AI Provider Revenue: Reconciling OpenAI, Anthropic and the Economics Behind Run Rates

Review period: March 17, 2016–October 13, 2026 Scope: Comparability of OpenAI and Anthropic revenue, channel accounting, gross margins, infrastructure obligations and valuation, with implications for the wider AI industry. Source policy: Prioritize company statements, SEC filings, accounting standards and government evidence; use original reliable reporting for private financial information, and label reconstructions, estimates and unresolved disclosures.

The commercial analysis focuses on calendar-year 2024–2025 results and 2026 disclosures. Earlier documents establish the accounting framework. All financial amounts are nominal U.S. dollars.

Executive summary

The headline revenue run rates are not presently an apples-to-apples measure of provider economics. Reuters reported on October 8, 2026 that OpenAI’s September annualized revenue was almost $50 billion, rather than the approximately $70 billion previously indicated. The discrepancy chiefly reflected an attempt to compare OpenAI with Anthropic’s broader inclusion of cloud-partner sales—not evidence, by itself, of a comparable decline in sales. Anthropic’s independently reported milestone was more than $65 billion at the end of July, a different measurement date. (investing.com)

Three findings matter most:

• Gross revenue, retained commercial proceeds and profit are different quantities. A provider can legitimately recognize gross sales while paying a marketplace commission; a licensing or reseller arrangement can instead produce revenue limited to the provider’s contractual share. Principal-versus-agent conclusions depend on control and contractual obligations, not a preference for a larger headline. (storage.fasb.org)
• Expense placement adds a second comparability problem. Reuters’s prospectus reporting says Anthropic places channel fees in sales, marketing and partnerships expense, below gross profit. Revenue normalization alone therefore does not produce comparable gross margins. (thestar.com.my)
• Presentation changes can materially alter growth comparisons and sales multiples, but cannot manufacture operating profit or cash. Actual inference costs, partner compensation, training expenditure and capacity obligations must be analyzed separately. The available evidence supports useful historical and conditional reconstructions, but not a fully comparable current-quarter revenue-and-margin ranking.

1. Verified figures: distinguish dates, measures and evidence quality

Revenue milestones

Two corrections are essential. First, an exit-month run rate is not revenue earned during the year. Second, Anthropic’s July milestone and OpenAI’s September milestone do not establish a same-month competitive ranking.

What is—and is not—publicly inspectable

Anthropic confirmed a confidential draft S-1 submission on June 1, 2026. Reuters subsequently reported financial information from its prospectus. OpenAI’s historical financial figures were transcribed by journalist Ed Zitron, who stated that the Financial Times independently verified the documents. This review obtained those reports, not complete provider financial statements and signed distribution contracts. Consequently, the figures are stronger than unsupported market estimates, but the providers’ detailed accounting judgments cannot be independently reproduced here. (anthropic.com)

2. What the revenue labels actually mean

The recognized-revenue distinctions follow the revenue standards; the analytical labels above are deliberately separate from accounting presentation. (ifrs.org)

Four economically different routes to an AI customer

1. Direct subscription or API sale. The model provider contracts with the customer and buys the infrastructure necessary to deliver the service.
2. Marketplace-facilitated sale. The provider supplies the service while the cloud marketplace handles procurement, billing or distribution and charges a fee.
3. Reseller or managed-cloud service. The cloud company sells its own contractual offering, potentially combining model access, compute, security and support. The model developer earns its agreed license fee or revenue share—not necessarily the reseller’s entire customer bill.
4. A separate partnership-wide revenue-sharing agreement. A developer may owe a partner compensation extending beyond sales routed through that partner’s marketplace.

These routes must not be collapsed into a single assumed commission rate. The FTC’s partnership study documented revenue-sharing rights and cloud-spending obligations in multiple forms. Microsoft’s February 2026 statement explicitly said its revenue-sharing arrangement included OpenAI partnerships with other cloud providers. (search.ftc.gov)

Important inference: Adding a cloud reseller’s gross sales to a developer’s revenue can expand the business perimeter being measured. It is not necessarily just reversing a marketplace commission. Comparable margins would then also require the associated delivery costs.

3. Principal versus agent: the accounting test, not a public-relations choice

The governing questions

Under ASC 606, the central question is whether the entity controls the specified good or service before transfer to the customer. A principal recognizes its gross entitlement; an agent recognizes its fee or commission. The analysis is performed for the specified service, so a company can be principal for one component and agent for another. (storage.fasb.org)

Relevant evidence includes:

• What is promised: model-IP access, a managed inference service, software access, or an integrated bundle?
• Who is primarily responsible for fulfillment, acceptance and service performance?
• Who bears relevant capacity or inventory risk?
• Who has meaningful pricing discretion?
• What do the provider–cloud, cloud–customer and provider–customer agreements collectively establish?

The IFRS software-reseller interpretation likewise requires a facts-and-circumstances assessment. Pricing discretion and customer contact are indicators, not a substitute for determining control. (ifrs.org)

It is not a discretionary accounting-policy election. Nor is net presentation inherently the standard required of public companies. The contractual facts must support the conclusion. (dart.deloitte.com)

Evidence relevant to Anthropic

Reuters-linked prospectus reporting describes Anthropic as responsible for pricing and service provision, supporting its position that it is the principal in the relevant marketplace transactions. That is an explanation of its accounting judgment—not an independently reproduced audit conclusion. (itiger.com)

A useful first-party corroborating detail appears in AWS’s tax documentation: AWS describes Anthropic’s Bedrock products as third-party sales and, for specified Japanese transactions, describes itself as a marketplace facilitator issuing invoices on Anthropic’s behalf. This supports the distinction between billing intermediary and underlying seller, although tax status does not itself determine the ASC 606 conclusion. (aws.amazon.com)

Evidence relevant to OpenAI

In a memo sent on April 12 and published by The Verge on April 13, 2026, OpenAI revenue chief Denise Dresser said OpenAI reported Microsoft revenue sharing net. She also alleged that Anthropic’s treatment overstated its then-$30bn run rate by approximately $8bn. This is an attributable management claim, not a published accounting reconciliation. (theverge.com)

Crucially, that description does not establish whether OpenAI’s net treatment arises from agency status, a limited contractual royalty entitlement, consideration payable to a customer, or some combination across arrangements. Payments to a counterparty that is also a customer require a separate accounting analysis: generally, consideration payable reduces revenue unless it purchases a distinct good or service, subject to fair-value limits. (ey.com)

The Microsoft agreement changed during 2026

The April 27, 2026 announcement states that:

• Microsoft would no longer pay revenue sharing to OpenAI.
• OpenAI’s payments to Microsoft would continue through 2030, at the same percentage but subject to a total cap.
• OpenAI could serve its products across cloud providers.

The announcement does not disclose the numerical percentage or cap. Therefore, claims that the revenue-sharing obligation simply ended—or that every current revenue dollar can be grossed up with one uniform percentage—are not supported by those published terms. (blogs.microsoft.com)

4. Revenue reconstruction: what the evidence permits

A. Anthropic’s 2025 marketplace economics

The prospectus reporting provides approximately $2.16bn of cloud-marketplace sales and $351m of distribution fees, against nearly $4.6bn of total 2025 revenue. (itiger.com)

Revoqra-derived reconstruction: Use rounded total revenue of $4.60bn and assume the reported fees relate to the same annual sales cohort. The sample is N=1 company-year, not a customer-level sample.

Underlying inputs are source-reported; the ratios and fee-net values are Revoqra-derived. They are not source-published GAAP revenue or cash-collection figures. (itiger.com)

This is a distribution-only normalization. It does not deduct inference, training, other revenue-sharing payments or overhead. Subtracting the entire $2.16bn marketplace channel would incorrectly discard the value retained from those sales.

Nor should the combined 16.3% historical channel deduction be interpreted as a universal AWS contractual rate. AWS’s published standard fees vary by product and offer structure—for example, public SaaS and server/machine-learning listings have different schedules. Those published schedules do not establish Anthropic’s negotiated economics across AWS and Google. (docs.aws.amazon.com)

B. Recent run-rate scenarios—not a verified current leaderboard

The conditional Anthropic calculation is Revoqra-derived: $65bn × [1 − ($0.351bn / $4.60bn)] ≈ $60.0bn. Its fee assumption is based on one historical annual observation, not disclosed 2026 channel economics. (thestar.com.my)

For sensitivity, a 6%–10% deduction from the rounded $65bn anchor produces $61.1bn–$58.5bn. A September Axios source estimated a 6%–10% adjustment. (axios.com)

These scenarios show that accounting presentation can narrow or change an apparent comparison. They do not establish which provider currently has more comparable recognized revenue, retained revenue or gross profit.

C. Does the evidence substantiate OpenAI’s $8bn allegation?

Dresser’s April allegation implies a 26.7% deduction from Anthropic’s then-$30bn headline, leaving $22bn. That arithmetic is Revoqra-derived from the memo’s claim. (news.halloffame.workers.dev)

The documented 2025 distribution-fee fraction is much smaller. Nevertheless, the two cannot simply be declared contradictory measurements: they concern different periods, potentially different commercial payments and possibly different service perimeters. The memo provides no complete bridge identifying the contracts, partner-sales amounts, deductions and serving costs behind the $8bn estimate.

Assessment: The allegation is not sufficiently documented to serve as a verified adjustment. Conversely, the historical 7.6% distribution fraction is not sufficient to prove all 2026 partner economics remained unchanged.

5. Gross margin: revenue normalization is only half the problem

A simple example exposes the classification effect

Assume a hypothetical delivered service generates $100 of customer revenue, requires $50 of serving costs and incurs a $20 distribution charge. These are illustrative assumptions, not estimates for either provider.

All three have the same underlying operating contribution. But their revenue, gross profit and gross-margin percentages differ. This is an arithmetic illustration, not permission to elect any accounting presentation.

Practical conclusion: Compare contribution after both serving and distribution, then separately assess training/R&D and remaining operating expenditure.

OpenAI: reconstructable historical gross margins

Zitron’s reporting supplies revenue and cost-of-revenue lines for 2024 and 2025. The following calculations use those transcribed figures, not a directly obtained provider dataset. N=2 annual observations, yielding one year-on-year comparison. (wheresyoured.at)

Method: gross profit = revenue − cost of revenue; gross margin = gross profit / revenue. Rounded source inputs limit precision. On those inputs, recognized revenue increased approximately 253% and gross margin improved approximately 14.2 percentage points. (wheresyoured.at)

Cost of revenue should not be relabeled pure inference expenditure. Moreover, the reported 2025 gross profit was before $19.18bn of R&D and other operating expenses; the reported operating loss was $20.92bn. Positive gross margin therefore did not establish profitability. (wheresyoured.at)

Anthropic: a conditional margin reconstruction

Reuters Breakingviews described an implied 2025 gross margin of roughly 40%. This is a source interpretation, not a gross-margin figure independently verified from a complete financial statement in this review. (breakingviews.com)

Using 40% solely as a conditional baseline, together with the documented fee classification, gives:

These are Revoqra-derived scenario values, not Anthropic disclosures. The 40% starting margin is approximate, while the fee input and its operating-expense classification are reported from the prospectus. (breakingviews.com)

The reconstruction demonstrates how a roughly 40% headline margin could correspond to a materially lower post-distribution contribution margin. It does not establish a definitive profitability ranking against OpenAI: other partner payments, cost classifications and product mixes remain unmatched.

Do not mistake total compute spending for cost of revenue

Reuters reported Anthropic’s 2025 compute and infrastructure spending at $7.33bn, within $12.65bn of total operating expenses. That aggregate is not a disclosed inference-only cost line. Subtracting it wholesale from revenue would not reconstruct gross margin because training and serving must be separated. (marketscreener.com)

Likewise, net loss is not cash burn. Anthropic’s reported approximately $42bn 2025 net loss included roughly $34bn of financing-related valuation charges, whereas operating losses exceeded $8bn. (marketscreener.com)

Reports of Anthropic’s Q2 operating profit excluded stock-based compensation; that does not demonstrate GAAP net profitability. (axios.com)

6. Inference costs and infrastructure commitments

The economic cost perimeter

A defensible comparison needs three successive layers:

1. Retained commercial proceeds: deduct specified marketplace fees and revenue sharing consistently.
2. Post-serving contribution: deduct inference infrastructure, associated delivery/support costs and relevant capacity costs on a matched basis.
3. Company profitability and cash generation: account for training/R&D, selling expenses, administration, stock compensation, financing and cash investment requirements.

A reseller may bear serving costs that the model developer does not. Grossing up the developer to the reseller’s customer-sales level without adding corresponding costs would overstate the economics attributed to the developer.

Selected commitments illustrate scale, not annual expense

These figures must not be added together mechanically: later aggregates may include individual announced deals.

OpenAI’s CFO describes commitments as tranched and contracts as offering flexibility across providers and hardware. Amazon claims its custom silicon improves cost economics. These are relevant management explanations, not independent demonstrations of future provider margins. (openai.com)

Three accounting and economic cautions

• Commitment value is not recognized revenue or expense. Delivery, availability obligations, prepayments and contractual minimums determine when amounts enter accounts and cash flows.
• Variable pass-through costs can expand reported revenue without equivalent economic value added. One OpenAI compute supplier’s SEC revenue note explicitly recognizes delivered data-center pass-through costs gross and excludes some uncertain future costs from remaining performance obligations. (sec.gov)
• Cloud investment and cloud revenue can be linked without being fictitious. The FTC documented arrangements requiring developers to spend substantial portions of cloud-partner investments on the partner’s services. This warrants scrutiny of external demand and financing dependence, but does not itself prove improper revenue recognition. (search.ftc.gov)

7. Growth, profitability and valuation: material effects, different conclusions

Growth

Observed evidence: Realized annual revenue expanded substantially: OpenAI’s reported annual figures rose from $3.7bn to $13.07bn, while Reuters reported approximately twelvefold Anthropic revenue growth in 2025. It would be unsupported to explain all industry growth as accounting presentation. (wheresyoured.at)

Revoqra synthesis: Nevertheless, a changing channel mix can alter reported growth relative to retained proceeds. A valid growth comparison needs both companies’ prior and current periods restated to the same service perimeter and deduction policy. The available 2025 Anthropic fee data cannot fully restate its 2026 growth, and the OpenAI gross-up does not provide a complete historical bridge.

Profitability

A pure gross-versus-net presentation change, with the corresponding fee expense adjusted consistently, leaves operating income and cash economics unchanged. But a real distribution fee is an economic deduction, and placing it below gross profit changes the apparent gross-margin comparison.

Accordingly:

• Higher revenue does not necessarily imply higher retained revenue.
• Higher reported gross margin does not necessarily imply higher contribution after distribution.
• Positive contribution does not establish profitability after frontier-model training.
• Adjusted operating profit does not establish GAAP net profit or positive free cash flow.

Valuation sensitivity

The following holds historical disclosed post-money equity values constant to isolate denominator effects. These are not contemporaneous enterprise-value multiples, and the funding dates differ from the run-rate dates.

Equity-value anchors are company disclosures; reported run rates and the historical fee assumption supply the denominators. All ratios are Revoqra-derived. (openai.com)

For OpenAI, changing a denominator from $70bn to $50bn increases the apparent sales multiple by 40%, holding valuation fixed. A historical 7.6% revenue deduction raises a sales multiple by approximately 8.3%. These are arithmetic presentation sensitivities—not changes in intrinsic value.

September 29 reports of approximately $1.4tn OpenAI fundraising talks were not a completed financing in the cited report. (techcrunch.com)

Valuation implication: Revenue multiples should be reconciled with retained commercial revenue, post-serving contribution, training requirements and contractual cash obligations. A broader gross-sales denominator cannot, by itself, justify a lower economic valuation multiple.

Across the AI value chain

Microsoft disclosed $24.1bn of fiscal-2026 revenue from commercial arrangements with OpenAI, including revenue-sharing payments. That is Microsoft revenue from its counterparty—not OpenAI’s customer revenue—and its fiscal year ended June 30, unlike the calendar-2025 provider results above. It should not be netted against those results or used as a same-period demand comparison. (sec.gov)

Amazon’s shareholder letter separately reported an AWS AI revenue run rate above $15bn in Q1 2026. This measures a cloud-platform business, not an independent model developer’s revenue. (sec.gov)

Revoqra synthesis: Adding revenue across model developers, cloud infrastructure and application resellers measures activity at multiple stages of the same value chain. It does not measure unduplicated end-customer spending. Comparisons involving integrated providers must specify the business perimeter rather than equating cloud AI revenue with standalone model-provider revenue.

8. What would settle the comparison?

The most decision-useful disclosures would be:

1. A billings-to-revenue bridge: direct subscriptions, direct API, marketplaces, managed-cloud resellers and royalties, with deferred revenue and credits separated.
2. A revenue-to-retained-proceeds bridge: partner deductions by arrangement, distinguishing distribution compensation from inference purchases and broader partnership sharing.
3. Principal-versus-agent policy by channel: specified services, control judgments and contractual fulfillment responsibilities.
4. A consistent gross-profit bridge: serving costs, free-user costs, capacity utilization, depreciation and distribution fees, with training separated.
5. Actual-period financial statements: matched quarterly and annual results, rather than only annualized milestones or adjusted profit claims.
6. A capacity-obligation schedule: minimum purchases, cancellability, delivery milestones, prepayments, financing and overlapping announcements.

Final assessment

Bottom line: Revenue-recognition and presentation differences materially affect perceived scale, growth rankings, gross-margin optics and valuation multiples. They do not eliminate the evidence of rapid commercial expansion, nor do they resolve whether that expansion will generate durable profits. The strongest comparison is matched-period retained revenue plus consistently measured serving contribution, followed by training and contractual cash requirements—not ARR alone.

Sources

Primary statements, filings and accounting evidence

• OpenAI: January 18, 2026 CFO discussion of historical run rates; March 31, 2026 revenue and financing announcement. (openai.com)
• Anthropic: April 20, 2026 AWS agreement; May 28 funding/run-rate disclosure; June 1 confidential S-1 announcement. (anthropic.com)
• Microsoft: October 2025 agreement filed with the SEC; February and April 2026 partnership updates; fiscal-2026 related-party revenue disclosure. (sec.gov)
• FASB: ASU 2016-08, principal-versus-agent considerations. (storage.fasb.org)
• IFRS Foundation: Software-reseller principal-versus-agent interpretation and IFRS 15 framework. (ifrs.org)
• FTC: Cloud-provider/AI-developer partnership study; findings principally reflect information through September 2024 and public material through January 2025. (search.ftc.gov)
• AWS: Marketplace fee schedules and first-party descriptions of third-party Bedrock transactions. (docs.aws.amazon.com)
• SEC supplier disclosure: Gross presentation of OpenAI-related infrastructure pass-through costs. (sec.gov)
• Amazon: Shareholder-letter discussion of AWS AI revenue. (sec.gov)

Original reporting and specialist interpretation

• Reuters: OpenAI September run-rate clarification; Anthropic July investor update; Anthropic prospectus finances and commitments. (investing.com)
• Reuters, carried by The Star: Anthropic distribution fees and expense classification. (thestar.com.my)
• Reuters Breakingviews: Approximate implied Anthropic gross margin. (breakingviews.com)
• The Verge: Published Dresser memo, providing the original text of OpenAI’s allegation. (theverge.com)
• Ed Zitron: Original reporting and transcription of OpenAI historical financial statements; the article identifies Financial Times verification. (wheresyoured.at)
• Deloitte and EY: Original technical accounting interpretation. (dart.deloitte.com)
• TradingKey/Tiger Brokers: Additional Reuters-linked prospectus details. (itiger.com)
• Axios: Fee-adjustment estimate and adjusted-profit qualification. (axios.com)
• Bloomberg via TechCrunch: Fundraising-talk qualification. (techcrunch.com)