OpenAI's Reported Revenue Was $70B. Then $50B. What Changed?

The number was enormous before it became controversial.

Reports in September 2026 suggested OpenAI's annualized revenue was approaching $70 billion. Then, in early October, the Financial Times and Reuters reported that the company's September annualized figure was closer to $50 billion. The headline difference was roughly $20 billion.

An ordinary reader might reasonably ask whether OpenAI had lost a huge amount of business in a matter of weeks.

But that is not what these two figures establish. They come from reporting about a private company's revenue run rate, and the mismatch appears to involve how sales through cloud partners were counted when OpenAI was compared with competitor Anthropic.

The question is less sensational and more important: what exactly are investors measuring when they compare the fastest-growing AI companies in the world?

A run rate is not a year's audited revenue

Annualized revenue figures try to express a recent sales pace as an annual number. In a simplified example, if a company sells $1 billion of services in one month and multiplies that by twelve, the resulting annualized run rate would be $12 billion.

That calculation can be useful for a rapidly growing business. An old fiscal-year total may understate how much the company is selling today.

But it is also easy to misunderstand. A run rate is not the same as cash received, a full year's realized revenue, profit, or a guarantee of next year's sales. Seasonal effects, contracts, pricing, changing products, and the precise snapshot date can all matter.

For the same reason, a difference between two figures described as annualized revenue does not automatically mean that actual sales fell by that amount.

A more basic test comes first: Were the two figures built using the same counting rules?

The customer does not always buy directly from the model developer

Consider an enterprise that wants to use an advanced AI model. It might pay the model developer directly through its API. Or it might buy access through a cloud provider that bundles computing infrastructure, support, billing, and integration.

Those are different commercial routes to what may look like a similar service from the customer's point of view.

Now imagine a hypothetical customer pays $100 for model access delivered through a cloud platform. Depending on which company is the principal in the transaction, its contractual obligations, and the accounting interpretation, one party could report the gross customer amount while another could report only its own share of that transaction.

This is an illustration of the measurement question, not a claim that an actual $100 OpenAI or Anthropic contract is allocated in a particular way.

The important point is that gross customer spending, the amount paid to a model supplier, and a company's recognized revenue need not be identical.

Why Anthropic entered the comparison

The October reporting said the larger OpenAI number had been shared in an effort to produce a more comparable figure against Anthropic's reported revenue pace. According to Reuters, Anthropic includes revenue from certain cloud-partner sales in the figures under discussion, while OpenAI's more narrowly stated amount excluded such sales.

Axios also described the difference as a reporting-boundary issue rather than a simple disappearance of customer demand.

That distinction matters, but it should not be overstated.

It does not establish that either company fabricated sales, that one breached accounting rules, or that a precise $20 billion pile of cash was lost. The public reporting does not provide the complete underlying private-company ledgers needed to independently reconcile the numbers.

There may be valid reasons for gross-versus-net treatment in particular transactions, including who controls a product or service before transfer, who is responsible for delivery and how contracts are structured. A detailed accounting judgment depends on the actual agreements—not merely on a headline.

So the safe conclusion is narrower: the publicly discussed numbers were not necessarily prepared on an identical basis.

Even the dates must be compared carefully

Private-company growth headlines also arrive at different moments.

Anthropic's own May 28 announcement said its annualized run rate had crossed $47 billion earlier that month. That is a company-reported snapshot from May, not a September figure and not a substitute for audited financial statements.

Similarly, media references to OpenAI's near-$70 billion number and its subsequently reported near-$50 billion September run rate need to be read with the source, period, and intended inclusion rules attached.

It is tempting to assemble every headline into a single league table. But mixing dates and definitions produces a ranking that may be more precise-looking than meaningful.

A better comparison asks how much actual customer demand each company serves, how much of that demand it reports as revenue, what it must pay partners to deliver the service, and how efficiently it converts that activity into cash.

The cloud platforms have their own business to protect

The disputed boundary points to a larger fact about the AI industry.

The model developer is not always the company operating the server, owning the customer relationship, or billing the end user. Large cloud platforms can occupy the distribution and infrastructure layers between a powerful model and the business that pays to use it.

That creates important strategic questions beyond headline revenue.

Who owns customer access? Who sets the effective retail price? Who funds the data center, chips, electricity and networking? And which company earns the remaining economic margin after everyone has been paid?

A laboratory can have a remarkable model and fast-growing usage while still depending heavily on partners to make that capability available at scale.

A cloud platform can gain influence because it controls access, procurement, and the operational relationship—even when it did not build the underlying model.

The distinction between the gross amount a customer spends and the net economics retained by each participant is fundamental to understanding these relationships.

Four questions behind every giant AI revenue headline

When another private AI company announces—or is reported to have reached—an extraordinary annualized-revenue figure, ask four things before comparing it with a rival.

What period? A full-year realized result, an end-of-month snapshot, and a projected future year are different measurements.

What scope? Does the headline include sales through cloud partners, resellers and other channels, and are those revenues counted gross or net?

What source? Is the number taken from an audited filing, a company announcement, an investor discussion, or anonymous-source reporting?

What economics? How much of the reported revenue is retained after computing, hosting, partner fees and other costs?

OpenAI's reported $70 billion versus $50 billion discrepancy is a useful reminder that very large numbers do not automatically provide a clear picture.

In an industry where customers, model developers, and cloud platforms share the same commercial transaction, the ruler matters as much as the number it measures.