OpenAI’s $20 Billion Gap Tests Faith in the AI Build-Out

OpenAI’s revenue run rate, restated to about $50 billion, reflects accounting rather than lost sales. The episode shows how much of the listed AI complex is priced on one private company’s numbers.

Key Takeaways

• The $20 billion gap arose when investors grossed up OpenAI’s net revenue to match Anthropic’s method, then applied a growth rate measured on a different basis to that inflated figure.
• On its own terms OpenAI is still compounding quickly: 77% run-rate growth in the third quarter and 107% in enterprise imply a net base near $28 billion entering July.
• Oracle, CoreWeave and Microsoft sold off because OpenAI’s revenue underwrites their backlogs. Oracle alone reports $664 billion in remaining performance obligations.

A Number That Described No Company

For much of the past year, investors have built their view of OpenAI from scattered disclosures: a finance chief’s blog post, a funding announcement, a slide deck passed between backers. On Thursday one of those pieces was replaced. OpenAI recently told investors its annualized revenue was approaching $50 billion at the end of September, well short of the roughly $70 billion that Reuters and other outlets had reported late last month, according to the Financial Times. CNBC confirmed the lower figure, which it said came from an investor presentation, and contrasted it with a $68 billion number that circulated widely in September. Within hours, a correction to a private company’s metric had become a verdict on the public AI trade.

The missing $20 billion was never lost revenue. A person familiar with the documents told CNN the larger figure did not come from OpenAI but from investors trying to line its numbers up against Anthropic’s, which include gross revenue from cloud providers, whereas OpenAI reports net. The method compounded the problem. Investors first estimated OpenAI’s revenue at about $40 billion in August; when the company later reported growth of more than 70 percent, they applied that rate to the base and arrived at $70 billion. A growth rate measured on one definition was multiplied onto a level built on another. The result described no company that actually exists.

Growth That Survives the Restatement

Read on its own terms, the corrected figure is still formidable. Alongside the $50 billion, OpenAI cited 77 percent run-rate growth in the third quarter and 107 percent growth in its enterprise run rate. Working backward, that pace implies a net run rate of roughly $28 billion entering July, which fits the company’s earlier disclosures. In January, CFO Sarah Friar wrote that the annual run rate had climbed from $2 billion in 2023 to $20 billion in 2025. When its funding round closed in March, OpenAI said it was bringing in about $2 billion a month after recording $13.1 billion of revenue in 2025.

The trajectory, in short, holds. Enterprise revenue roughly doubling within a single quarter should steady nerves rather than rattle them, because corporate contracts tend to be stickier and more predictable than consumer subscriptions. What changed on Thursday was not the slope of OpenAI’s growth but the ruler used to measure it, and in this market the ruler matters considerably. Private AI valuations are argued largely by comparison, and a company seen as pulling ahead of its rival commands a different multiple from one seen as falling behind. The restatement did nothing to OpenAI’s revenue. It did a great deal to the story investors had been telling about it.

Two Rulers, One Comparison

That comparison now rests on uncertain ground. Anthropic’s last company-confirmed run rate was $47 billion, disclosed with its Series H on May 28 alongside a $965 billion valuation; Bloomberg later reported, without company confirmation, that the figure topped $65 billion by the end of July. Both companies are GAAP compliant but treat partner sales differently: under Anthropic’s approach, a $100 sale through a cloud provider is booked in full and the provider’s cut is expensed, while OpenAI records only its share. Place a gross number beside a net one and OpenAI appears to trail. Whether it does on a consistent basis cannot be determined from anything in the public record.

This is the more durable lesson. The two most scrutinized revenue figures in technology are not commensurable, and the attempt to make them so produced a distortion that stood for weeks. As Axios noted, annualized revenue is a Silicon Valley convention better suited to startups, and public-market investors prefer reported revenue. Clarity will arrive only through registration statements. OpenAI confidentially submitted draft IPO documents in June, and Anthropic did the same that month. Until audited figures appear, the sector’s two most valuable private businesses will be priced from numbers that require translation.

The Market Prices Its Exposure

The reaction was swift and selective. The Nasdaq Composite fell 1.25 percent to 27,193.34 and the S&P 500 slipped 0.47 percent to 7,765.36, the Nasdaq’s steepest one-day decline since mid-August. CoreWeave fell nearly 8 percent, Oracle nearly 6 percent, Nvidia 3 percent, and AMD and Broadcom 4 percent each. Rising oil prices and Treasury yields added to the pressure. The losses tracked exposure, not arithmetic. Oracle’s 10-Q shows $664 billion of remaining performance obligations at August 31, about 13 percent of which it expects to recognize within twelve months, along with $11.4 billion of customer prepayments carrying a significant financing component in the quarter.

In that same quarter Oracle spent $28.5 billion on capital projects and raised about $20 billion through an at-the-market stock sale. Analysts estimate that roughly half the backlog traces to OpenAI, a share Oracle does not disclose. CoreWeave has announced OpenAI contracts worth up to $22.4 billion. Microsoft accounts for an approximately 25 percent as-converted stake under the equity method, and OpenAI has contracted an incremental $250 billion of Azure services. For these companies, OpenAI’s revenue is the base on which counterparty risk rests. Baird’s Ross Mayfield warned that any crack in the revenue story could spread through the AI supply chain, while D.A. Davidson’s Gil Luria said he was not concerned. Both can be right. The revenue has not fallen; confidence in its measurement has.

Time, Capital and the Prospectus

OpenAI has the balance sheet to wait. It raised $122 billion at an $852 billion valuation, with Amazon committing $50 billion, $35 billion of it contingent on an IPO or a technical milestone, and Nvidia and SoftBank $30 billion apiece. Friar told CNBC last week that the company remains “very well capitalized.” Its spending, however, is heavy: OpenAI posted a net loss of $38.5 billion on $13.07 billion of revenue in 2025, according to reported figures the company has not confirmed. Its product cadence has been shaped by safety as much as by competition. The GPT-5.6 Sol model went public on July 9, but in August OpenAI said it could not rule out its unreleased Astra model reaching the Critical cyber threshold, and it later paused reinforcement-learning training for two weeks.

Sam Altman then told Fortune the company would not go public in 2026, citing safety. Each decision is defensible. Together they extend the period in which markets must price OpenAI from secondhand disclosures, even as the company holds early-stage talks about raising new money. Thursday offered a preview of what a prospectus will eventually impose: defined, audited line items in place of a contested headline, separating gross from net and showing what each dollar of revenue costs in compute. For OpenAI, that discipline may prove clarifying. For shareholders of Oracle, CoreWeave and Microsoft, the question is narrower and harder: not whether OpenAI is growing, but whether its revenue, measured on its own terms, can grow into the obligations already booked against it.