WhitmanTrading

OpenAI's $20 Billion Revenue Gap, Explained

The $20 billion gap in OpenAI's annualized revenue is a counting difference, not lost sales. Late September reports put it near $70 billion by adding partners' gross sales. OpenAI's own figure, reported by the Financial Times and confirmed by CNBC on 8 October 2026, was close to $50 billion.

None of the reports said that OpenAI lost $20 billion of annualized revenue. Two ways of counting the same business produced two different totals, and the larger one had been repeated widely before OpenAI’s own figure surfaced.

This page sets out what was reported and by whom, how gross and net counting differ, what the stocks tied to OpenAI did that day, and what the public filings of its suppliers show. OpenAI is a private company, so every OpenAI revenue figure here is a reported figure, not an audited one.

The two figures and who reported them

The higher figure. On 29 September 2026, Reuters cited a source who said OpenAI’s annualized recurring revenue was approaching $70 billion. TechCrunch and AFP both describe that figure as one reported late last month from information shared with investors. CNBC’s own late-September figure was $68 billion, and AFP refers to that number as the one the network had reported.

The lower figure. On 8 October the Financial Times reported that OpenAI had told investors its annualized revenue was approaching $50 billion at the end of September. CNBC said it confirmed the roughly $50 billion figure, and TechCrunch and AFP both carried the FT report. The difference is $70 billion minus $50 billion, or $20 billion, against the widely repeated number, and $18 billion against CNBC’s own $68 billion.

The explanation that came with it. A person familiar with the matter told CNBC that the $68 billion figure included gross revenue from OpenAI’s partners, which made it easier to compare OpenAI with its main rival, Anthropic. TechCrunch, citing the FT, says the $70 billion figure came from attempts by OpenAI’s own investors to produce that comparison, and notes that Anthropic counts sales made by its cloud partners while OpenAI does not.

What the company has filed. OpenAI confidentially submitted paperwork for a stock market listing in June, according to both CNBC and AFP, and CNBC and TechCrunch both report that a debut is now expected in 2027 rather than this year. Until a prospectus is public, there are no audited OpenAI accounts for anyone outside the company to read.

How it works: gross and net revenue

Annualized revenue is a pace. Reuters describes the measure as often involving multiplying one month’s revenue by 12. AFP says it typically extrapolates a single month’s sales across a full year. Both note that the metric is sometimes criticized as misleading. A strong month makes a strong annualized figure, even if the next month is weaker.

Gross counting takes the whole sale. When a customer buys a product through a partner, such as a cloud provider reselling an AI model, gross counting records everything the customer paid.

Net counting takes only the company’s share. The partner keeps part of the sale, and net counting records only what is left for the company that made the product. The customers, the usage and the cash the company receives are the same under both methods. Only the headline total changes.

Why the choice matters for comparisons. If one company reports gross and another reports net, a side-by-side table makes the net reporter look smaller than it is. That is the comparison investors were trying to fix, according to the FT as cited by TechCrunch. The trouble came when the adjusted number circulated as if it were OpenAI’s own.

The calm reading and the worried reading

The calm reading. Nothing in the reports says customers left or that sales fell. The gap comes from the method, and net counting simply gives the smaller of the two totals. A run rate near $50 billion is still a very large business, and the suppliers’ backlogs below rest on signed customer contracts.

The worried reading. TechCrunch notes that OpenAI’s revenue is used to justify the very large investments being made on its behalf. AFP says investors track OpenAI’s and Anthropic’s annualized revenue as the clearest signal of demand for AI, and that those figures help justify spending on data centers. If the most-watched figure in the trade was a generous version, investors may ask which other numbers were counted the same way.

Both readings rest on the same facts. Neither of them tells you what OpenAI will earn next year.

The original data

We built three charts from figures we could trace. The first shows the reported revenue figures side by side, labeled as reports rather than results.

Three horizontal bars of OpenAI's reported annualized revenue: approaching $70 billion and $68 billion in late September 2026 reports, which counted partners' gross sales, and approaching $50 billion as OpenAI told investors at the end of September.
OpenAI annualized revenue as reported, not audited. The higher figures added partners' gross sales; the lowest is OpenAI's own net figure. Source: Reuters, CNBC, TechCrunch citing the Financial Times, AFP (m78-openai-reported-annualized-revenue-2026.csv).

None of the three bars is a sales total for a calendar year. Each is a monthly pace scaled up to twelve months, and the gap between the top two and the bottom one is the counting method.

The second chart is the first day’s market reaction, from Yahoo Finance closes on 7 and 8 October 2026. Nvidia closed at $230.48, down 2.94% from $237.47. Advanced Micro Devices fell 3.90%, Broadcom 4.35%, Oracle 5.48% and CoreWeave 7.77%. Nasdaq’s own price history gives the same closes, and the moves match the rounded falls CNBC and AFP reported that day.

Horizontal bars of one-day price falls on 8 October 2026: Nvidia down 2.94%, AMD down 3.90%, Broadcom down 4.35%, Oracle down 5.48% and CoreWeave down 7.77%.
Close-to-close change from 7 to 8 October 2026 for five companies that sell chips or computing capacity to AI developers. One day of prices, moved by all of that day's news. Source: Yahoo Finance daily closes (m78-ai-stocks-one-day-moves-2026-10-08-yahoo.csv).

To size Nvidia’s fall in dollars, we multiplied the $6.99 price drop by the 24.1 billion shares on the cover of Nvidia’s latest 10-Q, counted on 21 August 2026. That comes to about $168 billion of market value. As a cross-check, Nasdaq listed Nvidia’s market value at $5.53 trillion at the 9 October close of $229.28, which works out to the same 24.1 billion shares. The share count is rounded to a tenth of a billion, so treat the result as an estimate.

The third chart comes from Oracle’s own SEC filings. Remaining performance obligations are revenue Oracle has under contract but has not yet recognized. At the end of each quarter, Oracle reported:

Nine horizontal bars of Oracle's remaining performance obligations at quarter ends from August 2024 to August 2026, rising from $99.1 billion to $664 billion, with the largest single jump to $455.3 billion in August 2025.
Oracle remaining performance obligations, contracted revenue not yet recognized, at each quarter end from 31 August 2024 to 31 August 2026. Source: Oracle 10-Q and 10-K filings, SEC EDGAR XBRL (m78-oracle-rpo-sec-xbrl-2024-2026.csv).

The figure was $99.1 billion at 31 August 2024, $455.3 billion a year later and $664 billion at 31 August 2026, a rise of $208.7 billion in the latest year and about 6.7 times its level two years earlier.

In between, the filings show $97.3 billion (November 2024), $130.2 billion (February 2025), $137.8 billion (May 2025), $523.3 billion (November 2025), $552.6 billion (February 2026) and $638 billion (May 2026). The biggest single quarter added $317.5 billion, in the three months to 31 August 2025. 24/7 Wall St., reading Oracle’s 8-K of 10 September 2026, gives the same $664 billion and a $209 billion increase.

CoreWeave, which rents out computing power for AI, reported a revenue backlog of about $104 billion at 30 June 2026 in the results it filed with the SEC.

The company says that total leaves out more than $25 billion of net new customer commitments added early in the third quarter, and Pulse 2.0 reported the same two figures. CoreWeave’s backlog includes amounts it estimates under committed contracts, subject to delivering the capacity, so it is a wider measure than Oracle’s.

A worked example: one $100 cloud sale

Take a made-up sale to show the mechanics. A business pays a cloud provider $100 to use an AI model through the provider’s platform. The provider keeps $30 and passes $70 to the model’s maker.

  1. Counted gross, the model maker books $100 of revenue from this sale.
  2. Counted net, it books $70, the part it actually keeps.
  3. Annualized, if this happened every month, the gross pace is $1,200 a year and the net pace is $840 a year. Same customer, same usage, a $360 gap.
  4. Compared across companies, a gross reporter selling exactly the same amount would look about 43% bigger than a net reporter, because $100 is 43% more than $70.

The real split between OpenAI and its partners has not been published, so these numbers only show how the two methods pull apart. They are not an estimate of OpenAI’s figures.

What it means for you

Read revenue headlines with two questions. Is the figure a full year of sales or a recent month multiplied by 12? Does it count what partners collect? A private company’s figure that arrives through investor documents and news reports deserves more caution than a number in an audited filing. Our revenue page explains why the top line is the least informative number on an income statement even when it is audited.

You may already own this story. In State Street’s holdings file for the SPDR S&P 500 ETF Trust dated 8 October 2026, Nvidia was the largest holding at about 8% of the fund, and the ten largest lines added up to about 39% (Alphabet appears twice, as two share classes). Slickcharts’ S&P 500 weights, read on 12 October, give a similar picture: Nvidia first at 7.78% and a top ten of about 39%.

Broadcom and Advanced Micro Devices were both among the 15 largest holdings in each list. A broad index fund spreads money across hundreds of companies, but a few large ones carry much of the weight, which is what concentration risk describes.

Check your own exposure once. Look up the top ten holdings of each fund you own. If the same few technology names lead every list, that overlap is your real exposure to AI news.

Put one day in context. In the 12 months to 9 October 2026, by both Yahoo Finance and Nasdaq closing prices, Nvidia closed 2.94% or more lower on 28 of 251 trading days, including 8 October, and moved at least that much in either direction on 52 days. A fall that size has been routine for this stock.

When it fails: the limits of these figures

Reported is not audited. Every OpenAI number on this page comes from what the company told investors and what reporters were told about it. No public filing confirms either figure. If a prospectus is published, its numbers may differ again.

Annualized figures move with one month. A run rate built from a single strong month can overstate a full year, and one built from a weak month can understate it. The $50 billion is a pace at the end of September, not OpenAI’s sales for 2026.

Backlogs are promises, not cash. Oracle’s remaining performance obligations and CoreWeave’s backlog turn into revenue only as the work is delivered, and CoreWeave says its own figure depends on delivering the capacity. A customer that pays less, later or not at all would change those totals.

One day of prices mixes many causes. A closing price reflects everything traders reacted to that day, not one report alone. Our chart shows what happened to each stock, not how much of the move the OpenAI news caused.

Our own estimates have limits. The $168 billion market value figure uses a share count from 21 August, rounded to a tenth of a billion. The SPY weights are one fund on one day, and Slickcharts' index weights differ by a few tenths of a point, so the exact share depends on the source and the date.

Start with revenue, which explains what the top line does and does not tell you. If the weight of a few holdings in your fund worries you, concentration risk explains how many positions can still depend on one driver, and index funds covers choosing the index first and the fee second. When the word bubble starts appearing in headlines, the stock market bubble page measures the dot-com run against today’s, and explains why a bubble is only confirmed by the fall.

The practical check

When a company’s revenue headline sounds huge, find out whether it is a full year of sales or one month multiplied by 12, and whether it counts what partners collect. If your money is in an S&P 500 fund, look up its top ten holdings once so you know how much of it rides on a few chip and cloud companies.

— Michael Whitman

This page is educational, not financial advice. Test every idea on your own charts before risking money.