The dispute over how large artificial-intelligence companies really are is partly an accounting dispute. When a service is sold through a cloud partner, companies can report the gross amount billed, the net amount retained, or a run rate annualised from a single month, producing very different headlines.
Financial Times reporting on OpenAI, summarised by Reuters and CNN, centred on exactly this comparability problem against rival Anthropic. A September run rate is not full-year revenue, and a net figure is not a gross figure; comparing unlike numbers flatters one side or the other.
For readers, the useful questions are consistent: is the number audited, over what period, gross or net, and what costs sit beneath it. Training and inference are expensive, so revenue alone does not establish profit.
This explainer synthesises reputable financial reporting on measurement methods. It draws no conclusion about any companys private valuation.
Reporting is based on statements and reporting available at publication time. Digital News Point verified the central facts against at least two reputable sources and attributes claims to their sources in the text. This story will be updated if confirmed new information materially changes the account, and corrections will follow the site corrections policy.
Gross, net, and why the distinction is not a technicality
Consider a customer who pays one hundred dollars for an artificial-intelligence service delivered through a cloud marketplace. If the provider reports the full one hundred as revenue and records the marketplaces share as a cost, that is a gross presentation. If the provider reports only the amount it retains after the marketplace share, that is a net presentation. Both can be prepared consistently, but they produce different growth rates and different implied market shares when placed side by side without disclosure.
Accounting standards provide frameworks for deciding which presentation fits a given contractual relationship, based on who controls the service, who sets the price, and who bears primary responsibility to the customer. Private companies have more latitude in what they choose to emphasise publicly before audited statements are available, which is why readers should always ask which definition stands behind a headline figure before comparing two companies.
Run rates and the limits of a single month
A monthly run rate annualises one month of activity by multiplying it by twelve. The calculation is simple and can be useful for describing scale quickly, but it embeds strong assumptions. It treats the chosen month as typical, even though enterprise contracts can be signed unevenly, academic and fiscal calendars shift demand, and introductory offers can inflate early adoption that later normalises. A run rate is therefore best read as a snapshot scaled up, not as a forecast and not as revenue that has actually been earned across a year.
Full-year revenue, by contrast, records what was recognised over twelve months under the applicable accounting policy. It will usually be the more reliable basis for comparing companies of different sizes, provided the underlying definitions align. Where only run rates are available, careful reporting names the month, the source, and the fact that the figure is unaudited and annualised.
Revenue is not profit, especially where compute is costly
Artificial-intelligence services carry unusual cost structures. Training a large model requires concentrated spending on specialised chips, data centres, networking, and research staff. Serving the model to customers, known as inference, creates ongoing costs that rise with usage. A company can therefore post rapid revenue growth while investing heavily ahead of that revenue, and the path to profitability depends on utilisation, pricing, model efficiency, and the mix between training and inference over time.
For readers, a complete picture pairs revenue with gross margin, operating expense, capital expenditure, and cash flow where those figures are disclosed. When they are not disclosed, the honest conclusion is limited: the size of the top line is only partly known, and its quality cannot yet be judged. Digital News Point reflects that limit by attributing figures, naming their period and basis, and avoiding any conclusion about private valuation that the published evidence does not support.
A short checklist for any AI revenue headline
When a striking revenue number appears, four questions bring it into focus. What period does it cover, and is it actual or annualised? Is it gross or net of partner shares? Has it been audited, or is it an operational update? What major costs sit beneath it before profit? Applying that checklist to each company in a comparison, using the same answers for each, is the simplest protection against a misleading ranking built on unlike numbers.
