NextFin News - The artificial-intelligence boom is being priced on a measurement system that does not exist. The largest hyperscalers are committing more than $1 trillion to AI infrastructure across 2025 and 2026, but the revenue figures used to justify that spending are voluntary, self-defined, and in some cases circular — a disclosure gap that is turning the AI trade into a bet on accounting as much as on technology.
Merriam-Webster named "slop" its word of the year for 2025, defining it as digital content of low quality produced in quantity by artificial intelligence. The sharper use of the term now circulating through markets is different. The slop is not only the AI-generated content flooding the internet. It is the revenue reporting itself.
The most bullish number in circulation comes from Exponential View, a research firm, whose mid-2026 "State of the AI Economy" report estimated that the global ex-China generative-AI economy generated $110 billion in trailing 12-month revenue, with a current run-rate of $175 billion. The report described the figure as deduplicated revenue across apps, foundation models, and hosting — excluding chips, AI advertising uplift, legacy software AI features, and financing. It is also the most contested. Ed Zitron, an independent analyst of the AI industry, argued the total was assembled "by smashing together all AI revenues, including both OpenAI and Anthropic's customer spend and compute spend," and that the firm declined to explain how it deduplicated. He estimated that OpenAI and Anthropic alone accounted for 68% of the total.
Whether that critique lands fully or not, it points to the underlying problem: there is no agreed definition of what counts as AI revenue, no standard for how a run-rate is calculated, and no requirement that companies disclose the figure at all. The result is a market in which the denominator of a trillion-dollar investment thesis is a metric that management can change, or withdraw, at its discretion.
The growth is real. The measurement is not.
The revenue growth is not fictional. Microsoft reported in its fiscal third quarter of 2026 that its AI business surpassed an annual revenue run-rate of $37 billion, up 123% year over year — the first update to the metric since it disclosed a $13 billion run-rate in January 2025. Alphabet's Google Cloud said enterprise AI solutions revenue grew 800%, and chief executive Sundar Pichai told investors:
Our enterprise AI solutions have become our primary growth driver for cloud.
The unit's annual run-rate crossed $80 billion on quarterly revenue of $20.02 billion, up 63% year over year.
But these figures are run-rates, not GAAP revenue. A run-rate extrapolates a recent period — a month, or in some cases the last four weeks — across a full year. A technology-industry publication reported in March 2026 that OpenAI had topped $25 billion in annualized revenue as of the end of February, a 17% increase from $21.4 billion at the end of 2025, citing a person familiar with the figure. The same reporting noted that the company calculates annualized revenue by multiplying its most recent four weeks of revenue by 12 — a methodology in which a single strong week can move the annual figure by billions.
Anthropic, for its part, was reported to have reached roughly $65 billion in annualized revenue by May 2026, a number whose construction has not been publicly defined. The problem is not that these numbers are fabrications. It is that they are private definitions of a public question. Run-rates are voluntary non-GAAP metrics that companies can introduce, modify, or stop reporting altogether at any time. Microsoft disclosed its AI run-rate in the third quarter of fiscal 2026 but did not update the figure when it reported full-year results three months later. When investors build models on a metric that can be changed at management's discretion, the models inherit that discretion.
There is also the question of what sits inside the number. Microsoft's fiscal 2026 annual report, filed with the Securities and Exchange Commission, disclosed $24.1 billion of revenue from commercial arrangements with OpenAI, inclusive of revenue-sharing payments, plus $6.0 billion in accounts receivable from OpenAI as of June 30, 2026. One analysis of the disclosure estimated that between half and two-thirds of Microsoft's reported AI business derived from OpenAI. That concentration is the circularity critics point to: compute spend by a loss-making model lab shows up as revenue for the cloud provider that hosts it, and both sides can point to the same dollar as evidence of demand. Microsoft's total funding commitments to OpenAI reached $13.0 billion, of which $11.9 billion had been funded by the end of fiscal 2026 — meaning a portion of the revenue Microsoft books from OpenAI is funded, indirectly, by Microsoft itself.
The capex is audited. The return is not.
Against these fluid revenue figures sits spending that is concrete, depreciated, and unforgiving. The Bank for International Settlements said in its 2026 annual economic report that the five largest hyperscalers are set to spend more than $1 trillion on AI-related capital expenditure from 2025 through 2026 — commitments that are "outpacing earnings and the free cash flow of these firms," leading some to issue debt to raise additional financing. The BIS warned:
Disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust.
The report noted that the scale and pace of the current AI investment boom bear resemblance to canal mania, railway mania, the Roaring Twenties, and the dotcom boom — episodes that "ended with an eventual reversal in investment, inducing economy-wide recessions."
Combined 2026 capital expenditure across the four largest hyperscalers is projected at roughly $730 billion, compiled from company earnings guidance: Amazon guiding about $200 billion, Alphabet roughly $195 billion to $205 billion, Microsoft roughly $190 billion, and Meta roughly $130 billion to $145 billion. That spending is already compressing cash flow. Amazon's trailing 12-month free cash flow collapsed to approximately $1.2 billion, down about 95% from roughly $26 billion a year earlier, as capital expenditure consumed essentially all of the cash the business generates. Oracle's free cash flow turned negative $23.7 billion for fiscal 2026, with operating lease liabilities of $26.6 billion on its balance sheet and roughly $260 billion in additional data-centre lease commitments signed but not yet reflected as liabilities.
Here is the asymmetry that defines the trade. The spending is GAAP — depreciated, audited, and due on schedule. The revenue used to justify it is a self-defined run-rate. If the two converge on the company's terms, the boom continues. If they converge on an auditor's or a regulator's terms, the denominator shrinks.
The accounting gap is structural, not temporary
The root of the problem is that AI revenue does not map cleanly onto existing revenue-recognition rules. Under ASC 606, the accounting standard for revenue from contracts with customers, companies must identify distinct performance obligations, determine standalone selling prices, and time recognition as obligations are satisfied. AI arrangements — bundled model access, usage-based pricing, training credits, revenue-sharing partnerships — strain each of those steps. A usage-based API call may be recognized as it is consumed; a multi-year committed capacity contract may be recognized over the commitment period; a revenue-sharing arrangement with an investee sits in a different part of the income statement entirely.
Because the economics are genuinely hard to capture, companies have reached for metrics that are easier to communicate than to audit. Run-rates, annualized revenue, "AI business" revenue that blends infrastructure, software, and services — none of these appears in a GAAP income statement, and none is subject to the same disclosure discipline. The SEC's Division of Examinations has made AI a focus area for fiscal 2026, saying it will review registrant representations regarding their AI capabilities for accuracy, and will assess whether operations and controls are consistent with disclosures made to investors. But the division has authority to review accuracy; it does not have a standardized definition to review it against. That gap is precisely what lets run-rates, annualized figures, and bundled cloud revenue coexist without contradiction.
Enforcement so far has targeted capability claims more than revenue math. The SEC and the Justice Department brought parallel actions in April 2025 against the founder of Nate Inc., alleging he misrepresented an app's AI automation when orders were in fact routed to overseas contract workers. The agency also settled with two investment advisers, Delphia and Global Predictions, over false and misleading statements about their purported use of AI. The logic, however, extends naturally: if a company cannot define what its AI revenue is, it cannot represent it accurately. Some companies are already moving toward more granular disclosure. Analysts note that when hyperscaler management shifts from describing AI revenue as a "run-rate" or "trajectory" to reporting discrete product-line figures — Bedrock, Copilot, Gemini Workspace add-on revenue — that shift signals the monetization curve is maturing. Only 11% of companies on first-quarter 2026 earnings calls quantified AI productivity gains, and just 2% quantified AI's impact on actual earnings, according to one analysis of earnings transcripts.
The counter-thesis: the backlog proves the demand is real
The bull case is not weak, and it deserves to be stated at full strength. Exponential View's report is explicit that its $110 billion figure excludes chips, advertising uplift, legacy software AI features, and financing — the very categories that would invite circularity charges. Luke Lango, the report's author, described it as "real customer demand, not circular GPU spend," pointing to growth running roughly 35% quarter over quarter, or 3.2 times annualized — faster, the report argues, than the internet, mobile, or cloud waves that preceded it.
The bulls also have the backlog, which is the strongest evidence on their side because it is contractual rather than extrapolated. Alphabet reported a Google Cloud remaining performance obligation of $462 billion as of March 31, 2026, of a company-wide total of $467 billion; by the second quarter that cloud backlog had risen to $514 billion. Google said it expects to recognize just over half of its revenue backlog over the next 24 months. Microsoft reported commercial remaining performance obligations of $678 billion at the end of fiscal 2026, up 84% year over year, from a base of $627 billion just three months earlier. Committed future revenue is real revenue — it is not a run-rate extrapolation, and it is contractually owed.
And the growth rates are not accounting artifacts. Microsoft's AI business up 123% year over year, Google's enterprise AI revenue up 800%, Google Cloud operating income up 203% to $6.6 billion in the first quarter of 2026 — this is cash coming in the door. The bull case concedes the measurement problem and argues it does not decide the outcome: the customers are real, the contracts are signed, the conversion is visible in the backlog, and standardization will come when the market demands it. One analysis of Microsoft's backlog noted that the entire $51 billion sequential increase came from customers outside the frontier model companies, meaning the backlog grew even excluding OpenAI entirely — evidence, in the bull's view, that demand is broadening beyond the circular core.
The bear case, at its strongest, concedes the growth is real and argues it is priced for perfection on numbers that cannot be stress-tested. The question is not whether AI revenue is growing. It is whether the revenue that can be audited, recognized, and converted to cash grows fast enough to cover a capital base that may require close to $1 trillion in annual spend by 2027.
What to watch
The near-term signal is disclosure granularity. If hyperscalers begin reporting discrete, audited AI product-line revenue and those figures reconcile cleanly with their run-rate claims, the slop thesis weakens materially. If companies quietly stop reporting run-rates — as Microsoft effectively did at its full-year print — or if the gap between run-rate and GAAP-recognizable revenue widens, the opposite follows. A specific threshold: if reported AI revenue growth decelerates below 30% year over year while capital expenditure remains above $700 billion annually, the return math breaks down.
The medium-term signal is the conversion of backlog into recognized revenue. Alphabet's $514 billion cloud backlog and Microsoft's $678 billion commercial backlog are the bull case's strongest evidence; the pace at which they convert into revenue, relative to the pace of new capex and to the depreciation schedule on the assets being built, is the scorecard. Google expects just over half of its backlog to convert within 24 months — a claim that will be testable against actual cloud revenue over the next two years.
The long-term signal is regulatory. The SEC has the authority to demand accuracy but not yet a standard to measure it against. If the agency or the Public Company Accounting Oversight Board moves from examination priorities to prescriptive AI revenue disclosure rules — and companies are forced to restate or reclassify what counts as AI revenue — the measurement system would harden, and with it the denominator of the entire trade.
Base case: AI revenue keeps growing fast, disclosure stays voluntary and inconsistent, and the market tolerates the ambiguity as long as capex keeps producing growth. Upside case: standardized disclosure arrives, backlog converts ahead of schedule, and the run-rates are vindicated by audited numbers. Downside case: growth decelerates, companies quietly stop reporting run-rates, and a regulator or auditor forces a redefinition that shrinks the reported AI economy.
The AI boom will be decided by technology, but it will be priced by accounting. Right now the accounting is slop — and a market built on self-defined numbers is not a market that can be stress-tested, only believed.
Explore more exclusive insights at nextfin.ai.
