Introduction: The October 2026 Revenue Confusion That Shook the Artificial Intelligence Economy
On Thursday, October 8, 2026, Wall Street confronted a question that would have seemed absurd in any previous technology cycle, and yet one that goes to the heart of how the artificial intelligence economy is currently being measured, financed, and valued: had the revenue of OpenAI, one of the two most closely watched private companies in the world, suddenly fallen by roughly $20 billion, or had investors simply been comparing two different financial measurements of the same underlying business? The Financial Times reported that morning, on the basis of investor documents it had reviewed, that OpenAI had told its backers its annualized revenue was approaching $50 billion at the end of September, a figure that stood roughly $20 billion below the approximately $70 billion number that had circulated widely among investors and in the financial press only two weeks earlier.[1] CNBC independently confirmed the $50 billion figure that afternoon, and reported that a person familiar with the matter had described the earlier $68 billion number as one that included gross revenue from OpenAI’s partners, a presentation that had been constructed to allow a more direct comparison with its principal rival, Anthropic.[2]
The market reaction was immediate and, in its own way, deeply revealing. Shares of Nvidia, Oracle, and CoreWeave, three companies whose valuations rest to a meaningful degree on the expectation that frontier-model developers will continue to purchase accelerators, cloud capacity, and datacenter space at an accelerating pace, fell on the day the $50 billion figure became public.[3] Nothing about OpenAI’s underlying business had changed between the last week of September and the second week of October. The company’s product portfolio, its customer base, its enterprise contracts, its consumer subscriptions, and its multi-year compute commitments were all exactly what they had been. What changed was the denominator: the definition of what the word “revenue” was being asked to describe, and the method by which a monthly figure had been converted into an annualized one. That a measurement convention, rather than a change in economic substance, could move hundreds of billions of dollars of public-market capitalization in a single session is the phenomenon this paper sets out to examine.
By Friday, October 9, Bloomberg’s additional reporting offered a more layered picture. OpenAI, it emerged, was still telling investors that it expected annualized revenue to reach or exceed $70 billion by the end of 2026, with its enterprise segment serving as the primary engine of that growth, and the September figure of roughly $50 billion was consistent with the company’s own internal presentation rather than a sign of faltering demand.[3] The earlier $70 billion figure, Bloomberg explained, had originated from investors attempting to calculate OpenAI’s performance using the same methodology as Anthropic, which tallies the full gross value of sales made through cloud partners such as Amazon Web Services and Google Cloud, whereas OpenAI books only its own share of proceeds from certain partner arrangements.[3] Reuters, reporting the following day, added the detail that would become central to the subsequent debate: OpenAI, unlike Anthropic, does not include revenue from sales made through cloud partners, and Anthropic pays those partners roughly 16 cents of every dollar earned through them, a channel that accounted for close to half of Anthropic’s revenue in the prior year.[4]
The September measurement discrepancy and OpenAI’s year-end forecast were, strictly speaking, two separate matters. One concerned how a single month’s revenue had been annualized and whether partner sales were included in the base; the other concerned a forward-looking projection for the final quarter of the year. But their appearance in the same forty-eight-hour news cycle, layered on top of a months-long and increasingly public dispute between the two leading frontier laboratories over whose revenue was “real,” amplified a confusion that had been building since the spring. In April 2026, OpenAI’s chief revenue officer had circulated an internal memo asserting that Anthropic’s then-$30 billion run rate was overstated by roughly $8 billion through gross revenue accounting on cloud-partner sales.[9] In June, the two companies filed confidential registration statements with the Securities and Exchange Commission seven days apart, making side-by-side financial comparison inevitable for every institutional investor evaluating either listing.[47] In early September, Axios published an explanatory piece in which an accounting professor laid out, in plain language, why two companies with the same business model could legitimately arrive at different revenue conclusions.[5] And at the end of September, Reuters obtained a copy of Anthropic’s confidential prospectus and reported that 47 percent of the company’s 2025 sales had been routed through Amazon and Google, that it had paid roughly $351 million back to those platforms in distribution fees, and that it had recognized the gross amount because it regarded itself as the principal in the transaction.[6]
The controversy therefore exposed a structural problem that extends far beyond the reporting practices of two private companies, however large and however consequential. Across the emerging artificial intelligence economy, revenue originates through consumer subscriptions, enterprise software agreements, application programming interfaces priced per token, cloud marketplaces, model distribution partnerships, autonomous agents that consume compute on behalf of users, and increasingly sophisticated commercial arrangements between model developers and the infrastructure providers that are simultaneously their suppliers, their distributors, their investors, and their customers. Some of those transactions are recorded gross and some are recorded net, depending on the contractual responsibilities of each party and the judgment each company’s management and auditors apply under the applicable accounting standards. Annualized revenue run rates, recognized revenue under generally accepted accounting principles, bookings, deferred revenue, remaining performance obligations, and contracted future payments can each describe a different dimension of the same rapidly expanding industry, and in the autumn of 2026 those dimensions were being quoted interchangeably in headlines, investor decks, and analyst notes.
This raises a question of real consequence for what I have described in earlier work as the Five-Layer AI Economy, the vertically integrated system of Energy, Chips, Datacenters, Models, and Applications and Agents through which artificial intelligence is produced and delivered. If investors cannot consistently determine how much revenue the companies at the top of that stack are actually generating, how should they assess the trillions of dollars of infrastructure investment being planned beneath them to support their future growth? Nvidia reported $89.0 billion of datacenter revenue in a single quarter ending in July 2026, and its chief executive declared that compute had become revenue.[16] Oracle disclosed $664 billion of remaining performance obligations in September, more than thirty times its quarterly revenue.[20] The four largest hyperscalers guided to combined 2026 capital expenditure in the range of $725 billion to $760 billion.[24][25] Every one of those numbers rests, directly or indirectly, on an expectation about how much money the Models and Applications layers will eventually collect from customers, and in October 2026 the market discovered that it did not have a settled definition of the figure it had been using to anchor those expectations.
The October 2026 controversy suggests that understanding the denominator underlying AI revenue comparisons may become as important as tracking the headline numbers themselves. It provides the starting point for this paper’s examination of how financial measurement shapes company valuations, cloud relationships, semiconductor demand expectations, infrastructure financing, and ultimately the economic sustainability of artificial intelligence through 2030. The argument proceeds in six sections. Section 1 reconstructs the October episode and situates it within the accounting framework that governs gross-versus-net presentation. Section 2 traces the AI distribution chain from end customer to model developer and develops the Revenue Denominator framework. Section 3 examines how inconsistent revenue definitions distort frontier-model valuation. Section 4 follows the transmission mechanism from revenue expectations down through datacenters, chips, and energy. Section 5 considers public-market disclosure, regulatory oversight, and a proposed reconciliation framework for the road to 2030. Section 6 distills the lessons into eight pillars, and the conclusion returns to the question of why a measurement problem may become one of the defining financial frameworks of the decade.
Why I Chose the Title “Revenue Denominator”
I chose the title “Revenue Denominator” because the October 8–9, 2026, reporting on OpenAI and Anthropic revealed a fundamental weakness in how investors interpret the commercial performance of rapidly expanding artificial intelligence companies, and because that weakness is best understood through the arithmetic metaphor the title invokes. When an analyst computes a revenue multiple, a growth rate, a market-share estimate, or a return on invested capital, revenue sits in the denominator of the first calculation and in the numerator of the others, and in every case the result is only as meaningful as the definition of the revenue figure that was fed into it. Revenue figures that appear directly comparable may in fact reflect different accounting treatments, different reporting periods, different distribution arrangements, and different allocations of economic responsibility among the parties to a transaction. A valuation of $852 billion placed on one company and a valuation of $965 billion placed on another cannot be meaningfully compared on a revenue-multiple basis if the two revenue figures are not measuring the same thing, and in the autumn of 2026 they were not.[46][47]
The term emphasizes that before comparing revenue growth, valuation multiples, profitability, or market leadership, analysts must first establish a consistent measurement basis, exactly as a student must find a common denominator before adding two fractions. It introduces a framework for understanding the relationship between gross and net revenue, between annualized and recognized sales, between customer contracts and cash collected, between cloud marketplace transactions and the economic value actually retained by the frontier-model developer at the end of the chain. None of these distinctions is new to accountants. All of them became newly urgent in 2026 because the magnitudes involved had grown so large, the growth rates so steep, and the infrastructure commitments resting on them so enormous, that a difference of definition could no longer be treated as a footnote.
The title also connects financial transparency to the broader Five-Layer AI Economy of Energy, Chips, Datacenters, Models, and Applications and Agents. Although commercial AI revenue is generated primarily through the two upper layers, expectations about future revenue in those layers drive investment decisions throughout the physical infrastructure beneath them. A semiconductor company sizing its production commitments, a hyperscaler deciding whether to raise its capital expenditure guidance for a third consecutive quarter, a private-credit fund underwriting a datacenter campus, and a utility planning a transmission upgrade are all, whether they articulate it this way or not, making bets on a revenue figure that originates at the top of the stack. If that figure is defined inconsistently, the error propagates downward and compounds at each layer. By examining how revenue measurements affect capital allocation, hyperscaler partnerships, semiconductor procurement, datacenter financing, and public-market valuations, this paper seeks to establish a forward-looking framework for investors, corporate executives, accountants, regulators, and policymakers preparing for the AI economy of 2027 through 2030, when the first generation of frontier laboratories will be reporting as public companies and the first generation of AI-era infrastructure debt will be coming due.
Section 1: The October 2026 Revenue Confusion — When Financial Measurements Become Market-Moving Information
The deepest lesson of the October 2026 episode is not that one company’s revenue was lower than reported, nor that another company’s revenue was higher than it “should” have been, but that the financial vocabulary the market had been using to describe the artificial intelligence boom had quietly become unfit for purpose. For three years, the frontier-model industry had communicated its commercial progress almost exclusively through a single metric, the annualized revenue run rate, and for three years that metric had been quoted in headlines, repeated in funding announcements, and extrapolated in analyst models with very little attention paid to how it was constructed. The run rate had the virtue of being simple, dramatic, and always rising. It had the vice of being undefined. When two companies with different distribution strategies and different accounting judgments began to be compared against one another on that metric at valuations approaching a trillion dollars, the absence of a common definition stopped being an academic concern and became a market-moving one. This section reconstructs what happened, explains why the two companies’ presentations diverge, situates that divergence within the accounting standards that govern it, and traces how a dispute about measurement translated into volatility in the shares of companies several layers removed from the original transaction.
1.1 October 8–9, 2026: Reconstructing the OpenAI Revenue Discrepancy
The sequence of figures that confused the market can be reconstructed from the reporting of the Financial Times, CNBC, Bloomberg, Reuters, and Axios between October 8 and October 10. According to the Financial Times, OpenAI had informed investors that its annualized revenue was approaching $50 billion at the end of September 2026, and the company had told them that revenue had grown by more than 70 percent over the period in question.[1] CNBC added that OpenAI had touted 77 percent total run-rate growth during its third quarter and 107 percent run-rate growth for its enterprise business over the same period.[2] Neither of these growth figures was in dispute. What was in dispute was the base to which they applied.
The $70 billion figure that had circulated in late September had a traceable genealogy. Earlier in the year, investors seeking to compare OpenAI with Anthropic on a like-for-like basis had attempted to “gross up” OpenAI’s reported figures to include the full value of sales flowing through partners, and that adjustment had produced reports that OpenAI’s annualized revenue reached $40 billion in August.[1] When OpenAI subsequently told investors that revenue had grown by more than 70 percent, applying that growth rate to the grossed-up $40 billion base produced the roughly $70 billion number that appeared in multiple outlets.[1] The Financial Times reported that OpenAI’s July annualized revenue on its own narrower basis had been closer to $30 billion, and Data Studios’ reconstruction of the investor materials placed the narrower measure at approximately $30 billion in July and $50 billion in September.[10][11] In other words, the $70 billion was never a figure OpenAI itself had reported as its September run rate; it was a derived estimate built on a comparability adjustment that had then been compounded by a growth rate that applied to a different base.
Bloomberg’s October 9 reporting clarified the forward-looking dimension. OpenAI continued to tell investors that it expected annualized revenue to reach or exceed $70 billion by the end of 2026, driven primarily by enterprise adoption.[3] The $70 billion was therefore simultaneously a misattributed September estimate and a legitimate December target, and the collision of those two uses of the same number in the same week was a large part of what unsettled the market. KuCoin’s summary of the episode noted a third figure that further illustrates the vocabulary problem: OpenAI’s projected actual revenue for calendar 2026 had reportedly been revised upward to approximately $35 billion, a figure that is entirely consistent with a September run rate near $50 billion and a December run rate near $70 billion, because a company growing at 70 percent or more per year will always report full-year recognized revenue far below its exit run rate.[55] Three numbers, $35 billion, $50 billion, and $70 billion, were all simultaneously “OpenAI’s 2026 revenue” depending on whether one meant recognized revenue for the fiscal year, the annualized rate at the end of the third quarter, or the projected annualized rate at year-end.
The following table reconstructs the key figures and the basis on which each was stated.
Table 1
| Figure | What it measured | Basis of presentation | Source |
| ~$30 billion | OpenAI annualized revenue, July 2026 | OpenAI’s own narrower (net-of-partner) presentation | FT via Investing.com; Data Studios[1][11] |
| ~$40 billion | OpenAI annualized revenue, August 2026 | Investor “gross-up” adjustment for comparability with Anthropic | FT via Investing.com[1] |
| ~$68–70 billion | Reported OpenAI run rate, late September 2026 | Grossed-up August base compounded by >70% growth; not an OpenAI-reported figure | CNBC; Bloomberg via Quartz[2][3] |
| ~$50 billion | OpenAI annualized revenue, end of September 2026 | OpenAI’s own presentation to investors | FT; CNBC; Reuters[1][2][4] |
| ~$35 billion | OpenAI projected recognized revenue, calendar 2026 | Full-year revenue, not a run rate | KuCoin summary of reporting[55] |
| ≥$70 billion | OpenAI projected annualized revenue, December 2026 | Forward target, enterprise-led | Bloomberg via Quartz[3] |
| ~$65 billion | Anthropic annualized revenue, end of July 2026 | Includes gross value of sales through cloud partners | Bloomberg via Quartz; Reuters[3][4] |
The lesson of the table is that every figure in it was, on its own terms, accurate. The confusion arose entirely from the juxtaposition of figures computed on different bases, over different periods, and for different purposes, without the labels that would have allowed a reader to distinguish them.
1.2 OpenAI Versus Anthropic: Two Different Revenue Presentations
The divergence between the two companies’ presentations arises from a single structural fact about how frontier models reach enterprise customers in 2026. A very large share of enterprise consumption of Claude, and a smaller but growing share of consumption of OpenAI’s models, is purchased not directly from the model developer but through a cloud marketplace operated by Amazon Web Services, Google Cloud, or Microsoft Azure, where the enterprise customer already holds an account, a negotiated enterprise agreement, a committed-spend contract, and a procurement relationship. The customer is billed by the cloud provider. The cloud provider collects the cash. The cloud provider then remits a portion to the model developer and retains a portion as a distribution or marketplace fee. The question of whose revenue the gross amount represents is precisely the principal-versus-agent question that revenue-recognition standards were written to answer, and the two companies have answered it differently.
Anthropic records the full value of Claude sales made through its cloud partners as revenue, and then records the cloud providers’ share as an expense. OpenAI records only its own share of certain sales made through partners such as Microsoft.[5] The Reuters analysis of Anthropic’s prospectus quantified the arrangement with unusual precision: sales through the cloud marketplaces totaled about $2.16 billion, or 47 percent of Anthropic’s 2025 revenue of nearly $4.6 billion, and the company paid roughly $351 million back to the platforms in distribution fees, implying that the cloud providers collected approximately 16 cents of every dollar of those sales.[6] The share of Anthropic’s revenue routed through Amazon and Google had risen from 11 percent in 2023 to 32 percent in 2024 and 47 percent in 2025, and the same third parties were responsible for collecting 60 percent of the $909 million in outstanding customer bills at the end of 2025.[7] The channel is therefore not a peripheral feature of Anthropic’s business model but its primary route to the enterprise.
Anthropic’s justification for gross presentation has been stated consistently. When Reuters reported in June that OpenAI had told investors and employees that this approach inflated Anthropic’s reported revenue by billions of dollars, Anthropic responded that it follows established accounting practices and recognizes gross revenue because it is the principal in the transaction.[6] A person close to the company offered the Financial Times the same formulation in April, adding that the cloud partners function as the distribution channel.[9] OpenAI’s position, as expressed in the April memo from its chief revenue officer, was that net presentation is the more accurate reflection of economic reality.
“Their stated run rate is inflated.”
— Denise Dresser, Chief Revenue Officer, OpenAI, internal memo via The Verge and CNBC[9]
It is essential to recognize, as the Axios analysis did, that neither approach is necessarily wrong. The accounting professor Francine McKenna explained to Axios that the criteria turn on who is the principal and who is the agent, which in turn depends on who controls the client relationship and who is ultimately responsible for delivering the product.
“You can have two companies with the same business model come to a different conclusion.”
— Francine McKenna, accounting professor and author of The Dig, to Axios[5]
McKenna’s point is that the conclusion depends both on the specific terms of the customer contracts and on the argument each company’s accounting team makes to the Securities and Exchange Commission in justifying its treatment. Lucas Ward, chief executive of the financial verification platform Oath, told Axios that accounting in this area is closer to an art than a science, and said he planned to look past the revenue line to EBITDA for a clearer view of the companies’ true profitability.[5] The same analysis noted a crucial quantitative limit on the accounting explanation: if Anthropic were to switch from gross to net presentation, a source familiar with its financials estimated the hit at only 6 to 10 percent, which would still leave the company roughly $19 billion to $21 billion ahead of OpenAI on the figures available in early September.[5] The accounting difference, in other words, is real and material, but it does not explain the whole gap, and Anthropic may genuinely be growing faster than its rival.
The arithmetic of a single $100 transaction routed through a cloud marketplace, using the 16 percent distribution fee implied by the Reuters analysis, makes the mechanics concrete.
Table 2
| Line item | Gross presentation (principal) | Net presentation (agent) |
| Customer pays cloud provider | $100 | $100 |
| Cloud provider retains distribution fee | $16 | $16 |
| Amount remitted to model developer | $84 | $84 |
| Revenue recognized by model developer | $100 | $84 |
| Distribution fee recorded by model developer | $16 expense (cost of revenue or sales and marketing) | Not recorded (netted against revenue) |
| Gross profit before compute costs | $84 | $84 |
| Cash collected by model developer | $84 | $84 |
| Implied headline run-rate difference | +19% relative to net | — |
The final two rows are the ones that matter most for the Revenue Denominator argument. Under either presentation, the model developer ends up with the same $84 of gross profit before compute costs and the same $84 of cash. The economic substance is identical. What differs is the top line, and because the industry has chosen to communicate its progress primarily through the top line, a difference that leaves profit and cash unchanged nevertheless changes the headline by nearly a fifth. An analyst who applies a revenue multiple to the gross figure and another who applies the same multiple to the net figure will arrive at valuations that differ by 19 percent for two businesses with identical economics.
1.3 Annualized Revenue Run Rates Versus Recognized Revenue
The second source of confusion in October was not the gross-versus-net question at all, but the practice of annualization itself. An annualized revenue run rate is typically computed by multiplying the most recent month’s revenue by twelve, or occasionally by annualizing a quarter. It is an operating metric rather than an accounting measure, it is not defined by any standard-setter, and it is not revenue in the sense that appears on an audited income statement. The glossary definition used across private-market diligence makes the point crisply: ARR is a snapshot of recurring contracts at a date, whereas revenue is recognized under accounting standards over a period, and diligence normally reconciles the former to recognized revenue, billings, and cash collections.[50]
For a business growing at a steady and modest rate, the distinction between run rate and recognized revenue is small. For a business whose revenue grows by 70 percent or more in a year, the distinction is enormous, because the exit run rate will always be far above the recognized revenue for the year that produced it. OpenAI’s own trajectory illustrates the point: the company started 2026 with roughly $20 billion in annualized revenue, up from roughly $6 billion at the equivalent point in 2024, reached $25 billion in February, approached $50 billion by September, and targeted $70 billion by December, while its recognized revenue for calendar 2025 was approximately $13 billion and its projected recognized revenue for 2026 was approximately $35 billion.[4][45][55] Anthropic’s figures are even more dramatic: recognized 2025 revenue of nearly $4.6 billion, a run rate of $9 billion at the end of 2025, $30 billion in March, $65 billion at the end of July, and a target of $100 billion by year-end.[4][6][9] Reuters reported that Anthropic’s quarterly revenue eclipsed OpenAI’s for the first time in the second quarter of 2026, at $11.5 billion against $6.7 billion.[4]
Three distinct measurement problems lurk inside these figures. The first is the base-month problem: a run rate annualizes one month, so any lumpiness in that month, whether from a large enterprise deal, a seasonal consumption spike, or a one-time credit purchase, is multiplied by twelve. The second is the recognition-timing problem: usage-based revenue is recognized as tokens are consumed, but prepaid credits, annual subscriptions, and committed-spend contracts create deferred revenue that is recognized over the service period, so cash received, revenue recognized, and run rate can all diverge in the same month. The third is the forecast-contamination problem, which is what actually happened in October: when a company publicly guides to a year-end run rate, that forward figure can be mistaken for a current one, and when investors apply a stated growth rate to a base they have themselves adjusted, the result is a number that no one has reported and no auditor has reviewed.
The distinction among the terms is set out in the following table, which will be referred to throughout the paper.
Table 3
| Term | What it measures | Who defines it | Appears on audited financial statements? | Primary risk when misused |
| Annualized revenue run rate (ARR) | Most recent month or quarter × 12 or × 4 | The company, not a standard-setter | No | Overstates recognized revenue for fast-growing firms; sensitive to one month |
| Recognized revenue | Revenue earned over the fiscal period under ASC 606 / IFRS 15 | FASB / IASB | Yes | Lags commercial momentum; depends on gross-versus-net judgment |
| Bookings | Value of contracts signed in a period | The company | No | Signed is not delivered; includes optional and conditional commitments |
| Deferred revenue | Cash received for services not yet delivered | FASB / IASB | Yes (balance sheet) | Can inflate cash flow while revenue lags |
| Remaining performance obligations (RPO) / backlog | Contracted revenue not yet recognized | FASB / IASB (RPO); company (backlog) | RPO disclosed in notes; backlog not | Conditional on capacity delivery; multi-year; concentration |
| Cash collected | Actual cash received from customers | — | Yes (cash flow statement) | Timing differs from revenue; third-party collection risk |
A reader who keeps this table in mind will find that most of the confusion in the autumn of 2026 reduces to the quotation of one row as though it were another.
1.4 Gross-versus-Net Accounting Under Existing Standards
The accounting framework that governs the Anthropic–OpenAI divergence is neither new nor obscure. The converged revenue standard, Accounting Standards Codification Topic 606 in the United States and IFRS 15 internationally, was issued in 2014 and became effective for public companies in 2018. In March 2016, the Financial Accounting Standards Board issued Accounting Standards Update 2016-08, “Principal Versus Agent Considerations (Reporting Revenue Gross Versus Net),” specifically to clarify how an entity determines whether it is a principal or an agent for each specified good or service promised to a customer, in response to implementation questions raised by the joint FASB–IASB Transition Resource Group.[12] The amendments were identical in substance to the clarifications the IASB made to IFRS 15, so the two frameworks are aligned on this point.[12]
The core principle is control. An entity is a principal if it controls the specified good or service before that good or service is transferred to the customer, and it is an agent if it does not; a principal recognizes revenue on a gross basis, and an agent recognizes revenue on a net basis.[13] This represented a deliberate shift from the “risks and rewards” model of legacy GAAP to a control model, and the standard’s indicators of control, which include primary responsibility for fulfilling the promise, inventory risk before or after transfer, and discretion in establishing prices, are to be used to support a conclusion about control rather than as a freestanding checklist.[13] ASU 2016-08 further clarified that the unit of account is each specified good or service, so an entity can be a principal for some promises in a contract and an agent for others, and must determine the nature of each specified good or service before assessing control.[56]
Applied to a cloud marketplace sale of a frontier model, the analysis asks: what is the specified service the enterprise customer is buying, and who controls it before it is transferred? Anthropic’s position is that the specified service is access to Claude, that Anthropic controls that service, sets its capabilities, bears primary responsibility for its performance, and has substantial discretion over its pricing, and that the cloud provider is arranging for Anthropic’s service to be provided through the provider’s marketplace, in the way that an app store arranges for a developer’s app to be sold. On that reading Anthropic is the principal and gross presentation follows. OpenAI’s position with respect to certain Microsoft arrangements appears to be that the customer is purchasing an Azure service that embeds OpenAI’s models, that Microsoft controls the bundled offering, bills the customer, and bears responsibility for the service level, and that OpenAI’s share is more in the nature of a royalty or revenue share on a service controlled by another party. On that reading OpenAI is closer to an agent, or more precisely a licensor, and net presentation follows.
Both readings are defensible applications of the same standard to arrangements whose contractual terms differ. That is precisely McKenna’s point. The standard does not ask whether the two companies sell similar products; it asks who controls the specified service in each specific contract, and the contracts between Anthropic and Amazon, between Anthropic and Google, between OpenAI and Microsoft, and between OpenAI and Amazon, which became the exclusive third-party cloud distributor for OpenAI’s Frontier agent platform in February 2026, are not identical.[9] The implication for investors is uncomfortable but unavoidable: the accounting standards are functioning exactly as designed, and the lack of comparability is a consequence of genuine differences in commercial structure rather than a defect in the rules. Any remedy must therefore come through disclosure rather than through a change in recognition.
1.5 From Accounting Confusion to Financial-Market Volatility
The final element of the October episode is the one with the broadest implications: the speed with which uncertainty about frontier-model revenue translated into price movements in companies several layers removed from the original transaction. Nvidia, Oracle, and CoreWeave all fell on October 8 when the $50 billion figure became public.[3] None of those companies sells a single token to an enterprise customer through a cloud marketplace. Each of them, however, sells to, finances, or is financed by the companies that do, and each of them carries a valuation that embeds an assumption about how fast the top of the stack is growing.
The mechanism is straightforward. Oracle’s remaining performance obligations of $664 billion at the end of its fiscal first quarter, up $209 billion year over year, are dominated by multi-year AI cloud contracts, and the company’s ability to convert that backlog into revenue depends on its customers’ ability to pay for the capacity they have reserved.[20] CoreWeave’s revenue backlog of approximately $104 billion as of June 30, 2026, against quarterly revenue of $2.575 billion and total debt of roughly $35 billion, represents about forty quarters of current revenue, and its 2025 annual filing disclosed that Microsoft alone generated 67 percent of its revenue, with OpenAI the second pillar.[22][23] Nvidia’s hyperscale datacenter revenue of $48.7 billion in a single quarter, more than doubling year over year, is purchased by companies whose own capital expenditure budgets are justified to their shareholders by reference to AI demand.[17] When the market’s estimate of that demand falls by $20 billion in a single morning, even for reasons of definition rather than substance, every one of those valuations is mechanically repriced.
The Financial Times made the connection explicit in its original reporting, noting that the annualized revenue metric serves as a key indicator of overall demand for AI technology and influences infrastructure spending decisions across the sector.[1] That observation, offered almost in passing, is the thesis of this paper. The run rate had become the single most important number in the AI economy not because it was the best measure of anything in particular but because it was the number everyone used, and a number everyone uses acquires systemic importance regardless of its analytical merits. The October episode demonstrated that the AI infrastructure complex, with its trillions of dollars of planned capital expenditure and its hundreds of billions of dollars of newly issued debt, had been built on a metric that two of its most important participants defined differently and that the market had never required either of them to reconcile. The remainder of this paper is an attempt to work out what follows from that discovery.
Section 2: Inside the Artificial Intelligence Distribution Chain — Who Generates Revenue, Who Recognizes It, and Who Retains the Economic Value?
To understand why two well-advised, well-audited companies can present the same category of transaction so differently, it is necessary to follow a dollar of enterprise AI spending from the moment it leaves a customer’s budget to the moment it settles, in whatever diminished form, on the income statement of the company whose model actually produced the output. That journey is longer and more contested in 2026 than it was in any previous software cycle. In the era of on-premises enterprise software, the vendor sold a license, booked the revenue, and bore the cost of building the product; distribution partners, where they existed, were clearly resellers. In the era of software as a service, the vendor hosted the product on infrastructure it rented from a cloud provider, and the cloud provider was clearly a supplier. In the era of frontier models, the cloud provider is simultaneously the supplier of the compute on which the model runs, the distributor through whose marketplace the enterprise buys access, the collector of the customer’s cash, an equity investor in the model developer, a customer of the model developer, and in several cases a direct competitor offering its own models on the same shelf. The question of who generates revenue, who recognizes it, and who retains the economic value is therefore not a technicality but a description of the industry’s power structure, and this section traces it channel by channel before assembling the five dimensions of the Revenue Denominator framework.
2.1 Direct Consumer Subscriptions and Enterprise Contracts
The simplest and most comparable channel is the direct one, in which a consumer or an enterprise pays the model developer directly for access to ChatGPT, Claude, Gemini, or a competing product. Consumer subscriptions are recognized ratably over the subscription period, enterprise seat licenses are recognized over the contract term, and prepaid usage credits are recognized as consumed, with the unconsumed balance sitting in deferred revenue. There is no principal-versus-agent ambiguity in this channel because there is no intermediary; the model developer controls the service, bills the customer, and bears the delivery risk. Were the entire industry to operate through this channel, the October confusion could not have arisen.
The difficulty is that the direct channel is a shrinking share of the enterprise business for at least one of the two leading laboratories and an uncertain share for the other. Anthropic’s prospectus disclosed that about $3.8 billion of its 2025 revenue came from usage-based billing while subscription revenue amounted to $789 million, and that 47 percent of the total was routed through cloud marketplaces; the direct enterprise and consumer business was therefore slightly more than half of the whole and declining as a proportion.[7] OpenAI’s business is weighted more heavily toward consumer subscriptions and direct enterprise contracts, which is one reason its gross-versus-net adjustment is smaller in relative terms, but its February 2026 arrangement making Amazon the exclusive third-party cloud distributor for its Frontier agent platform suggests that its marketplace share will rise.[9] Even within the direct channel, the vocabulary problem persists: a $10 million annual enterprise contract signed in September produces $10 million of bookings in September, roughly $833,000 of recognized revenue per month thereafter, $10 million of run rate immediately if the company annualizes contract value, and cash collections that depend entirely on the billing schedule negotiated with the customer. A company that reports bookings-based run rate and a company that reports recognition-based run rate will show different trajectories for identical commercial activity.
2.2 API Sales, Token Consumption, and Usage-Based Economics
The second channel is the application programming interface, through which developers and enterprises purchase model output priced per million input and output tokens. This channel is the engine of the frontier-model business and the source of its most distinctive measurement challenges. Usage-based revenue is recognized as the service is consumed, which aligns revenue closely with activity but makes it volatile from month to month and highly sensitive to pricing. A price reduction of 50 percent on a given model, which has occurred repeatedly across the industry as newer models displace older ones, halves the revenue from a fixed volume of tokens even as it may stimulate demand; a run rate computed in the month before a price cut and one computed in the month after describe the same customer base very differently.
Three further distinctions matter in this channel. The first is between commercial activity and revenue earned: tokens generated under free tiers, under promotional credits, under developer sandboxes, or under the enterprise’s own evaluation programs represent compute consumed and cost incurred but not revenue recognized. The second is between revenue earned and cash collected: enterprise API customers are typically invoiced in arrears on net-thirty or net-sixty terms, and when the API is consumed through a cloud marketplace the cash is collected by the cloud provider and remitted later, which is why 60 percent of Anthropic’s outstanding customer bills at the end of 2025 were in the hands of third parties.[7] The third is between reserved and on-demand consumption: OpenAI’s announcement in May 2026 of Guaranteed Capacity, offering customers one-, two-, and three-year commitments for long-term compute access, introduces a committed-spend structure into the API channel that creates bookings and remaining performance obligations in a business that had previously been almost entirely pay-as-you-go.[47] Each of these distinctions widens the gap between what the market calls “revenue” and what an auditor would recognize as such.
Nvidia’s chief executive framed the economic significance of the token in a single sentence in his company’s second-quarter results.
“Now, compute is revenue.”
— Jensen Huang, founder and CEO, NVIDIA, Q2 FY2027 results[16]
The statement is a precise description of the usage-based channel’s economics from the chipmaker’s perspective: every token generated consumes accelerator cycles, and every accelerator cycle was purchased with capital. It is also, read in the light of the October episode, an acknowledgment of the dependency this paper is concerned with. If compute is revenue, then the measurement of revenue at the top of the stack is the measurement of the return on compute at the bottom of it, and an undefined numerator produces an undefined return.
2.3 Cloud Marketplaces and Distribution Partnerships
The third channel is the one at the center of the dispute. Amazon Web Services, Microsoft Azure, Google Cloud, and Oracle Cloud Infrastructure each operate marketplaces through which enterprise customers can purchase third-party software and model access, drawing down against their existing committed cloud spend and consolidating the purchase onto a single invoice. For the enterprise buyer, the attraction is procurement simplicity and the ability to apply pre-committed cloud dollars to AI consumption. For the model developer, the attraction is access to tens of thousands of enterprise accounts that have already cleared security review, legal review, and budget approval with the cloud provider. For the cloud provider, the attraction is a distribution fee, increased consumption of its own infrastructure, and a deeper lock-in of the enterprise relationship.
The accounting consequences flow from the division of responsibilities. The cloud provider bills the customer, collects the cash, and typically bears the credit risk. The model developer provides the service, sets its technical characteristics, and in most arrangements retains substantial discretion over list pricing. Whether the model developer “controls” the specified service before it is transferred to the customer, and therefore whether it is the principal, depends on the specific contractual allocation of responsibility for service levels, pricing, customer support, and fulfillment, which is why Anthropic’s prospectus framed its relationships with Amazon, Google, and Microsoft as an advantage while simultaneously acknowledging that they create complex dynamics that could give rise to conflicts of interest and could adversely affect its access to compute.[7] The same filing noted that the cloud companies are also Anthropic’s customers, which means that a single counterparty can appear on the model developer’s financial statements as a supplier of compute, a distributor collecting a fee, a customer purchasing model access, and an investor holding equity.[6]
The commercial reality is that the cloud marketplace has become the dominant enterprise distribution channel for frontier models faster than the market’s analytical vocabulary has adapted. Anthropic’s channel share rose from 11 percent to 47 percent in two years.[7] Anthropic’s May 2026 announcement that Claude was available across all three major clouds, and OpenAI’s February 2026 Amazon distribution agreement, indicate that both companies regard multi-cloud marketplace distribution as strategically necessary regardless of how it complicates their reported revenue.[9][11] The Reuters analysis quantifying the 16 percent distribution fee is therefore not merely a detail about one company but an early data point on the economics of an entire channel through which a growing share of the AI economy’s revenue will flow.[6]
2.4 The Economic Structure of Revenue-Sharing Agreements
Beyond the marketplace fee lies a denser thicket of revenue-sharing arrangements whose terms are largely confidential and whose effects on reported revenue are correspondingly difficult for outsiders to assess. OpenAI’s arrangement with Microsoft, under which Microsoft has historically received a share of OpenAI’s revenue and OpenAI has received a share of certain Azure OpenAI Service revenue, is the archetype; Microsoft’s fiscal 2026 results disclosed that net gains from its investments in OpenAI added $4.963 billion to full-year net income, a reminder that the two companies’ financial statements are entangled at the level of equity-method accounting as well as revenue.[18] Anthropic’s prospectus, as reported by Reuters, disclosed $54.6 billion in non-cancellable hosting and computing commitments at the end of 2025, plans to spend at least $518 billion over a decade on infrastructure with six partners, and a provision under which Anthropic must pay Google the difference if its actual spend falls short of a committed level.[6][8] The company’s net loss of $42 billion for 2025 included roughly $34 billion of accounting charges tied to the estimated value of financing that could convert into shares, with the operating loss excluding that charge widening to $8.06 billion on operating expenses of $12.65 billion.[8]
These arrangements shape the distribution of economic value in ways that the revenue line alone cannot capture. A model developer that receives a $50 billion equity investment from a cloud provider and simultaneously commits to purchase $30 billion of compute from that provider over several years has, in economic substance, obtained vendor financing; the cloud provider’s investment returns to it as infrastructure revenue, and the model developer’s compute cost is partially funded by its own equity dilution. OpenAI’s March 2026 round of $122 billion at an $852 billion valuation included up to $50 billion from Amazon and $30 billion each from Nvidia and SoftBank, all of which are also suppliers to or distributors for the company.[47] The Bank of England’s Financial Policy Committee has described such financing arrangements as opaque and at times circular, and has warned that they could complicate risk analysis and amplify losses if the underlying revenue fails to materialize.[41]
Anthropic’s own filing offered the most candid statement of the dependency at the heart of these structures.
Demand for AI will be “limited principally by the availability of compute.”
— Anthropic IPO prospectus, as reported by Reuters[8]
If demand is limited principally by compute, then the party that controls compute controls the ceiling on the model developer’s revenue, and the terms on which compute is supplied, whether as a cost of revenue, a distribution fee, a take-or-pay commitment, or an equity-for-capacity exchange, determine how much of each dollar of customer spending the model developer ultimately retains. Gross margin, in such a structure, is not merely a financial ratio but a measure of bargaining power between the Models layer and the Datacenters layer of the Five-Layer AI Economy.
2.5 Developing the Revenue Denominator Framework
The preceding four subsections suggest that a consistent analytical comparison of AI companies requires more than a choice between gross and net. It requires a set of dimensions along which every revenue figure can be located before it is compared with any other. I propose five.
The first dimension is the measurement period. Is the figure a monthly run rate annualized, a quarterly figure annualized, a trailing-twelve-month recognized total, a fiscal-year recognized total, or a forward projection? The October confusion arose in significant part because a September run rate, a December target, and a calendar-year recognized total were all in circulation under the same label.
The second dimension is the reporting entity and its contractual role. For each channel through which the figure was generated, is the reporting entity the principal, the agent, a licensor receiving a royalty, or a supplier of inputs to someone else’s service? This is the gross-versus-net question, but framed to acknowledge that a single company can occupy different roles in different contracts.
The third dimension is distribution-channel treatment. What share of the figure flows through direct sales, through cloud marketplaces, through resellers and integrators, and through embedded distribution in third-party products? Anthropic’s 47 percent marketplace share and OpenAI’s lower but rising share are the single most important inputs to any gross-to-net adjustment, and neither would be visible from the headline figure alone.[6]
The fourth dimension is revenue retained after identifiable partner payments. Whatever the presentation, what does the model developer keep after distribution fees, revenue shares, marketplace commissions, and contractually committed payments to partners? This is the figure that most closely approximates the economic value accruing to the company, and under both gross and net presentation it is the same number, which is why it provides the common denominator the title of this paper calls for.
The fifth dimension is conversion of recognized revenue into cash. What share of recognized revenue has been collected, what share is held by third-party collectors, what is the aging of receivables, and how do deferred revenue and prepaid credits distort the relationship between the income statement and the cash flow statement? Anthropic’s disclosure that 60 percent of its outstanding bills were in third-party hands at the end of 2025 illustrates why this dimension cannot be ignored.[7]
Table 4
| Dimension | Question to ask | What the October 2026 episode revealed | Primary data needed |
| 1. Measurement period | Run rate, recognized, or projected? Over what window? | September run rate, December target, and calendar-year total were conflated | Monthly revenue series; fiscal-period recognized revenue; stated forecast dates |
| 2. Reporting entity and contractual role | Principal, agent, licensor, or supplier in each channel? | Anthropic principal (gross); OpenAI net on certain Microsoft arrangements | Contract summaries; auditor’s principal-agent conclusions; SEC correspondence |
| 3. Distribution-channel treatment | What share flows through direct, marketplace, reseller, and embedded channels? | 47% of Anthropic’s 2025 sales through AWS and Google | Channel mix disclosure by period |
| 4. Revenue retained after partner payments | What remains after fees, revenue shares, and commitments? | ~16% distribution fee; ~6–10% gross-to-net adjustment estimated | Partner payments by counterparty; revenue-share terms |
| 5. Conversion to cash | How much recognized revenue has been collected, by whom? | 60% of receivables held by third parties | Receivables aging; third-party collections; deferred revenue roll-forward |
The purpose of the framework is not to override the legally required accounting treatment of any company. Anthropic’s gross presentation and OpenAI’s net presentation may both be correct applications of ASC 606 to their respective contracts, and nothing in this framework asks either company to change its audited financial statements. The purpose is instead to give investors, lenders, and infrastructure planners a way to locate every revenue figure on a common grid before they compare it with another, so that the arithmetic of valuation, growth, and capital allocation is performed on a common denominator. Applied to the October 2026 figures, the framework would have immediately flagged that the $70 billion and the $50 billion differed on dimension 1 (period), dimension 2 (role), and dimension 3 (channel treatment), and that the correct comparison with Anthropic’s $65 billion required adjusting Anthropic’s figure down by 6 to 10 percent on dimension 4 or adjusting OpenAI’s figure up by its own, undisclosed, marketplace share, rather than applying a growth rate to an estimate.
Section 3: The Frontier-Model Valuation Problem — When Revenue Definitions Influence Trillion-Dollar Expectations
Valuation is the discipline in which the Revenue Denominator problem does its most visible damage, because valuation is the activity in which revenue is most directly converted into a dollar figure that is then transacted upon. In private markets, where neither OpenAI nor Anthropic has yet published audited financial statements, the revenue run rate has been the primary input into the multiples at which successive funding rounds have been priced, and those rounds have in turn set the reference points against which public-market infrastructure companies are valued. OpenAI closed a round at an $852 billion post-money valuation in March 2026 and has reportedly resisted any initial public offering valuation below $1 trillion; Anthropic was valued at $965 billion in a May 2026 Series H round and was reported to be targeting a valuation above $2 trillion in a listing expected after the November 2026 midterm elections.[8][46][47] At those magnitudes, a 19 percent difference in the definition of the revenue denominator corresponds to a difference in implied enterprise value measured in the hundreds of billions of dollars. This section examines how revenue multiples are applied to frontier laboratories, why revenue growth alone cannot establish sustainability, how bookings and commitments are being conflated with earned revenue, what customer concentration implies for revenue quality, and what a disciplined reset might look like between 2027 and 2030.
3.1 Revenue Multiples and the Valuation of Frontier AI Companies
The revenue multiple is the dominant valuation heuristic for companies that are growing rapidly and are not yet profitable, for the simple reason that there is no earnings figure to which a price-earnings ratio could be applied and no free cash flow from which a discounted cash flow could be constructed without heroic assumptions. In the software industry the multiple is conventionally applied to forward annual recurring revenue, and a decade of public-market data has established rough relationships between growth rate, gross margin, net revenue retention, and the multiple investors will pay. The frontier-model industry has imported this heuristic wholesale, but with three complications that make it far less reliable than it was for subscription software.
The first complication is the one this paper is about: the run rate to which the multiple is applied is not consistently defined across companies. On the figures available in early September 2026, OpenAI at an $852 billion valuation and a run rate exceeding $40 billion on its own basis was trading at roughly twenty times run rate, while Anthropic at $965 billion and a $65 billion gross run rate was trading at roughly fifteen times.[5][46][47] Adjusting Anthropic’s run rate down by 6 to 10 percent to a net basis raises its multiple toward seventeen; grossing up OpenAI’s run rate by its marketplace share lowers its multiple toward the mid-teens. The two companies’ valuations are therefore closer in multiple terms than the headline figures suggest, but no outside analyst can say precisely how close without channel-mix data that neither company has published.
The second complication is that the run rate is itself a moving target growing at rates that no public software company has sustained at comparable scale. A multiple of twenty times a run rate growing at 70 percent per year is a multiple of roughly twelve times the run rate twelve months hence, if the growth is realized, and a multiple of thirty-five times if growth halves. The sensitivity of the valuation to the growth assumption dwarfs the sensitivity to the gross-versus-net adjustment, which means that the most important denominator question is not the September 2026 figure but the credibility of the trajectory from $50 billion to $70 billion to the $100 billion that OpenAI’s chief executive has targeted for 2027.[46]
The third complication is gross margin, to which the next subsection turns. The valuation professor Aswath Damodaran has argued that AI companies come in many shapes and will need to be valued in tiers, and that for the majority of firms the technology may simply become a cost of doing business rather than a source of durable advantage.
“If everybody has AI, then nobody will have AI.”
— Aswath Damodaran, Professor of Finance, NYU Stern School of Business[37]
Damodaran’s observation applies with particular force to the revenue multiple. A multiple embeds an assumption about the durability of the revenue stream and the margin it will eventually earn, and if frontier-model capability converges across providers, the pricing power that would justify a software-style multiple may not persist. The October episode, in which a definitional change moved the market, is an early signal that investors are beginning to ask what sits behind the run rate rather than simply extrapolating it.
3.2 Gross Margins, Infrastructure Expenses, and Economic Profitability
Revenue growth, however defined, cannot by itself establish financial sustainability, and the frontier-model industry’s own disclosures make the point starkly. Anthropic’s 2025 operating loss of $8.06 billion on revenue of nearly $4.6 billion implies that the company spent roughly $2.75 for every dollar of revenue, and its compute and infrastructure spending of $7.33 billion in 2025 exceeded its entire revenue.[8][53] OpenAI reportedly burned $3.7 billion in the first quarter of 2026 against revenue of $5.7 billion, with both figures tripling year over year, and its 2025 revenue of $13 billion was accompanied by spending of $8 billion according to a source cited by Reuters.[45][46] These are not the economics of software businesses at comparable revenue scale; they are the economics of capital-intensive infrastructure businesses whose primary input, compute, is purchased from a concentrated set of suppliers that are also their investors and distributors.
The gross-versus-net question interacts with margin analysis in a way that is frequently overlooked. Under gross presentation, the distribution fee paid to a cloud partner is recorded as an expense, and depending on classification it may sit in cost of revenue, depressing gross margin, or in sales and marketing, depressing operating margin but leaving gross margin intact. Under net presentation, the fee is simply absent from both revenue and expense, and gross margin is computed on the smaller base. Two companies with identical economics can therefore report different gross margins as well as different revenues, and an analyst who compares gross margins across the two without adjusting for classification will draw the wrong conclusion about relative efficiency. The Fortune commentary on flawed AI valuations observed that many AI companies have true gross margins in the low teens to twenties once the full cost of inference is allocated, far below the 70 to 80 percent margins that justify software multiples, and that the reporting of annual recurring revenue has itself been a weakness in the sector’s valuation practices.[54]
Beyond gross margin lie the costs that no presentation choice can hide: model training, which must be repeated for each generation at escalating cost; research and development; depreciation on owned infrastructure; stock-based compensation, which is enormous at companies competing for a small pool of researchers; and the financing costs embedded in take-or-pay compute commitments. Anthropic’s prospectus reportedly projected breakeven by 2028 and OpenAI’s internal projections reportedly extended to 2030, with OpenAI carrying $1.4 trillion in long-term compute commitments.[47] A valuation that rests on a revenue multiple is, implicitly, a bet that these costs will eventually fall as a share of revenue, and the credibility of that bet depends on a margin trajectory that the headline run rate reveals nothing about.
3.3 Bookings, Contracted Revenue, and Future Commercial Commitments
The third distortion in frontier-model valuation is the increasingly common practice of citing future commitments as though they were current revenue. The practice is most visible in the infrastructure layer, where Oracle’s $664 billion of remaining performance obligations and CoreWeave’s $104 billion revenue backlog have become headline numbers that dwarf the companies’ recognized revenue, but it has migrated upward into the Models layer as well.[20][22] OpenAI’s $1.4 trillion of compute commitments, Anthropic’s $518 billion of planned infrastructure spending, and the multi-year enterprise contracts both companies increasingly sign are all forward commitments, and the market has shown a tendency to treat them as evidence of revenue that has already been earned.[8][44]
The accounting distinctions here are precise and important. Remaining performance obligations under ASC 606 represent the transaction price allocated to performance obligations that are unsatisfied or partially unsatisfied at the reporting date; they are disclosed in the notes to the financial statements, they are subject to the entity’s judgment about which contracts are enforceable, and they do not include optional purchases or contracts the customer can terminate without penalty. Backlog, as CoreWeave defines it, includes remaining performance obligations plus other amounts the company estimates will be recognized under committed customer contracts, in each case subject to the satisfaction of delivery and availability requirements, which is to say it is a management measure that extends beyond the audited figure and is conditional on the company building the capacity to serve it.[22] Oracle’s own disclosure that only 12 percent of its $638 billion of remaining performance obligations at the end of fiscal 2026 was expected to be recognized in the following twelve months, with a further 34 percent between thirteen and thirty-six months, illustrates how far the headline number sits from near-term revenue.[21]
Purchase commitments on the model developer’s side are the mirror image. Anthropic’s $54.6 billion of non-cancellable hosting and computing commitments at the end of 2025 are liabilities, not assets, and the provision requiring it to pay Google the difference if its actual spend falls short is a take-or-pay obligation that converts a cost of revenue into something closer to fixed debt service.[6][8] OpenAI’s chief executive has stated that the company is committed to roughly 30 gigawatts of compute with a total cost of ownership of about $1.4 trillion.
OpenAI is “comfortable with this given what we see on the horizon.”
— Sam Altman, CEO, OpenAI, on $1.4 trillion of compute commitments[44]
The risk the Revenue Denominator framework highlights is that a market which does not distinguish clearly between recognized revenue and run rate is unlikely to distinguish clearly between a supplier’s backlog and a customer’s commitment either, and the two figures describe the same contracts from opposite sides. Oracle’s $664 billion is in large part OpenAI’s and other laboratories’ future spending; CoreWeave’s $104 billion is in large part Microsoft’s and OpenAI’s. If the revenue of the laboratories disappoints, the backlog of the infrastructure providers is impaired, and valuing the backlog as though it were earned revenue compounds the original error.
3.4 Customer Concentration and the Quality of AI Revenue
The fourth dimension of the valuation problem is concentration, which affects the quality of revenue independent of how it is presented. Anthropic’s prospectus disclosed that 12 percent of its 2025 revenue came from two unnamed customers, that 47 percent was routed through two cloud partners, and that those partners were also its largest suppliers and among its largest investors.[7] CoreWeave’s 2025 revenue was 67 percent dependent on Microsoft.[23] Oracle’s remaining performance obligations are widely understood to be dominated by a small number of frontier-laboratory contracts. Nvidia’s datacenter revenue, while diversified across hyperscalers, neoclouds, and sovereign buyers, is ultimately purchased by a customer base whose own revenue is concentrated in the two laboratories whose accounting this paper has examined.
Concentration of this kind has three implications for the revenue denominator. First, it means that the channel through which revenue flows is also a point of control over the customer relationship, pricing, and data, so that revenue recognized gross through a marketplace is of lower quality than revenue recognized gross through a direct contract even if the accounting treatment is identical. Second, it means that a change in the commercial terms between a model developer and a single cloud partner, whether a renegotiated distribution fee, a revised revenue share, or a shift from marketplace to direct billing, can move reported revenue by a double-digit percentage without any change in end-customer demand. Third, it means that the exposure of the infrastructure layers to the revenue of the model layers is far more concentrated than the number of counterparties suggests, because the same two or three companies sit at the end of most of the chains. Investors assessing revenue durability must therefore look past the recognized figure to retention, pricing competition among models that are converging in capability, the risk that a cloud partner promotes its own model over a partner’s, and the sensitivity of consumption to the next generation’s benchmark results.
3.5 The 2027–2030 Frontier-Model Valuation Reset
The forces examined in this section point toward a reset in how frontier laboratories are valued between 2027 and 2030, and the October 2026 episode may be remembered as its first visible tremor. The catalyst will be the transition from private to public markets. Anthropic filed its confidential registration statement on June 1, 2026, and OpenAI on June 8; Anthropic’s listing is expected after the November midterms and OpenAI’s is now reportedly more likely in 2027 after advisers warned that the $1 trillion valuation its chief executive insists upon may not be achievable in current conditions.[46][47] When those listings occur, three things will happen that have not happened before in this industry. Audited financial statements prepared under ASC 606 will replace run rates as the primary revenue figure. The correspondence between each company and the SEC’s Division of Corporation Finance will be made public, including, as Axios noted, the justifications each company offered for its revenue-recognition treatment.[5] And quarterly reporting will impose a discipline of comparability over time, if not yet across companies, that private funding rounds never did.
The reset will not necessarily mean lower valuations. It will mean valuations constructed from different inputs: recognized revenue rather than run rate, gross margin after full allocation of inference cost, operating cash flow after take-or-pay commitments, customer concentration disclosed by counterparty, and remaining performance obligations reconciled to delivery capacity. Some companies will emerge from that scrutiny with their valuations intact or enhanced, because durable demand and improving unit economics will be visible for the first time. Others will find that the run rate had been doing a great deal of work that the audited figures cannot sustain. The Revenue Denominator framework is intended to help investors anticipate which is which before the filings are public rather than after, and the next section turns to why that anticipation matters for the trillions of dollars of infrastructure being built on the assumption that the run rates are real.
Section 4: The Infrastructure Feedback Loop — How AI Revenue Expectations Influence Chips, Datacenters, Energy, and Capital Expenditure
If the preceding sections have established that the revenue figures at the top of the AI economy are defined inconsistently and that this inconsistency distorts the valuation of the companies that report them, this section establishes why the problem does not stay at the top. The Five-Layer AI Economy is a vertically coupled system in which the revenue expectations of the Models and Applications layers are the demand signal to which every lower layer responds. Semiconductor manufacturers size production against it. Hyperscalers set capital expenditure guidance against it. Datacenter developers and their lenders underwrite construction against it. Utilities and power producers plan generation and transmission against it. In each case the planning horizon is years and the commitments are largely irreversible, which means that the signal is being acted upon long before it can be verified, and an error in the signal is embedded in physical capital before anyone discovers it. The scale of the commitments now being made against that signal is without precedent in the history of corporate investment, and the October 2026 episode was the first occasion on which the market was forced to confront the possibility that the signal itself had been misread.
4.1 From Revenue Forecasts to Semiconductor Procurement
The most immediate recipient of the revenue signal is the semiconductor layer, and the magnitude of the response is visible in Nvidia’s results for the quarter ended July 26, 2026. Revenue of $96.2 billion was up 106 percent from a year earlier and 18 percent from the prior quarter; datacenter revenue of $89.0 billion was up 117 percent and accounted for roughly 92 percent of the total; GAAP and non-GAAP gross margins were both 75.0 percent; and the company guided to third-quarter revenue of $108.0 billion, plus or minus 2 percent, while assuming no datacenter compute revenue from China.[16] Within the datacenter segment, hyperscale customers contributed $48.7 billion, more than doubling year over year, while AI cloud providers, industrial, and physical-enterprise customers contributed $40.3 billion, growing 138 percent.[17] Management indicated that it expected fiscal 2028 revenue growth of approximately 70 percent and described even that figure as supply-constrained.[17] The company returned approximately $26 billion to shareholders in the quarter and ended it with a market capitalization that reached $5.49 trillion following a single-day gain of $441.5 billion.[16][17]
Nvidia’s chief financial officer was explicit that the forward outlook excluded an entire geography.
The outlook “assumes no Data Center compute revenue from China.”
— Colette Kress, CFO, NVIDIA, Q2 FY2027 earnings call[52]
Every one of these figures is downstream of an assumption about demand for tokens. Nvidia does not sell to enterprises that use Claude or ChatGPT; it sells to the hyperscalers, neoclouds, and laboratories that serve those enterprises, and its customers’ willingness to purchase accelerators at 75 percent gross margins depends on their expectation that the compute will be consumed at prices that justify the capital. A 70 percent growth guidance for fiscal 2028 is, in effect, a forecast that the aggregate run rate of the Models and Applications layers will continue to compound at a rate sufficient to absorb a doubling of installed capacity. The October episode did not change Nvidia’s order book, but it changed the market’s confidence in the signal on which the order book rests, and the share price responded accordingly.[3] The same logic applies, with varying intensity, to AMD’s accelerator business, to Broadcom’s custom silicon for the hyperscalers’ in-house chips, to the high-bandwidth memory suppliers whose pricing Microsoft cited as adding $25 billion to its 2026 capital budget, and to the networking vendors whose products connect the clusters.[24]
4.2 Cloud Contracts and Long-Term Infrastructure Obligations
The second recipient of the signal is the cloud layer, which converts the laboratories’ revenue expectations into multi-year contracts and then into capital expenditure. The second-quarter 2026 earnings season made the scale visible. Microsoft reported that Azure and other cloud services revenue grew 43 percent in the quarter ended June 30, 2026, that Azure surpassed $100 billion in annual revenue for the first time, and that Microsoft Cloud revenue for the fiscal year exceeded $214 billion; it guided to approximately 45 percent Azure growth in constant currency for the following quarter and to calendar-2026 capital expenditure of roughly $190 billion, including $25 billion attributable to higher component pricing.[18][19][24] Alphabet raised its 2026 capital expenditure guidance to $195 billion to $205 billion after reporting second-quarter spending of $44.9 billion, double the year-earlier figure, and signaled that 2027 spending would increase significantly.[26][27] Amazon indicated that its plan to approach $200 billion in 2026 capital expenditure remained largely intact, with some reporting placing the revised figure as high as $220 billion; Meta raised its range to $125 billion to $145 billion and then lifted the floor to $130 billion.[25][26] Combined 2026 capital expenditure guidance from the four companies was variously estimated at $725 billion, up 77 percent from a record $410 billion in 2025, and as high as $760 billion.[24][25]
Not every investor regarded this as cause for concern. One sell-side analyst dismissed the skeptics in terms that captured the prevailing mood in the spring.
“The bear thesis is garbage.”
— Brent Thill, analyst, Jefferies, to the Financial Times[24]
The contracts underlying this spending are the mechanism through which the revenue signal becomes a fixed obligation. Oracle’s remaining performance obligations rose from $638 billion at the end of fiscal 2026 to $664 billion at the end of its fiscal first quarter, after it booked more than $30 billion of additional AI cloud contracts in a single quarter; its cloud infrastructure revenue grew 121 percent to $7.4 billion; its operating cash flow was a record $23 billion; and its free cash flow was negative $5 billion as it continued to invest, having raised $43 billion of debt and $5 billion of equity in fiscal 2026 and completed a $20 billion at-the-market equity sale in the first quarter.[20][21] Oracle’s fiscal 2026 free cash flow was negative $23.7 billion.[21] A company with $664 billion of contracted future revenue and negative free cash flow is, in the most literal sense, borrowing against its customers’ revenue forecasts, and those customers are the laboratories whose revenue denominators the market cannot agree upon.
The allocation of risk in these contracts is therefore the central question. Take-or-pay arrangements, minimum-spend commitments, and capacity reservations shift demand risk from the infrastructure provider to the model developer, which is why Anthropic’s filing discloses that it must pay Google the difference if its spending falls short.[8] But the shift is only as good as the model developer’s ability to pay, and the model developer’s ability to pay depends on the revenue that the October episode showed to be less well-defined than assumed. In the limit, a take-or-pay commitment from a customer whose revenue disappoints is not a transfer of risk but a deferral of its recognition.
4.3 Datacenter Financing and the Economics of Expected Demand
The third recipient of the signal is the capital market that finances datacenter construction, and here the transmission mechanism runs through credit rather than equity. CoreWeave’s second-quarter 2026 results offer the clearest window. Revenue of $2.575 billion grew 112 percent year over year; the company spent $9.4 billion on capital expenditure in the quarter and raised its full-year revenue guidance to $12.4 billion to $13.2 billion while raising its capital expenditure guidance by $4 billion; it ended the quarter with roughly $35 billion of debt against $5.5 billion of cash; it raised more than $10 billion of unsecured debt and convertible bonds in the quarter including an inaugural Eurobond; and its net loss of $626 million was, as one analysis observed, within a rounding error of the net interest expense of $640 million it paid on the debt that built the datacenters.[22][23] Contracted power stood at 3.7 gigawatts at quarter-end and 4.2 gigawatts by the date of the earnings release.[23]
The broader credit picture is consistent. Reuters reported that AI datacenter and project financing deals surged to $125 billion in the first portion of 2025 from $15 billion in the same period of 2024, that Morgan Stanley estimated private credit could supply more than half of the $1.5 trillion needed for datacenter construction through 2028, that high-yield issuance for AI projects had set records, and that Oracle’s credit default swaps had closed at their highest level since 2009.[43] The Bank of England’s Financial Policy Committee recorded in September 2026 that the growing indebtedness of AI firms and their opaque and at times circular financing arrangements could complicate risk analysis and amplify losses, and noted Morgan Stanley’s estimate that $700 billion of datacenter capital expenditure between 2026 and 2028 would be funded by private credit.[41] The Governor of the Bank of England subsequently wrote to G20 finance ministers warning that a future market correction could ripple across borders through the financial ties between AI developers and the hyperscalers building their infrastructure, and the committee’s minutes have repeatedly described AI-focused valuations as stretched.
Valuations for AI-focused technology firms “appear stretched.”
— Bank of England Financial Policy Committee, as cited by Capacity[42]
A lender underwriting a datacenter loan is making a judgment about the creditworthiness of the tenant, and the tenant’s creditworthiness is a function of its revenue. When the tenant is a hyperscaler with a diversified business and an investment-grade balance sheet, the lender is one step removed from the laboratories’ revenue. When the tenant is a neocloud whose own revenue is two-thirds dependent on a single hyperscaler customer and whose second-largest customer is a laboratory, the lender is two steps removed. When the tenant is the laboratory itself, as in several of the largest recent projects, the lender is directly exposed to the denominator problem. In none of these cases has the lender had access to a reconciled, consistently defined revenue figure for the ultimate source of demand.
4.4 The Five-Layer AI Economy’s Financial Transmission Mechanism
The three preceding subsections can be assembled into a single description of how a revenue expectation formed at Layer 5 and Layer 4 propagates downward. An enterprise commits to a multi-year consumption contract for a frontier model, perhaps through a cloud marketplace. The model developer books the contract, raises its run rate, and uses the enlarged run rate to justify a larger compute commitment to a cloud provider or neocloud, often financed in part by an equity investment from that same provider. The cloud provider books the commitment as remaining performance obligations, raises its capital expenditure guidance, and places orders with the semiconductor layer. The semiconductor layer books the orders, guides to supply-constrained growth, and commits wafer capacity with its foundry and memory suppliers. The datacenter developer signs a lease with the cloud provider or neocloud, uses the lease to secure construction financing from banks and private credit, and contracts with the energy layer for power. The energy layer plans generation and transmission against a load forecast derived, ultimately, from the enterprise’s consumption contract several layers above.
Table 5
| Layer | How the revenue signal arrives | Form of commitment made against it | Q2–Q3 2026 scale indicator | Reversibility |
| 5. Applications / Agents | Enterprise and consumer demand; run rate | Multi-year consumption contracts; guaranteed capacity | OpenAI ~$50B run rate; Anthropic ~$65B (gross)[3][4] | High (contracts cancellable with notice) |
| 4. Models | Run rate; bookings | Take-or-pay compute commitments; equity-for-capacity | Anthropic $54.6B non-cancellable commitments; OpenAI $1.4T[6][44] | Low–medium |
| 3. Datacenters / Cloud | Customer commitments; RPO; backlog | Capex guidance; construction leases; debt issuance | Hyperscaler capex $725–760B; Oracle RPO $664B; CoreWeave $35B debt[20][22][24][25] | Low (multi-year construction) |
| 2. Chips | Hyperscaler and neocloud orders | Foundry and memory capacity; supply-constrained guidance | Nvidia $89.0B quarterly DC revenue; FY2028 ~70% growth guide[16][17] | Low (wafer commitments) |
| 1. Energy | Datacenter load forecasts | Generation, transmission, and interconnection investment | CoreWeave 4.2 GW contracted power; Bain 200 GW incremental demand by 2030[23][39] | Very low (decades) |
The critical feature of this mechanism is that reversibility declines as the signal descends. The enterprise at the top can reduce its consumption or cancel its contract with ninety days’ notice. The utility at the bottom cannot un-build a substation. An error in the revenue signal that is corrected within a quarter at Layer 5 may already have been converted into a thirty-year asset at Layer 1. This asymmetry is why the quality of the revenue signal, and not merely its growth rate, matters so much. A $20 billion definitional revision at Layer 5 that reflects no change in economic substance is harmless if everyone understands it as such; it is harmful if even a fraction of the lower layers have been planning against the unrevised figure.
The macroeconomic significance of the mechanism has been quantified by economists on both sides of the debate. Harvard’s Jason Furman calculated that investment in information-processing equipment and software, though only about 4 percent of U.S. GDP in the first half of 2025, accounted for 92 percent of GDP growth over that period, implying growth of only 0.1 percent without it.[32]
“Investment in information processing equipment & software is 4% of GDP.”
— Jason Furman, Professor of the Practice of Economic Policy, Harvard University, on X[33]
Fortune’s coverage of Furman’s analysis quoted a widely circulated quip that captured the concentration of the growth story.
“Our economy might just be three AI data centers in a trench coat.”
— Rusty Foster, Today in Tabs, as quoted by Fortune[32]
The International Monetary Fund’s World Economic Outlook authors noted in early 2026 that information-technology investment as a share of U.S. output had reached its highest level since 2001, and estimated that AI could add roughly 0.3 percentage points to annual growth if it delivers, or trigger a major equity correction if it does not.[51]
IT investment as a share of US output has “surged to the highest level since 2001.”
— Tobias Adrian and Pierre-Olivier Gourinchas, International Monetary Fund[51]
The IMF’s Managing Director, speaking on October 7, 2026, days before the OpenAI revenue episode, warned that the AI investment boom was one of several simultaneous inflationary forces, that AI investment as a share of GDP was likely to exceed the railroad, electricity-grid, and telecommunications build-outs, that AI hardware now accounted for more than a tenth of global goods trade, and that if markets were disappointed the fallout could be far-reaching.[30][31] A year earlier she had cautioned that valuations were approaching the levels seen in the internet bubble.
Valuations are “heading toward levels we saw during the bullishness about the internet.”
— Kristalina Georgieva, Managing Director, International Monetary Fund[28]
The IMF’s chief economist offered the more measured assessment that a dot-com-style bust was possible but unlikely to be systemic, observing that AI-related investment had increased by less than 0.4 percent of U.S. GDP since 2022 against 1.2 percent during the 1995–2000 build-out, and that the current boom was elevating demand and inflation without associated productivity gains.[29] The disagreement among these assessments is itself instructive: an economy whose growth is concentrated in a single investment category is an economy whose macroeconomic trajectory depends on the quality of that category’s revenue signal, and in October 2026 the signal was shown to be undefined at its source.
4.5 The Risk of Financing Infrastructure Against Incomparable Metrics
The final subsection of this section draws the threads together into an assessment of the specific risk that incomparable revenue metrics pose to infrastructure financing, and proposes a set of stress scenarios. The risk is not that AI demand is illusory. The adoption data are unambiguous: Stanford’s 2026 AI Index reported that global corporate AI investment more than doubled in 2025 to $581.7 billion, that generative AI reached 53 percent population adoption within three years of mass-market introduction, faster than the personal computer or the internet, that 70 percent of organizations now use generative AI in at least one business function, and that the estimated consumer value of generative AI tools in the United States reached $172 billion annually by early 2026.[35] The risk is rather that the conversion of adoption into revenue, revenue into cash, and cash into debt service has been planned against figures that were not measured consistently.
The gap between adoption and monetization has been quantified from several directions. Bain & Company estimated in September 2025 that $2 trillion of annual revenue would be required by 2030 to fund the computing power needed to meet anticipated demand, that even with AI-related savings the world was roughly $800 billion short, and that global incremental compute requirements could reach 200 gigawatts with the United States accounting for half.[39][40]
“If the current scaling laws hold, AI will increasingly strain supply chains globally.”
— David Crawford, Chairman, Global Technology Practice, Bain & Company[39]
MIT’s Project NANDA reported in 2025 that 95 percent of organizations in its research saw no measurable return from generative AI initiatives despite enterprise spending of $30 billion to $40 billion, and that the barriers were organizational rather than technological.[38]
“Only 5% of custom enterprise AI tools reach production.”
— MIT Project NANDA, The GenAI Divide: State of AI in Business 2025[38]
Stanford’s Erik Brynjolfsson, who has argued that AI’s productivity gains will arrive on a J-curve after a period of investment that depresses measured productivity, has cautioned against the assumption that the gains must come from labor substitution.
“The only way that you get productivity from A.I. is by removing labor costs.”
— a fallacy identified by Erik Brynjolfsson, Director, Stanford Digital Economy Lab[34]
MIT’s Daron Acemoglu, by contrast, has estimated that the total factor productivity gains from AI over the coming decade are likely to be modest, on the order of well under one percentage point cumulatively, and far smaller than the figures implied by the investment boom.[36] Sequoia Capital’s David Cahn framed the same tension in mid-2024 as a $600 billion question, the annual revenue the industry would need to generate to justify the infrastructure then being purchased, and the question has grown with the capital expenditure since.[48] Goldman Sachs Research has put cumulative capital spending by the largest technology companies on AI and datacenters from 2025 through 2030 at $5.3 trillion.[49]
Against this backdrop, the following stress scenarios illustrate how a mismatch between revenue expectations and realization would propagate through the layers. The scenarios are illustrative constructions rather than forecasts, and they are deliberately framed in terms of the denominator rather than the growth rate.
Table 6
| Scenario | Trigger at Layers 4–5 | Transmission to Layer 3 | Transmission to Layers 1–2 | Financial-stability channel |
| A. Definitional correction only | Market adopts a net, recognized-revenue basis; headline run rates fall 10–20% with no change in cash | Multiples compress; capex guidance unchanged; RPO unchanged | Order books unchanged; equity volatility only | Limited; a repricing of expectations |
| B. Slower revenue growth | Laboratory run-rate growth halves from ~70% to ~35%; year-end targets missed | Take-or-pay commitments become burdensome; some RPO deferred or renegotiated; neocloud backlog conversion slows | Fiscal 2028 chip growth guidance cut; wafer commitments renegotiated | Neocloud and single-tenant datacenter credit spreads widen |
| C. Price compression | Per-token pricing falls faster than volume rises; gross margins fall toward the teens | Cloud AI revenue grows but margin per gigawatt falls; capex returns extend | Demand for capacity persists but at lower value per unit; memory and accelerator pricing under pressure | Equity-like risk in debt instruments, as BoE and UBS have warned |
| D. Higher financing costs | Rates rise or credit spreads widen; equity-for-capacity rounds become harder to close | Oracle, CoreWeave, and private developers face higher debt service against unchanged RPO | Construction delayed; power contracts stranded | Private credit marks fall; refinancing risk concentrated in 2027–2029 |
| E. Delayed profitability | Breakeven slips from 2028–2030 to beyond 2030; laboratories require further equity | Suppliers that are also investors face dilution and receivable risk simultaneously | Capex becomes contingent on laboratory funding rounds | Circular financing unwinds in reverse; concentrated losses |
The purpose of the table is to make a specific point: scenario A, the pure definitional correction, is the only one that is costless, and it is costless only if it happens before the lower layers have embedded the uncorrected figures in irreversible commitments. The October 2026 episode was a small instance of scenario A. Its lesson is that the industry would be well served by completing the definitional correction deliberately, through disclosure, rather than experiencing it repeatedly through market shocks, before scenarios B through E have the opportunity to arrive.
Section 5: The Road to 2030 — Financial Transparency, Investor Accountability, and the Next Reporting Framework for the AI Economy
The argument of this paper has so far been diagnostic. It has shown that the revenue figures on which the AI economy is being financed are defined inconsistently, that the inconsistency arises from legitimate differences in commercial structure rather than from error or misconduct, that it distorts valuation at the top of the stack, and that it propagates downward into irreversible infrastructure commitments whose scale is without precedent. This section turns to remedy. It examines how the impending public listings of the leading laboratories will change the disclosure environment, reviews the existing regulatory framework and distinguishes what it requires from what it does not, proposes a voluntary reconciliation disclosure that would address the Revenue Denominator problem without altering any company’s audited accounting, considers how better comparability would reshape the behavior of each class of capital provider, and sketches three possible financial futures through 2030. The premise throughout is the one stated in the paper’s fifth pillar: the goal is not to force every AI company to account identically, which would be both impossible and wrong, but to make the differences legible.
5.1 Public-Market Disclosure and Frontier AI IPO Readiness
The transition of OpenAI and Anthropic from private to public companies will do more to resolve the Revenue Denominator problem than any voluntary framework, because it will replace run rates communicated to investors in presentations with audited financial statements filed with the SEC, and because the registration process itself forces the disclosure of the judgments that have so far been private. Anthropic submitted its confidential draft registration statement on June 1, 2026, and OpenAI on June 8; Reuters reported that Anthropic’s prospectus ran to 261 pages, roughly 80 of them devoted to risk factors, and that the company was targeting a valuation above $2 trillion in a listing likely to follow the November midterm elections.[8][47] OpenAI, having targeted a listing as early as September 2026, was by the autumn reportedly leaning toward 2027 after advisers presented a choice between waiting for a $1 trillion valuation and accepting a lower figure sooner.[46]
Three features of the registration process bear directly on the denominator. First, the audited financial statements will present recognized revenue under ASC 606, with the principal-versus-agent conclusion for each material channel disclosed in the revenue-recognition policy note and the disaggregation of revenue required by the standard providing at least some visibility into channel mix. Second, the SEC staff’s comment letters and the companies’ responses, which are released publicly after the registration becomes effective, will contain the staff’s questions about the gross-versus-net treatment and the companies’ justifications; Axios observed that this correspondence should finally show investors how Anthropic justifies its revenue treatment.[5] Third, the risk-factor and management’s-discussion sections will be required to disclose customer concentration, related-party transactions with cloud partners that are simultaneously investors, suppliers, distributors, and customers, and the non-cancellable commitments and take-or-pay provisions that Reuters has already reported from the confidential filing.[6][8]
What the registration process will not do, however, is impose comparability across the two companies. Each will present its revenue in accordance with its own contracts and its own auditor’s conclusions, and if those conclusions differ, as they appear to, the two sets of audited statements will be as incomparable on the headline line as the run rates were. Public disclosure converts the problem from one of opacity to one of legibility: investors will be able to see the difference, but they will still have to do the reconciliation themselves, and only if the disclosures provide the inputs the reconciliation requires.
5.2 SEC Oversight, Accounting Standards, and Independent Audits
It is worth being precise about what the existing framework already requires, because a good deal of commentary on the October episode implied either that the companies were doing something irregular or that new rules were needed to stop them. Neither is accurate. The principal-versus-agent assessment under ASC 606, as clarified by ASU 2016-08, is a required judgment that every registrant with intermediated sales must make, document, and have audited; it turns on control of the specified good or service before transfer, supported by indicators including primary responsibility for fulfillment, inventory risk, and pricing discretion.[12][13][56] A company that concludes it is the principal and presents gross, with the auditor’s concurrence, is complying with the standard; so is a company that concludes it is an agent and presents net. The standard contemplates that different contractual structures will produce different conclusions.
The run rate, by contrast, sits outside the audited financial statements entirely. When a public company presents an annualized revenue run rate or similar metric in its management’s discussion and analysis, it does so as a key performance indicator subject to the SEC’s interpretive guidance of January 30, 2020, Release 33-10751, which became effective on February 25, 2020.[14] That guidance does not define ARR or any other metric; it requires that a company presenting a metric provide a clear definition of it and how it is calculated, a statement of the reasons the metric provides useful information to investors, and a statement of how management uses the metric, and that if the company changes the method of calculation it disclose the change and its effects and consider recasting prior periods.[14][15] The guidance also reminds companies that metrics presented outside SEC filings, such as in investor presentations, warrant similar care, and that disclosure controls and procedures should extend to them.[15] Public software companies have long complied by defining ARR in their filings; Guidewire’s fiscal 2020 annual report, for example, defined ARR as the annualized recurring value of active customer contracts at period end and explicitly noted that this may not be the same as the timing and amount of revenue recognized.[50]
The implication is that once OpenAI and Anthropic are public, any run rate they present will have to be defined, and any change in definition will have to be disclosed. That alone would have prevented the October confusion, because a defined metric cannot be grossed up by third parties and reattributed to the company without the discrepancy being obvious. What the existing framework does not require, and what this paper argues is needed, is a reconciliation that would allow the defined metrics of different companies to be placed on a common basis.
5.3 Proposed AI Revenue Reconciliation Disclosures
I propose that frontier-model developers, and by extension the cloud providers and neoclouds whose revenue is derived from them, adopt a voluntary reconciliation disclosure organized around the five dimensions of the Revenue Denominator framework. The disclosure would be supplementary to, not a substitute for, the audited financial statements, and would be presented quarterly in the format of a reconciliation table in the style of the non-GAAP reconciliations that public companies already provide. Its elements would be as follows.
Table 7
| Disclosure element | Content | Dimension addressed | Rationale |
| 1. Recognized revenue by channel | ASC 606 revenue disaggregated into direct consumer, direct enterprise, API, cloud marketplace (by partner), and embedded/other | 2, 3 | Makes channel mix and the gross-versus-net perimeter visible |
| 2. Gross-to-net bridge | For each channel presented gross, the partner fees and revenue shares that a net presentation would exclude; for each channel presented net, the gross transaction value | 2, 4 | Produces a common-denominator revenue figure regardless of accounting conclusion |
| 3. Run-rate definition and calculation | Stated formula; base period; treatment of one-time items; reconciliation to trailing recognized revenue | 1 | Satisfies the spirit of Release 33-10751 and prevents third-party reattribution |
| 4. Material partner payments by counterparty | Distribution fees, revenue shares, and compute payments to each partner above a materiality threshold, identifying partners that are also investors or customers | 4 | Exposes circularity and bargaining power |
| 5. Customer and channel concentration | Share of revenue and receivables from top customers and top channels | 3, 5 | Revenue quality |
| 6. Contractual commitments and conditionality | Non-cancellable compute commitments; take-or-pay provisions; RPO where the entity is supplier; delivery conditions | 4 | Distinguishes commitments from earned revenue on both sides |
| 7. Cash conversion | Recognized revenue to billings to collections; receivables held by third parties; deferred revenue roll-forward | 5 | Links income statement to cash |
| 8. Forward metric labeling | Any forward-looking run-rate target labeled with its date and basis and separated from current-period figures | 1 | Prevents the September–December conflation of October 2026 |
The disclosure would allow an analyst to take Anthropic’s gross marketplace revenue and OpenAI’s net partner revenue and place both on a common basis in either direction, to compute each company’s revenue retained after partner payments, to compare run rates defined identically, and to assess cash conversion and concentration side by side. It would also allow lenders to the infrastructure layers to look through their borrowers’ backlogs to the reconciled revenue of the ultimate demand source. None of it requires a change to ASC 606, to IFRS 15, or to any company’s audited statements, and all of it is within the discretion of management to provide today.
The precedent for voluntary metric standardization in a fast-growing sector is encouraging. The software-as-a-service industry converged on common definitions of ARR, net revenue retention, and customer acquisition cost through investor pressure and competitive emulation rather than regulation, and companies that adopted the clearest definitions earliest were rewarded with lower costs of capital. The frontier-model industry, approaching its first public listings at valuations an order of magnitude larger than any software IPO, has a correspondingly larger incentive to do the same.
5.4 How Better Revenue Comparability Could Reshape Capital Markets
Each class of capital provider in the Five-Layer AI Economy would use a reconciled revenue figure differently, and the aggregate effect of their doing so would be a more efficient allocation of the trillions of dollars now being committed. Institutional equity investors in the public laboratories would be able to construct cross-company revenue multiples on a common denominator, which would narrow the dispersion of valuations attributable to definitional noise and widen the dispersion attributable to genuine differences in growth, margin, and durability. Sovereign wealth funds and the very large private investors that have anchored the recent funding rounds would be able to underwrite equity-for-capacity exchanges with a clear view of how much of their investment returns to them as infrastructure revenue and how much represents a net claim on the laboratory’s future cash flows.
Banks and private-credit funds financing datacenters would be able to look through a neocloud’s or hyperscaler’s backlog to the reconciled revenue of the tenants, assess concentration at the level of the ultimate demand source rather than the immediate counterparty, and price the equity-like risk that the Bank of England and UBS have identified in AI debt instruments more accurately.[41][43] Hyperscalers setting capital expenditure guidance would be able to calibrate their internal demand forecasts against a consistently measured external signal rather than against run rates that each laboratory defines for itself. Semiconductor companies guiding to supply-constrained growth would be able to distinguish demand that is backed by reconciled, collected revenue from demand that is backed by commitments contingent on future funding rounds. Venture investors in the Applications and Agents layer would be able to benchmark their portfolio companies’ metrics against those of the laboratories on which they depend. And utilities and energy developers, whose planning horizons are the longest of all, would have a somewhat firmer basis for the load forecasts against which they are committing generation and transmission capital that will outlive every contract in the chain.
The common thread is that comparability lowers the risk premium that each provider must charge for definitional uncertainty, and a lower risk premium across five layers compounds into a materially lower cost of capital for the system as a whole. In an industry that Bain estimates will need to deploy $500 billion of annual capital expenditure and find $2 trillion of annual revenue by 2030, the value of that reduction is not trivial.[39]
5.5 Three Possible Financial Futures Through 2030
The paper concludes its analytical sections by sketching three scenarios for how the Revenue Denominator problem might evolve between now and 2030. They are not predictions but structured possibilities, each of which follows from a different resolution of the tension between the industry’s communication practices and its capital requirements.
Table 8
| Dimension | Transparency-led future | Fragmented future | Market-correction future |
| Revenue disclosure | Laboratories adopt reconciliation disclosures voluntarily or under investor pressure; public filings make channel mix and gross-to-net bridges standard | Each company continues to emphasize the metric most favorable to it; run rates proliferate with incompatible definitions | Weak revenue realization forces disclosure retroactively through earnings misses, impairments, and restatements of expectations |
| Valuation behavior | Multiples converge on reconciled revenue; dispersion reflects fundamentals | Multiples remain noisy; definitional arbitrage persists between public and private markets | Multiples compress sharply; run-rate-based valuations are abandoned |
| Cloud partnerships | Partner economics disclosed by counterparty; distribution fees become a negotiated, visible variable | Opacity persists; circular financing continues to complicate risk analysis | Partnerships renegotiated under stress; take-or-pay provisions tested |
| Infrastructure financing | Lenders look through to reconciled demand; risk premia fall; capex paced to collected revenue | Lenders rely on backlog and RPO headlines; private credit continues to fund against unreconciled signals | Refinancing difficulties in 2027–2029 cohort; stranded capacity; private-credit marks fall |
| Macroeconomic outcome | Capex continues at a sustainable pace; productivity gains arrive on the J-curve Brynjolfsson describes | Growth remains concentrated in data-center investment; vulnerability to sentiment persists | Dot-com-style correction of the kind the IMF has described as possible but not systemic[29] |
| Who benefits | Companies with durable, collected, well-margined revenue | Companies with the most favorable presentation | Companies with the strongest balance sheets and least dependence on external funding |
The transparency-led future is the one this paper advocates, and it is also the one in which the companies with the strongest underlying economics have the most to gain, because comparability rewards substance. The fragmented future is the default if nothing changes, and it is the one in which the October 2026 episode is repeated at increasing scale as more companies list and more capital is committed against incomparable figures. The market-correction future is the one in which the definitional correction is forced rather than chosen, arriving as scenarios B through E of the preceding section rather than as scenario A, and in which the irreversibility of the lower layers converts a measurement problem into stranded capital. The choice among them is, to a greater degree than is commonly recognized, within the control of the companies whose revenue is being measured.
Section 6: What Have We Learned? Eight Pillars
The six sections of this paper have moved from a forty-eight-hour episode in October 2026 to a framework for the financial governance of an industry through 2030. It is useful, before concluding, to distill what has been learned into a set of pillars that can stand independently of the particular figures that prompted them, because the figures will change within months while the structural problems they revealed will persist for years. The original outline of this paper proposed five pillars. The research conducted in its preparation, and in particular the evidence from the second-quarter earnings season, the Anthropic prospectus, and the warnings of the central banks and international institutions, suggested three more, and all eight are set out below.
Pillar 1 — Revenue Comparability Is Becoming as Important as Revenue Growth
The October 2026 reporting on OpenAI and Anthropic demonstrated that headline revenue figures can create profoundly misleading comparisons when companies use different reporting conventions, different distribution structures, and different annualization practices, and that the market will react to a definitional revision as though it were a change in economic substance. For three years the frontier-model industry competed on growth rate alone, and the metric it used to communicate that growth was never defined in a way that would permit cross-company comparison. The $20 billion gap between a $70 billion estimate and a $50 billion reported figure contained no information about demand for OpenAI’s products, and yet it moved the shares of three infrastructure companies in a single session.[1][2][3] Investors must understand what a revenue figure measures, over what period, under which contractual role, and through which channels, before using it to determine a company’s growth, market share, or valuation. The Revenue Denominator framework’s insistence on consistent definitions, transparent reconciliation, and careful separation of reported results from forward-looking targets is not a refinement of good analysis but a precondition for it.
Pillar 2 — Cloud Partnerships Are Reshaping the Economics of Frontier AI
Cloud infrastructure providers are not passive vendors supplying servers and electricity to the Models layer. In 2026 they are simultaneously suppliers of compute, distributors through whose marketplaces nearly half of one laboratory’s revenue flows, collectors of the majority of its receivables, equity investors in rounds exceeding $100 billion, customers purchasing model access, and competitors offering their own models on the same shelf.[6][7][47] Their contractual roles determine whether the model developer is the principal or the agent in each transaction, and therefore whether revenue is presented gross or net; their distribution fees, estimated at roughly 16 cents per dollar, determine how much of each customer dollar the model developer retains; and their take-or-pay provisions determine whether the model developer’s compute cost is a variable expense or a fixed obligation.[6][8] Understanding how model developers and their cloud partners divide responsibilities and value is therefore essential to evaluating the long-term profitability of the AI economy, and the disclosure of those arrangements by counterparty is the single most valuable addition that public filings could make.
Pillar 3 — Frontier-Model Valuations Must Be Grounded in Financial Substance
Rapidly increasing annualized revenue does not establish sustainable profitability or justify a particular enterprise valuation. Anthropic’s 2025 operating loss of more than $8 billion on revenue of $4.6 billion, OpenAI’s first-quarter 2026 cash burn of $3.7 billion on revenue of $5.7 billion, and both companies’ projected breakeven dates of 2028 and 2030 respectively make clear that the revenue multiple is being applied to businesses whose unit economics are not yet those of software.[8][46][47] Investors must examine realized and recognized revenue, gross margin after full allocation of inference cost and distribution fees, customer retention and concentration, operating expenses including training and stock-based compensation, cash flow after take-or-pay commitments, and the durability of demand against converging model capability. Transparent and consistent reporting across these dimensions is what will distinguish sustainable commercial expansion from growth metrics that exaggerate financial strength, and the companies whose substance is strongest have the most to gain from providing it.
Pillar 4 — Revenue Expectations Influence Every Layer of the AI Infrastructure Economy
The Five-Layer AI Economy demonstrates that financial performance in the Models and Applications layers drives investment decisions throughout the stack beneath them. Expected future revenue at the top justifies Nvidia’s supply-constrained guidance of roughly 70 percent growth for fiscal 2028, the hyperscalers’ combined 2026 capital expenditure of $725 billion to $760 billion, Oracle’s $664 billion of remaining performance obligations, CoreWeave’s $35 billion of debt against $104 billion of backlog, and the gigawatts of power that utilities are contracting to supply.[16][17][20][22][24][25] The signal descends through layers of declining reversibility, so that an expectation that can be revised in a quarter at the top may already have become a thirty-year asset at the bottom. Uncertainty over revenue measurement therefore propagates across the entire AI industrial system, affecting investors, lenders, and infrastructure providers far removed from the original transaction, and the quality of the revenue signal matters as much as its growth rate.
Pillar 5 — The AI Economy Needs Greater Disclosure, Not Artificially Identical Accounting
The objective of improved comparability should not be to force every artificial intelligence company to recognize revenue identically regardless of its contractual obligations. ASC 606 and IFRS 15, as clarified by ASU 2016-08, appropriately distinguish arrangements in which an entity controls the specified service from those in which it merely arranges for another party to provide it, and the two leading laboratories’ different conclusions appear to reflect genuinely different contracts.[12][13] The priority should instead be transparent disclosure of the principal-versus-agent conclusion for each channel, consistently defined performance metrics with stated formulas and reconciliations to recognized revenue, clear explanation of partner economics by counterparty, independently audited financial statements, and the gross-to-net bridges that allow investors to construct a common denominator for themselves. By 2030, the companies best able to demonstrate reliable earnings and sustainable cash generation on a comparable basis may possess an important competitive advantage in attracting capital at a lower cost than their less transparent peers.
Pillar 6 — The Annualized Run Rate Has Acquired Systemic Importance Without Acquiring a Definition
The run rate became the single most important number in the AI economy not because it was the best available measure but because it was the number every participant used, and a number everyone uses acquires systemic importance regardless of its analytical merits. It is an operating metric defined by each company for itself, it is not audited, it annualizes a single month and therefore magnifies every lumpy contract and every price change twelve-fold, and for a business growing at 70 percent a year it will always sit far above the recognized revenue of the period that produced it.[50][55] The SEC’s 2020 guidance on key performance indicators already requires public companies to define any such metric, explain its calculation, and disclose changes in method, and the application of that guidance to the laboratories once they list will be the first real constraint the run rate has faced.[14][15] Until then, every investor, lender, and infrastructure planner who uses a run rate should do so with the explicit recognition that it is a company-defined extrapolation and not a measurement of anything that has yet occurred.
Pillar 7 — Circular Financing Converts a Measurement Problem into a Stability Problem
The most consequential feature of the 2026 AI financing structure is that the same small group of counterparties appears on every side of the laboratories’ transactions. Amazon, Nvidia, Microsoft, Google, and SoftBank invest equity in the laboratories; the laboratories commit that capital to purchasing compute from the investors; the investors book the commitments as backlog and raise capital expenditure; and the laboratories’ revenue, which is what will eventually repay all of it, flows in significant part back through the investors’ own marketplaces, where it is presented gross by one party and net by another.[6][9][47] The Bank of England’s Financial Policy Committee has described these arrangements as opaque and at times circular and has warned that they could amplify losses; its Governor has written to G20 finance ministers about the financial ties between AI developers and the hyperscalers building their infrastructure; the IMF has warned that disappointment could have far-reaching fallout; and Morgan Stanley has estimated that $700 billion of datacenter capital expenditure through 2028 will be funded by private credit.[30][41][42] A revenue denominator that cannot be reconciled across the parties to a circular structure is not merely an inconvenience to equity analysts but a gap in the risk analysis of a system that central banks now regard as a potential channel of financial instability.
Pillar 8 — Adoption Is Not Revenue, Revenue Is Not Cash, and Cash Is Not Return
The final pillar is the one that most directly connects the measurement problem to the question of whether the infrastructure build-out will pay. The adoption of generative AI has been faster than that of any prior general-purpose technology, with 53 percent population adoption within three years and 70 percent of organizations using it in at least one function, and global corporate AI investment doubled to $581.7 billion in 2025.[35] But adoption does not become revenue until a customer pays, revenue does not become cash until a cloud partner remits it, and cash does not become return until it exceeds the cost of the compute that produced it and the capital that built the compute. MIT’s finding that 95 percent of enterprise generative AI initiatives have produced no measurable return, Bain’s estimate of an $800 billion gap between the revenue the industry will need by 2030 and the revenue it is on track to generate, Acemoglu’s estimate of modest aggregate productivity gains, and Furman’s observation that data-center investment accounted for 92 percent of U.S. growth in the first half of 2025 all describe the same gap from different angles.[32][36][38][39] Brynjolfsson’s J-curve offers the reason for patience: the returns from general-purpose technologies arrive after the investment and after a period in which measured productivity falls.[34] Whether the AI economy is on the J-curve or in a bubble is a question that cannot be answered until the revenue denominator is defined, because the shape of the curve is drawn in units of revenue, and in October 2026 those units were shown not to be standard.
Conclusion: Why Revenue Denominator May Become a Defining Financial Framework of the Five-Layer AI Economy
The October 8–9, 2026, reporting concerning OpenAI and Anthropic revealed a fundamental challenge emerging beneath the extraordinary growth of the artificial intelligence industry. Although frontier-model developers are attracting capital at valuations approaching and exceeding a trillion dollars, expanding enterprise adoption at rates faster than any previous general-purpose technology, and driving investment throughout the semiconductor, cloud, datacenter, and energy sectors on a scale that the International Monetary Fund expects to exceed the railroad, electrical, and telecommunications build-outs as a share of output, the financial measurements used to evaluate their commercial progress are not directly comparable, and the market had been treating them as though they were.[31] The apparent difference between OpenAI’s approximately $50 billion September annualized revenue and the roughly $70 billion comparative figure that had circulated two weeks earlier demonstrated how easily different measurement conventions can create uncertainty, move the valuations of companies several layers removed from the original transaction, and complicate the assessment of the industry’s actual economic performance.[1][2][3]
The significance of the problem extends well beyond traditional accounting. Modern artificial intelligence is delivered through commercial networks in which the model laboratory, the hyperscale cloud provider, the neocloud, the enterprise software platform, the API distributor, and the autonomous agent each contribute a different combination of intellectual property, infrastructure, customer relationship, billing capability, and financial risk, and in which the same counterparty frequently occupies several of those roles at once. Existing revenue-recognition standards distinguish these arrangements according to their economic substance, asking who controls the specified service before it reaches the customer, and the different conclusions that OpenAI and Anthropic have reached appear to be legitimate applications of that principle to genuinely different contracts.[5][12][13] But headline financial comparisons conceal those differences, and a run-rate metric that no standard-setter defines has been allowed to carry the weight of trillion-dollar valuations and hundreds of billions of dollars of capital expenditure guidance. A meaningful evaluation of the industry’s future therefore requires a clearer understanding of how revenue is earned, reported, distributed, retained, and converted into cash, which is what the five dimensions of the Revenue Denominator framework are intended to provide.
Within the Five-Layer AI Economy, the consequences of this uncertainty are particularly significant because of the asymmetry between the speed at which revenue expectations can be revised and the irreversibility of the commitments made against them. Commercial revenue originates primarily in the Models and Applications layers, yet the expectations it generates drive capital allocation across Datacenters, Chips, and Energy, where construction periods run to years and asset lives to decades. When frontier-model companies forecast rapid commercial expansion, Nvidia guides to supply-constrained growth, Oracle books hundreds of billions of dollars of remaining performance obligations against negative free cash flow, CoreWeave issues tens of billions of dollars of debt against a backlog forty times its quarterly revenue, the hyperscalers raise capital expenditure guidance toward three-quarters of a trillion dollars, and utilities plan transmission against gigawatt-scale load forecasts.[16][20][22][24] If the revenue assumptions supporting those commitments are misunderstood, and the October episode demonstrated that they can be misunderstood even by sophisticated investors, the error is embedded in physical capital before it is detected, and the correction, when it comes, arrives through credit spreads and impairments rather than through a quiet revision of a spreadsheet.
This is precisely why I chose the title “Revenue Denominator.” The term captures a problem more fundamental than the absolute amount of money any AI company reports earning. It asks what financial activity is actually being measured, over what period, under which contractual role, through which channels, with how much retained after partner payments, and with what conversion into cash, and it insists that those questions be answered before any revenue figure is placed in the denominator of a multiple or the numerator of a growth rate. It also creates an analytical connection between the commercial performance of frontier models and the capital expenditure expectations supporting the physical infrastructure of artificial intelligence, a connection that the market has been acting upon for three years without naming. The title is therefore sufficiently focused to examine gross-versus-net accounting under ASC 606 and sufficiently broad to address valuation, cloud partnerships, infrastructure financing, financial stability, investor accountability, and the evolution of AI economic governance through 2030.
The remedy this paper proposes is deliberately modest in its demands on accounting and ambitious in its demands on disclosure. It does not ask Anthropic to present net or OpenAI to present gross; it does not ask the FASB or the IASB to amend a standard that is functioning as designed; it does not ask the SEC to regulate a metric that its 2020 guidance already requires public companies to define. It asks the companies at the top of the AI economy to publish, voluntarily and quarterly, a reconciliation that would allow every investor, lender, and infrastructure planner beneath them to place their revenue on a common denominator: recognized revenue by channel, a gross-to-net bridge, a defined run rate reconciled to recognized revenue, partner payments by counterparty, concentration, commitments and their conditionality, cash conversion, and forward targets labeled as such. The software industry converged on comparable disclosures without regulation because the companies that provided them were rewarded with a lower cost of capital, and the frontier-model industry, approaching its first public listings at valuations an order of magnitude larger, has a correspondingly larger incentive to do the same.
By the end of this decade, the defining financial question for artificial intelligence may no longer be which company announces the largest annualized revenue figure, secures the most capital, or commits to the greatest quantity of computing infrastructure. The more consequential question may be which companies can demonstrate durable customer demand, transparent revenue recognition, sustainable operating margins, and sufficient cash generation to support the commitments they have made and the commitments others have made on their behalf. In that environment, financial transparency will become not merely a matter of accounting compliance but a strategic foundation for long-term competitiveness, and the companies that define their denominators first will set the terms on which the rest of the industry is measured.
Ultimately, Revenue Denominator argues that the future of artificial intelligence cannot be valued accurately until the industry becomes more transparent about how it measures the economic activity on which its expansion depends. The Five-Layer AI Economy may be built upon energy, chips, datacenters, models, and applications, and the adoption of what those layers produce may be the fastest in the history of technology, but its long-term financial sustainability will depend on whether the revenues supporting those five layers are measurable, comparable, and economically real. In October 2026 the market learned that it had not yet established the first two. The purpose of this paper has been to show why the third cannot be known until it does.
Footnotes and Endnotes:
[1] Financial Times, via Investing.com. “OpenAI annualized revenue at $50bn, far below reports – FT.” October 8, 2026. https://www.investing.com/news/stock-market-news/openai-annualized-revenue-at-50bn-far-below-reports–ft-4939530
[2] CNBC. “Nvidia, Oracle, CoreWeave and other AI stocks sink on OpenAI revenue report.” October 8, 2026. https://www.cnbc.com/2026/10/08/open-ai-revenue-nvidia-oracle-coreweave.html
[3] Bloomberg, via Quartz. “OpenAI is still projecting $70 billion in annualized revenue by end of 2026.” October 9, 2026. https://qz.com/openai-annualized-revenue-70-billion-projection-2026-100926
[4] Reuters, via The Manila Times. “OpenAI’s Sept annualized revenue nears $50 billion.” October 10, 2026. https://www.manilatimes.net/2026/10/10/business/foreign-business/openais-sept-annualized-revenue-nears-50-billion/2442529
[5] Madison Mills (with Lucinda Shen), Axios. “Anthropic and OpenAI’s revenue chasm, explained.” September 3, 2026. Quoting Francine McKenna, accounting professor and author of The Dig, and Lucas Ward, CEO of Oath. https://www.axios.com/2026/09/03/anthropic-and-openais-revenue-chasm-explained
[6] Echo Wang and Sabrina Valle, Reuters, via Business Standard. “Anthropic IPO prospectus lays bare deep dependence on Big Tech partners.” September 30, 2026. https://www.business-standard.com/world-news/anthropic-ipo-prospectus-lays-bare-deep-dependence-on-big-tech-partners-126093000297_1.html
[7] Silicon UK (citing Reuters). “Anthropic prospectus details reliance on Amazon, Google cloud sales.” September 30, 2026. https://www.silicon.co.uk/e-enterprise/fund-raising/anthropic-ai-prospectus-631740
[8] The Next Web (citing Reuters). “Anthropic lost $42bn in 2025 as revenue grew 12-fold, Reuters reports.” September 2026. https://thenextweb.com/news/anthropic-ipo-prospectus-42bn-loss-518bn-compute
[9] Implicator.ai (citing The Verge and CNBC). “OpenAI CRO tells staff Anthropic inflates run rate by $8 billion.” April 13, 2026. Quoting Denise Dresser, Chief Revenue Officer, OpenAI. https://www.implicator.ai/openai-cro-tells-staff-anthropic-inflates-run-rate-by-8-billion/
[10] Channel Insider. “OpenAI’s Revenue Run Rate Is $20B Below Earlier Estimates. Here’s Why.” October 9, 2026. https://www.channelinsider.com/ai/news-openai-revenue-run-rate-cloud-partner-accounting/
[11] Data Studios. “OpenAI’s annualized revenue reportedly reaches $50 billion as accounting differences complicate comparison with Anthropic.” October 2026. https://www.datastudios.org/post/openai-annualized-revenue-50-billion-gross-net-anthropic
[12] Financial Accounting Standards Board, ASU No. 2016-08, “Principal Versus Agent Considerations (Reporting Revenue Gross Versus Net),” March 17, 2016, summarized by Deloitte IAS Plus. https://www.iasplus.com/en/news/2016/03/asu-2016-08
[13] Deloitte, Accounting Spotlight. “Revenue Recognition — Evaluating Whether an Entity Is Acting as a Principal or as an Agent.” 2019. https://www.iasplus.com/en/publications/us/accounting-spotlight/2019/revenue-principal-or-agent
[14] U.S. Securities and Exchange Commission. Release No. 33-10751, “Commission Guidance on Management’s Discussion and Analysis of Financial Condition and Results of Operations” (key performance indicators and metrics). January 30, 2020. https://www.sec.gov/rules/interp/2020/33-10751.pdf
[15] Proskauer Rose LLP. “SEC Issues Guidance on the Use of Key Performance Indicators and Metrics in MD&A.” February 2020. https://www.proskauer.com/alert/sec-issues-guidance-on-the-use-of-key-performance-indicators-and-metrics-in-md-and-a
[16] NVIDIA Corporation. “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027.” Form 8-K, Exhibit 99.1, August 26, 2026. Quoting Jensen Huang, founder and CEO. https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm
[17] Calvin Curdie, Rivkin. “NVIDIA Corporation (NVDA): July quarter revenue doubles, October guide lands at US$108bn.” August 31, 2026. https://rivkin.com.au/?p=37534
[18] Microsoft Corporation. “Microsoft Cloud and AI Strength Fuels Fourth Quarter Results.” Fiscal Year 2026 Q4 press release, July 29, 2026. https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast
[19] Free Press Journal. “Microsoft’s Azure Cloud Business Crosses $100 Billion In Annual Revenue, Up 41% As AI Demand Accelerates.” July 30, 2026. https://www.freepressjournal.in/amp/tech/microsofts-azure-cloud-business-crosses-100-billion-in-annual-revenue-up-41-as-ai-demand-accelerates
[20] Oracle Corporation. “Oracle Announces Q1 Results Driven by Triple Digit Growth in Cloud Infrastructure Revenues.” September 10, 2026. https://www.oracle.com/news/announcement/q1fy27-earnings-release-2026-09-10/
[21] Oracle Corporation. Fiscal Year 2026 Fourth Quarter and Full Year Results, Form 8-K, Exhibit 99.1. June 2026. https://www.sec.gov/Archives/edgar/data/0001341439/000119312526265848/orcl-ex99_1.htm
[22] CoreWeave, Inc. “CoreWeave Reports Strong Second Quarter 2026 Results.” August 11, 2026. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/
[23] Mike Thrift, Beancount. “CoreWeave FY2026 Q2 Earnings: $626M Loss, $104B Backlog.” September 22, 2026. https://beancount.io/blog/2026/09/22/coreweave-fy2026-q2-earnings-analysis
[24] Tom’s Hardware (citing the Financial Times). “Google, Microsoft, Meta, and Amazon capex spending to hit $725 billion in 2026, up 77% from last year.” 2026. Quoting Brent Thill, Jefferies. https://tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion
[25] Statista. “Big Tech’s AI Capital Expenditure Expected to Hit $760 Billion in 2026.” July/August 2026. https://www.statista.com/chart/35046/
[26] The Wealth Advisor (via Yahoo Finance). “‘Magnificent 7’ Earnings Rush Reveals AI Spending Surge, With Hyperscaler Capex Set To Reach $725 Billion In 2026.” July 29, 2026. https://www.thewealthadvisor.com/article/magnificent-7-earnings-rush-reveals-ai-spending-surge-hyperscaler-capex-set-reach-725
[27] Crypto Briefing. “Microsoft, Meta, Amazon face investor anxiety over AI spending as Big Tech capex tops $700B.” July 2026. https://cryptobriefing.com/big-tech-ai-spending-investor-anxiety/
[28] Associated Press, via The Korea Times. “Is there an AI bubble? Financial institutions sound a warning.” October 9, 2025. Quoting Kristalina Georgieva, Managing Director, IMF. https://www.koreatimes.co.kr/world/20251009/is-there-an-ai-bubble-financial-institutions-sound-a-warning
[29] Reuters, via Business Recorder. “AI investment boom may lead to bust, but not likely systemic crisis, IMF chief economist says.” October 2025. Pierre-Olivier Gourinchas, Chief Economist, IMF. https://www.brecorder.com/news/amp/40387359
[30] Asia News Network. “IMF warns of AI boom, energy shock and debt risks to global economy.” October 8, 2026. Kristalina Georgieva, IMF. https://asianews.network/?p=306549
[31] EM360Tech. “IMF chief warns AI boom puts global growth at risk.” October 2026. https://em360tech.com/tech-articles/imf-chief-warns-ai-boom-puts-global-growth-risk
[32] Fortune. “Without data centers, GDP growth was 0.1% in the first half of 2025, Harvard economist says.” October 7, 2025. Jason Furman, Harvard University; quoting Rusty Foster, Today in Tabs. https://fortune.com/2025/10/07/data-centers-gdp-growth-zero-first-half-2025-jason-furman-harvard-economist
[33] TechSpot. “Tech investments lift US GDP amid broader economic sluggishness.” 2025. Quoting Jason Furman (@jasonfurman) on X, September 27, 2025. https://www.techspot.com/news/109794-tech-investments-lift-us-gdp-amid-broader-economic.html
[34] Stanford Institute for Human-Centered Artificial Intelligence. News digest citing Erik Brynjolfsson, Director, Stanford Digital Economy Lab, in The New York Times, May 29, 2026. https://hai.stanford.edu/news/infographic-age-artificial-intelligence
[35] Stanford Institute for Human-Centered Artificial Intelligence. “The 2026 AI Index Report — Economy.” April 13, 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report/economy
[36] Daron Acemoglu. “The Simple Macroeconomics of AI.” NBER Working Paper No. 32487, May 2024. https://www.nber.org/papers/w32487
[37] Investment News NZ. “Call me Aswath: ESG, AI, valuation and other whales in the room.” Quoting Aswath Damodaran, Professor of Finance, NYU Stern. https://investmentnews.co.nz/investment-news/call-me-aswath-esg-ai-valuation-and-other-whales-in-the-room/
[38] David Ramel, Virtualization Review. “MIT Report Finds Most AI Business Investments Fail, Reveals ‘GenAI Divide’.” August 19, 2025. Citing MIT Project NANDA, “The GenAI Divide: State of AI in Business 2025.” https://virtualizationreview.com/Articles/2025/08/19/MIT-Report-Finds-Most-AI-Business-Investments-Fail-Reveals-GenAI-Divide.aspx
[39] PYMNTS. “AI Companies Face $800 Billion Funding Shortfall, Says Bain Report.” September 23, 2025. Quoting David Crawford, Bain & Company. https://www.pymnts.com/news/artificial-intelligence/2025/ai-companies-face-800-billion-dollar-funding-shortfall-bain
[40] Bloomberg. “An $800 Billion Revenue Shortfall Threatens AI Future, Bain Says.” September 23, 2025. https://www.bloomberg.com/news/articles/2025-09-23/an-800-billion-revenue-shortfall-threatens-ai-future-bain-says
[41] FStech. “Bank of England warns AI poses threat to interconnected financial stability.” October 2026. Citing the Financial Policy Committee record of September 25, 2026. https://fstech.co.uk/fst/Bank_of_England_warns_AI_poses_threat_to_interconnected_financial_stability.php
[42] Capacity. “Bank of England boss: AI bubble collapse would be a ‘disaster’.” 2026. Andrew Bailey, Governor, Bank of England; Financial Policy Committee minutes. https://capacityglobal.com/news/bank-of-england-boss-ai-bubble-collapse-would-be-a-disaster/
[43] Reuters, via MarketScreener. “Five debt hotspots in the AI data centre boom.” 2025. https://in.marketscreener.com/news/five-debt-hotspots-in-the-ai-data-centre-boom-ce7d50dbda80f120
[44] PYMNTS (citing Reuters). “OpenAI Mulling $1 Trillion IPO as AI Costs Compound.” October 2025. Quoting Sam Altman, CEO, OpenAI, on X. https://www.pymnts.com/artificial-intelligence-2/2025/openai-mulling-1-trillion-ipo-as-ai-costs-compound/
[45] Reuters, via Business Standard. “OpenAI expects $600 bn in compute spending through 2030 ahead of IPO.” February 21, 2026. https://www.business-standard.com/world-news/openai-expects-600-bn-in-compute-spending-through-2030-ahead-of-ipo-126022100286_1.html
[46] AI Weekly (citing The New York Times and The Information). “OpenAI Leans Toward 2027 IPO as Altman Holds $1 Trillion Line.” 2026. https://aiweekly.co/alerts/openai-leans-toward-2027-ipo-as-altman-holds-1-trillion-line
[47] AI Weekly (citing BNN Bloomberg). “OpenAI Disputes Anthropic Revenue by Billions.” June 2026. https://aiweekly.co/alerts/openai-disputes-anthropic-revenue-by-billions
[48] David Cahn, Sequoia Capital. “AI’s $600B Question.” June 20, 2024. https://www.sequoiacap.com/article/ais-600b-question/
[49] Startup Fortune (citing Goldman Sachs Research). “Meta and Microsoft walk into earnings week with $145 billion question marks hanging over them.” July 25, 2026. https://startupfortune.com/meta-and-microsoft-walk-into-earnings-week-with-145-billion-question-marks-hanging-over-them/
[50] Guidewire Software, Inc. Form 10-K for fiscal year ended July 31, 2020, “Key Business Metrics — Annual Recurring Revenue.” https://www.sec.gov/Archives/edgar/data/1528396/000152839620000053/gwre-20200731.htm
[51] Matt Mckenzie, The Nightly. “AI bubble or boom on the radar for economists as International Monetary Fund keeps an eye on tech.” March 2026. Quoting Tobias Adrian and Pierre-Olivier Gourinchas, IMF World Economic Outlook. https://thenightly.com.au/business/ai-bubble-or-boom-on-the-radar-for-economists-as-international-monetary-fund-keeps-an-eye-on-tech-c-21354539
[52] Webull. “Nvidia Q2 FY2027 Earnings Beats Revenue, EPS Estimates; Guides Q3 to $108B.” August 2026. Quoting Colette Kress, CFO, NVIDIA, Q2 FY2027 earnings call. https://www.webull.com/blog/304-Nvidia-Q2-FY2027-Earnings-Beats-Revenue-EPS-Estimates-Guides-Q3-to-108B
[53] InvestingLive (citing Reuters). “Anthropic IPO prospectus shows $42 billion 2025 loss and $518 billion spending plans.” September 2026. https://investinglive.com/stock-market-update/anthropic-ipo-prospectus-shows-42-billion-2025-loss-and-518-billion-spending-plans/
[54] Fortune Commentary. “When the music stops: the unravelling of AI companies’ flawed valuations.” February 6, 2026. https://dc.fortune.com/2026/02/06/when-the-music-stops-the-unravelling-of-ai-companies-flawed-valuations/
[55] KuCoin. “OpenAI Revenue in 2026: Why the $50 Billion Figure Is Below Earlier Estimates.” October 10, 2026. https://www.kucoin.com/blog/openai-revenue-2026-50-billion-estimate
[56] Deloitte, Roadmap: Revenue Recognition, Appendix C.9, “Principal-Versus-Agent Considerations” (TRG Agenda Papers 1 and 5; ASC 606-10-55-37 as amended by ASU 2016-08). https://dart.deloitte.com/USDART/home/codification/revenue/asc606-10/roadmap-revenue-recognition/appendix-c-summary-issues-addressed-in/c-9-principal-versus-agent-considerations
