2022, the additional electricity the U.S. grid is expected to need by 2030 has risen about 7x.
• Contribution of data centers to U.S. incremental load growth~55%The report estimates that data centers account for about 55% of U.S. incremental load growth.
• Absorption of AI revenue by depreciation81% / 68%Depreciation absorbs about 81% of hyperscaler and neo-cloud generative AI revenue, and 68% of total generative AI revenue, before accounting for additional operating costs.
• Industry revenue per GW of data center capacity>$7bn/GWThe report believes efficiency gains are increasing the monetization capacity of each GW.
• Generative AI revenue versus corporate profits32xEven using corporate profits as a relatively lenient benchmark, corporate profit scale is still about 32 times total generative AI revenue, indicating that the AI economy remains at an early stage.
• Mentions of AI impact by S&P 500 companies3-4x rise since 2023The report tracks that the frequency of AI impact mentions in S&P 500 company earnings calls has risen about 3 to 4 times since 2023.

Impact & implications

The investment implication is that the AI theme cannot be evaluated only through GPU supply and cloud capex; investors also need to track real external revenue, deduplicated demand, contract backlog, depreciation coverage, token price elasticity, and power constraints. If revenue, usage, and utilization continue to compound, AI infrastructure buildout may still generate returns; if falling prices fail to drive sufficient incremental demand, or if power and data center construction are constrained, industry profit pools will be squeezed.

Risks

• Capex and depreciation pressure are too high; if revenue, utilization, and pricing cannot keep compounding, AI infrastructure payback periods may lengthen.
• Third-party financing entering AI infrastructure may amplify cyclicality risk, and changes in financing conditions may affect construction pace.
• Falling token prices require sufficiently strong demand elasticity to offset them; otherwise, total revenue and margins may come under pressure.
• Data center electricity use, grid connection, and equipment delivery may become physical bottlenecks to AI expansion.
• Private model companies disclose limited information, so revenue estimates rely on cloud attribution, proxy indicators, and confidence scoring, leaving measurement error risk.
• If application, model, and hosting layers are not strictly deduplicated, the size of the AI economy can easily be overstated.
• Open-weight models and rapid commoditization may weaken the pricing power of closed-source models and some applications.

What to watch

• Whether deduplicated generative AI revenue can continue growing from the current $175 billion annualized base.
• RPO, contract backlog, GPU rental pricing, and data center utilization at hyperscalers and neo-cloud providers.
• The depreciation base, depreciation coverage ratio, and free cash flow changes after AI capex is put into service.
• Monthly token usage, per-token pricing, the share of agentic workloads, and token price elasticity.
• Power demand for data centers in the U.S. and globally, grid-connection approvals, and progress in new power capacity construction.
• Quantified disclosures from S&P 500 companies on AI impact, especially whether cost savings are turning into revenue growth.
• Shifts in value capture across the application, model, and infrastructure layers, and the impact of open-weight models on pricing power.