Reports of a $20B discrepancy in OpenAI annualized revenue sent NVIDIA shares down nearly 3%.

How did OpenAI's annualized revenue suddenly differ by 20 billion US dollars?

On October 8, the UK's Financial Times reported that documents provided by OpenAI to investors showed that, as of the end of September, the company's annualized revenue was close to 50 billion US dollars. This was lower than the figure of close to 70 billion US dollars that was widely reported at the end of September. The report intensified the market's concerns about AI growth. At the close of trading that day, NVIDIA fell by about 2.9%, Oracle by about 5.5%, and the NASDAQ Composite Index dropped by 1.25%.

Less than a month earlier, Anthropic had also made progress on profitability. On September 13, Reuters, citing a Financial Times report, said that Anthropic had told its shareholders it expects adjusted operating profit to be positive for a second consecutive quarter.

Now, the two sets of annualized revenue figures for OpenAI differ by about 20 billion US dollars. While the market questions whether AI can make money, it must first resolve a more fundamental question: how exactly are these revenue figures calculated?

Did OpenAI really earn 20 billion US dollars less than previously estimated?

This is not the correct interpretation.

On September 29, Reuters, citing people familiar with the matter, reported that OpenAI's annualized revenue was close to 70 billion US dollars. Documents disclosed by the FT on October 8 show that the figure is close to 50 billion US dollars. The discrepancy involves how a particular type of revenue is calculated: how should fees paid by customers using AI models through cloud platforms be counted?

This type of revenue may involve revenue sharing between cloud platforms and model companies.

CNN, citing people familiar with the matter, said that the earlier 70 billion US dollar figure did not come from OpenAI and may have resulted from outside parties adjusting the statistical method to compare it with Anthropic. According to this explanation, for relevant sales revenue from cloud platform partnerships, Anthropic uses a gross revenue basis, while OpenAI uses a net revenue basis.

Consider a hypothetical example: a customer pays 100 units of currency, the cloud platform takes 20 units, and the model company keeps 80 units. On a gross basis, the revenue is recorded as 100 units, and the 20 units distributed to the platform are counted as a separate cost. On a net basis, the revenue is recorded directly as 80 units.

For the same transaction, the amount paid by the customer remains unchanged, but different calculation bases can lead to different reported revenue figures. The revenue-sharing amounts here are purely illustrative and do not represent either company's actual situation. The 80 units retained are not profit, as other costs such as service provision must still be deducted.

Therefore, the 20 billion US dollar discrepancy cannot be directly counted as lost revenue for OpenAI, nor can it be entirely attributed to the share taken by cloud platforms. The currently public information is not sufficient to match each item one by one.

However, comparing the revenue scales of OpenAI and Anthropic requires first aligning the statistical period and calculation basis.

50 Billion US Dollars Annualized Does Not Mean Earning 50 Billion US Dollars in One Year

There is another easily overlooked term here: annualization.

Annualized revenue typically converts recent revenue into a full-year scale. A common method is to multiply one month's revenue by 12. For example, if a company has monthly revenue of 1 billion US dollars, its annualized revenue at this rate is 12 billion US dollars.

However, this does not mean the company generated 12 billion US dollars in revenue over the past year, nor does it guarantee that it will reach that figure in the coming year.

For fast-growing companies, annualized indicators can show the latest business scale. To see full-year performance, one still needs to look at actual full-year revenue.

Why Did NVIDIA and Oracle Also Fall?

OpenAI is not yet listed, but the market's judgment of its revenue has already affected the share prices of other companies, because its business is closely linked to theirs.

Consumers pay for AI subscriptions, enterprises pay for model calls, and that revenue flows into the accounts of model companies. To provide services and develop models, these companies need to purchase computing power, which drives demand for upstream cloud services, servers, chips and data centers.

NVIDIA's AI Factory platform covers data center design, computing power management software, and infrastructure such as servers, networking, power supply and cooling. Demand for computing power from model companies drives investment across these links, and whether costs can be recovered through user payments will affect how far the next round of expansion can go.

Investors are therefore concerned not only with how high OpenAI's revenue is, but also with whether it can rely on business revenue and financing to pay for continually expanding computing power bills.

According to CNN, US stock indexes had already fallen at the open that day, and the decline widened after the Financial Times report was released during trading hours.

However, the revenue controversy was only one factor that day. Reuters' closing report also mentioned rising oil prices and the resulting concerns about inflation and interest rate hikes, so the entire decline cannot be attributed to OpenAI.

Similarly, the market's concerns do not mean that AI demand has already weakened. Whether procurement has declined and whether contracts can be fulfilled will depend on operating information disclosed later. Still, the growth of model companies, the expansion of suppliers, and investors' expectations of returns are already closely linked.

The Next Round of Computing Power: Who Will Pay the Bill?

In January this year, OpenAI CFO Sarah Friar wrote in an official article that accessing top-tier computing power requires commitments made years in advance, and business growth is not stable: sometimes computing power is in place before demand arrives, and sometimes demand exceeds existing capacity.

Investment in computing power must be arranged in advance, but future revenue remains uncertain. This is why the market is closely watching AI companies' revenue.

When examining revenue figures, the following questions need to be clarified going forward: Which period is being counted? Is it an annualized scale or actual quarterly revenue? How are sales through cooperating platforms calculated?

Beyond that, one needs to see how much funding remains after deducting costs such as services, R&D and personnel, and when future computing power bills will be paid and what will be used to pay them.

Jensen Huang once summed up the business logic of AI in one sentence: "Computing power is revenue."

Elaborating on this in his GTC speech, Huang said that faster deployment to production, higher output efficiency, fewer operational interruptions and longer service life would all affect the revenue of AI factories.

Computing power supports more capable models, models attract paying users, and revenue then funds the next round of R&D and computing power construction.

The sustainability of this cycle depends on whether computing power can be converted into services that users are willing to pay for continuously, and whether it generates enough funds to support reinvestment.

If more and more computing power is purchased but payment demand cannot keep up, or revenue fails to cover costs over the long term, expansion will still depend on external financing.

That is why OpenAI's revenue figures affect the entire industrial chain: the market needs to see that invested funds can ultimately be recovered through user payments, so that the next round of expansion has firmer footing.

For everyday users, these huge investments will ultimately show up in the models we use every day: is the time they save and the work they complete worth the subscription fee?