U.S. Internet—AI Labs and Hyperscale Cloud Infrastructure Report Interpretation

Overview

Using hypothetical frontier AI Labs A and B as examples, the report breaks down how subscriptions, direct APIs, indirect APIs, training costs, inference costs, and cloud partnership revenue sharing flow through the income statement. It then traces how AI lab spending converts into revenue and operating profit for hyperscale cloud providers. The core conclusion is that underlying product unit economics are similar, but business mix, partnership terms, and accounting recognition methods can create significant apparent differences.

Core views

The report first notes that as AI labs begin reporting GAAP financial data, simple comparisons of revenue and gross margins may lead to misjudgments. The underlying unit economics of comparable products are broadly similar, but labs differ in their subscription and API mix, whether indirect APIs are recognized on a gross or net basis, revenue-sharing arrangements with hyperscale cloud providers, and whether training costs are classified as cost of sales or R&D expenses. Drawing an analogy to Uber and Lyft's different definitions of gross bookings and net revenue, the report argues that differences in reported revenue do not necessarily correspond to differences in actual business progress. Because it remains unclear how training costs will ultimately be allocated in the income statement, Barclays assumes in its model that approximately 10% of total training costs relate to final model training and are included in cost of sales, while the remainder is primarily reflected as R&D training costs. Business mix is the primary source of margin differences. For every $100 of revenue in 2026, hypothetical Lab A generates $30 from subscriptions, $45 from direct APIs, and $25 from indirect APIs, meaning APIs account for approximately 70% in total. Lab B generates $80 from subscriptions and $10 each from direct and indirect APIs. Lab A incurs $5 in indirect cloud fees and $30 in inference compute costs, with no strategic partner revenue share, resulting in paid inference profit of $65 and a margin of 65%. After deducting $10 in final model training costs, gross profit is $55 and gross margin is 55%. Lab B incurs $2 in indirect cloud fees, $30 in inference compute costs, and a $20 strategic partner revenue share, resulting in paid inference profit of $48 and a margin of 48%. After including final training costs, gross profit is $38 and gross margin is 38%. The modeled R&D training costs are $70 for Lab A and $96 for Lab B, also illustrating how training investment and revenue mix can significantly affect overall income statement performance. The improvement in 2026 is substantial compared with 2025. In 2025, Lab A's paid inference margin was only 14%, while its gross margin after training costs was 4%; the corresponding figures for Lab B were 18% and 8%. By 2026, paid inference margins had risen to 48%–65%, while gross margins after training costs had risen to 38%–55%, representing a year-over-year improvement of approximately 30–50 percentage points in adjusted gross margins. The report attributes this improvement to agentic workflows and enterprise products becoming must-have offerings in the market: enterprise and API customers typically generate higher inference margins, while a reduction in the number of tokens required by models to complete the same task also lowers unit service costs. Direct APIs were the earliest business model for most frontier labs, with developers paying based on token usage and embedding model capabilities into products such as Cursor or Figma. Barclays believes that improvements in model token efficiency and higher API list prices have jointly increased AI lab margins. On the infrastructure side, quantization, speculative execution, and next-generation compute resources improve model serving efficiency, thereby enhancing cloud provider margins. The report states that the current direct API inference margin in the second quarter of 2026 is estimated at above 80%, even higher than the level shown in the illustration, but believes this level may normalize downward in the future, although the timing remains unclear. Strategic partner revenue sharing transfers some profit from AI labs to cloud providers: in the example of every $100 spent by a direct API or subscription customer, AI lab inference profit is approximately $70, with a 70% margin, when there is no revenue share; with a 20% strategic partner revenue share, profit is approximately $50, with a 50% margin. Correspondingly, the cloud provider earns $30 in revenue and $9 in profit, with a 30% operating margin, when there is no revenue share; with revenue sharing, it can earn $50 in revenue and $29 in profit, with a 58% margin. However, excluding revenue-sharing fees, the actual underlying infrastructure profit per token may be similar. The end-user experience for indirect APIs is similar to that of direct APIs, but customer relationships and settlement are managed by hyperscale cloud providers. Lab A recognizes indirect API revenue on a gross basis, and this business is accounting for a growing share; Lab B generally recognizes it on a net basis after deducting cloud fees and may not recognize any revenue at all for indirect APIs hosted on its strategic partner's cloud. In a model based on every $100 of customer spending, indirect cloud fees are assumed to account for approximately 20%. Lab A recognizes $100 in gross revenue and, after deducting $20 in cloud fees and $30 in inference costs, generates paid inference profit of $50, with a 50% margin. When using other partners, Lab B recognizes $80 in net revenue and, after deducting a $16 revenue share and $30 in inference costs, generates $34 in profit, with a margin of approximately 43%. When operated directly by the strategic partner, Lab B may recognize zero revenue. The corresponding cloud provider revenue can reach $50, or $100 when operated directly by the strategic partner, with illustrative profits of $20 and $70, respectively. Consequently, as the share of indirect APIs increases, differences between gross and net recognition will increasingly distort comparisons of reported revenue scale across labs. Across the industry value chain, Barclays estimates that approximately $35–$40 of every $100 in AI lab revenue in 2026 flows into hyperscale cloud provider revenue and contributes nearly $10–$20 in operating profit, corresponding to a high operating margin of approximately 35%–45%. Partnership revenue sharing raises cloud providers' accounting margins, but after excluding the revenue share, underlying profit per token may not differ to the same extent. The report believes actual margins for AI labs and cloud providers in 2026 may be even higher than the examples suggest. However, as competition among frontier models intensifies and compute scarcity eases, margins are expected to decline gradually. Current training spending remains high relative to inference spending, so the report states that nearly every dollar of AI lab ARR ultimately flows into hyperscale cloud provider revenue. As inference profit grows, AI labs' inference profit should eventually exceed training costs, further improving their own margins. Over the next two years, AWS, Azure, and GCP are expected to continue controlling similar proportions of AI lab compute spending. However, beginning in 2028, traditional hyperscale cloud providers may start losing their share of training and inference compute as backstopped AI infrastructure projects come online and become the preferred option for labs. Lab B's strategic partner revenue share is expected to be eliminated after reaching a cumulative cap, and the report believes it may fall to zero after 2028.

Analysis framework

Barclays first standardizes different AI labs on the basis of every $100 of customer spending or revenue, separately breaking down subscriptions, direct APIs, and indirect APIs, and then deducting cloud fees, partner revenue sharing, inference compute, and final training costs item by item. The report subsequently compares changes in margins between 2025 and 2026 and analyzes the effects of business mix, token efficiency, pricing, and accounting recognition. Finally, it maps AI lab costs back into hyperscale cloud provider revenue and operating profit and discusses changes in infrastructure share around 2028.

Methodology notes

• Company Fundamentals and Financial Framework

  Unit Economics Model per $100 of Revenue

  The report standardizes the economics of labs and cloud providers on the basis of every $100 of customer spending or lab revenue to compare how different products, partnership structures, and cost items affect profit.

• Industry/Sector Analysis FrameworkVolume-price decomposition

  Decomposition of Business Mix, Token Efficiency, and API Pricing

  The report decomposes margin improvement into factors including a higher share of enterprise and API business, fewer tokens required to complete tasks, and increases in API list prices.

• Industry/Sector Analysis FrameworkUpstream-Midstream-Downstream Value Chain Transmission

  Transmission of AI Lab Revenue to Hyperscale Cloud Providers

  The report tracks how training, inference, indirect API fees, and revenue sharing convert from items on AI labs' income statements into cloud provider revenue and operating profit.

• Company Fundamentals and Financial Framework

  Comparison of Gross and Net Revenue Recognition

  The report separately models scenarios in which indirect APIs are recognized on a gross basis, recognized on a net basis after deducting cloud fees, or not recognized at all, illustrating that reported revenue does not necessarily directly reflect relative business progress.

• Industry/Sector Analysis FrameworkSupply-demand framework

  Impact of Compute Scarcity and Competition on Margins

  The report believes current compute scarcity supports high margins, but as compute supply increases and competition among frontier models intensifies, margins for AI labs and cloud providers may gradually decline.

Asset mapping & comparison

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

• Hypothetical Frontier AI Lab A

  Represents a lab model with a high API share, gross recognition of indirect API revenue, and no strategic partner revenue sharing.

  Strengths

  A 65% paid inference margin and a 55% gross margin after training costs in 2026, both higher than Lab B.

  Weaknesses

  Gross recognition of indirect API revenue inflates reported revenue, and modeled R&D training costs are $70.

  Comparison

  APIs total $70 per $100 of revenue, higher than Lab B's $20.

  Risks

  Intensifying competition and easing compute scarcity may cause current high margins to decline.

• Hypothetical Frontier AI Lab B

  Represents a lab model with a high subscription share, a 20% strategic partner revenue share, and primarily net recognition of indirect API revenue.

  Strengths

  Subscriptions account for 80% of revenue, and the revenue share is expected to potentially be eliminated after 2028 once the cumulative cap is reached.

  Weaknesses

  A 48% paid inference margin and a 38% gross margin after training costs in 2026, while indirect APIs on the strategic partner's cloud may not be recognized as revenue at all.

  Comparison

  Compared with Lab A, reported revenue and margins are more readily constrained by the subscription mix, net recognition, and partner revenue sharing.

  Risks

  Uncertainty remains regarding the timing of the elimination of revenue sharing and the income statement classification of training costs.

• Amazon.com, Inc. (AMZN) / AWS

  As a hyperscale cloud platform supporting AI lab training and inference, it captures lab compute spending.

  Strengths

  The report expects AWS to continue controlling a proportion of AI lab compute spending similar to Azure and GCP over the next two years.

  Weaknesses

  The report does not provide standalone unit economics data for AWS.

  Comparison

  Its share of AI lab compute spending is expected to be broadly similar to those of Azure and GCP over the next two years.

  Risks

  Beginning in 2028, it may lose share as backstopped independent AI infrastructure comes online.

• Microsoft (MSFT) / Azure

  As a hyperscale cloud platform supporting AI lab training, inference, and partnership revenue sharing, it captures related revenue and profit.

  Strengths

  Strategic partner revenue sharing can significantly raise a cloud provider's reported operating margin.

  Weaknesses

  Excluding revenue sharing, underlying profit per token may be similar to that under other cloud partnership models.

  Comparison

  Its share of AI lab compute spending is expected to be broadly similar to those of AWS and GCP over the next two years.

  Risks

  Beginning in 2028, it may face share pressure as alternative infrastructure comes online.

• Alphabet Inc. (GOOGL) / GCP

  As a hyperscale cloud infrastructure provider for AI lab training and inference, it captures a share of related compute spending.

  Strengths

  The report expects GCP to continue controlling a proportion of AI lab compute spending similar to AWS and Azure over the next two years.

  Weaknesses

  The report does not provide standalone revenue or margin data for GCP.

  Comparison

  Its near-term share of AI lab compute spending is expected to be broadly similar to those of AWS and Azure.

  Risks

  Independent AI infrastructure projects may divert training and inference demand after 2028.

Key data

• Lab A 2026 Revenue MixSubscriptions $30, direct API $45, indirect API $25, per $100 of revenueAPIs account for approximately 70% in total, representing a lab with a high API share
• Lab B 2026 Revenue MixSubscriptions $80, direct API $10, indirect API $10, per $100 of revenueSubscriptions account for approximately 80%, and the lab bears a strategic partner revenue share
• 2026 Paid Inference MarginLab A 65%; Lab B 48%14% and 18%, respectively, in 2025
• 2026 Gross Margin Including TrainingLab A 55%; Lab B 38%4% and 8%, respectively, in 2025, representing a year-over-year improvement of approximately 30–50 percentage points
• Final Model Training Cost AssumptionApproximately 10% of total training costsThe report includes the final training run component in cost of sales, while the remaining training investment is primarily reflected as R&D costs
• Current Direct API Inference MarginAbove 80%Estimated for the second quarter of 2026; the report expects it may normalize downward in the future
• Strategic Partner Revenue Share20%Borne by Lab B until the cumulative cap is reached and expected to potentially fall to zero after 2028
• Indirect API Cloud Fee AssumptionApproximately 20% of customer spendingGross versus net recognition results in materially different reported revenue for labs
• AI Lab Revenue Flowing to Cloud ProvidersApproximately $35–$40 per $100 of revenue2026 estimate
• Cloud Provider Operating Profit ContributionNearly $10–$20Per $100 of AI lab revenue, corresponding to an operating margin of approximately 35%–45%
• Timing of Existing Cloud Provider Share ChangesBeginning in 2028Traditional hyperscale cloud providers may start losing share after backstopped AI infrastructure projects come online

Impact & implications

The report argues that investors cannot assess relative competitive progress based solely on the ARR, revenue scale, or overall gross margins disclosed by AI labs. They must first standardize business mix, gross versus net recognition, training cost classification, and partner revenue-sharing treatment. In 2026, enterprise demand and improvements in model efficiency simultaneously enhanced the economics of labs and cloud providers, but the allocation of profit between the two depends on partnership terms. AWS, Azure, and GCP should continue benefiting from training and inference spending over the medium term, while the expansion of independent AI infrastructure after 2028 may alter this landscape.

Risks

• Intensifying competition among frontier models and easing compute scarcity may cause the currently high margins of AI labs and hyperscale cloud providers to decline gradually.
• The classification of training costs between cost of sales and R&D expenses remains unclear, potentially making gross margins across AI labs not directly comparable.
• Differences in whether indirect APIs are recognized on a gross basis, net basis, or not recognized at all may distort the market's assessment of labs' relative business progress.
• Lab B remains burdened by a 20% strategic partner revenue share until the cumulative cap is reached, and the precise timing of its elimination is uncertain.
• Backstopped AI infrastructure projects are expected to divert training and inference demand beginning in 2028, potentially reducing the share of existing hyperscale cloud providers.

What to watch

• Whether the share of enterprise and API business can continue to increase and whether agentic workflows continue to drive high-margin demand.
• Changes in the number of tokens required by models to complete individual tasks, API list prices, and infrastructure service efficiency.
• When and by how much the direct API inference margin of above 80% in the second quarter of 2026 will normalize.
• The specific gross, net, or non-recognition policies adopted by each lab as the share of indirect APIs rises.
• When Lab B's strategic partner revenue share reaches its cumulative cap and falls to zero, which the report expects may occur after 2028.
• The deployment progress of backstopped AI infrastructure projects around 2028 and changes in the compute shares of AWS, Azure, and GCP.