From Shubhi K | Product & Market Analysis

The AI Revenue Leaderboard: Who Actually Has Real Revenue in 2026

On this page

Announced commitments, annualised run rates and booked revenue are reported as if they were the same thing. They are not. One is a promise, one is a projection and one is an audited result. Ranking AI companies by headline number without separating those three produces a leaderboard that is confidently wrong, and almost every published version does exactly that.

Key takeaways

  • Three different measures get quoted interchangeably. Booked revenue, annualised run rate and announced commitments describe entirely different things.
  • Reporting basis changes the number materially. Anthropic reports on a gross basis counting reseller spend as revenue; OpenAI reports closer to net.
  • The distribution is a cliff, not a ladder. The top two companies hold the large majority of combined run rate across the leading names.
  • A top ten AI company is usually doing hundreds of millions. The median run rate across the leading names sits around $1 billion, which the headlines obscure.
~$47BAnthropic's reported annualised run rate as of late May 2026, on a gross basis.
~$25BOpenAI's approximate annualised revenue from roughly $2 billion per month, company confirmed.
$13.07BOpenAI's booked 2025 revenue per leaked audited financials, against a $20.9 billion operating loss.

The comparability problem

Three measures circulate and they are not interchangeable.

Booked revenue is what a company earned in a completed period, ideally audited. It is backward-looking and it is the only one with a defined meaning.

Annualised run rate takes a recent period, usually a month, and multiplies it out. It is forward-looking, flattering during a ramp and unflattering during a slowdown. It is not revenue.

Announced commitments are purchase agreements or investment frameworks, often spanning a decade. A $250 billion commitment across ten years is not $250 billion of anything this year.

Three measures, three meanings What each figure actually describes Time direction Verifiable? Common misuse Booked revenue Backward If audited Rarely misused Annualised run rate Forward No Quoted as revenue Announced commitment Multi-year Partly Quoted as annual
The second row is where most confusion originates. A run rate is a projection wearing the clothes of a result.

The leaderboard, with the caveats attached

Ranked by reported annualised run rate as of August 2026. Every figure carries a basis note because the basis changes the number.

Reported annualised run rate, August 2026 In $ billions. Basis differs between companies, which is the point. Anthropic (gross basis) ~$47B OpenAI (closer to net) ~$25B Cursor / Anysphere ~$2B Next tier, combined ~$6B across several Median across leading na… ~$1B Sources: company announcements and reported figures. Bases are not consistent and should not be compared directly.
The bottom bar is the one worth sitting with. Half the leading names sit at or below it.

Anthropic reported approximately $47 billion annualised as of late May 2026, up from $9 billion at the end of 2025. OpenAI has confirmed roughly $2 billion per month, near $25 billion annualised.

Both are extraordinary businesses. Neither figure is directly comparable to the other, for the reason set out below.

Growth rates need the same discipline

The revenue figures get scrutinised occasionally. The growth rates almost never do, and they carry the same problems.

A company moving from $9 billion to $47 billion in five months is growing at a rate with no software precedent, and both endpoints are run rates rather than results. The growth rate is therefore a projection divided by an earlier projection.

That does not make it false. It makes it a different kind of claim from a company reporting audited annual revenue growth, and the two get compared as though they were equivalent.

The practical test is the same one applied to the levels. Ask what period each endpoint describes, whether the basis was consistent across both, and whether either has been restated since.

Why nobody fixes this

The incentives run against clarity. A company with a favourable basis has no reason to flag that the comparison is unfair, and a company with an unfavourable one gains nothing by drawing attention to a smaller number.

Publications repeating the figures are not equipped to adjust them and would be criticised for editorialising if they tried. The result is a shared convention where everyone quotes what they were given.

Gross against net, and why it matters

Anthropic reports on a gross basis. Total end-customer spend flowing through cloud resellers is counted as revenue, with payments to those partners booked as an expense.

OpenAI reports closer to net. Reporting the same underlying business on the other basis would produce a materially different number.

Neither approach is wrong and both are used legitimately across software. What is wrong is placing them side by side in a ranking without saying so, which is what almost every published leaderboard does.

What this means in practice

If you are comparing two AI companies, the accounting policy note matters more than the growth rate. Until a public filing states the basis explicitly, any gap between two reported figures could be a difference in business size or a difference in accounting convention.

Both companies have filed confidentially for listings, which means this will be resolved. When the public documents appear, the accounting policy note will be the most valuable page in them, a point examined in the analysis of the enterprise listing.

The distribution is a cliff, not a ladder

The most useful finding from assembling this is structural rather than about any one company.

Where the combined run rate actually sits Approximate share of combined run rate across the leading AI companies 86% held by two 86% Top two companies 14% Everyone else combined Based on reported run rates across the leading AI companies as of August 2026. Bases differ between companies.
This is not a competitive market with a leader. It is two companies and a long tail.

Across the leading names, the top two account for the large majority of combined run rate. The median sits around $1 billion, and several companies routinely described as major AI businesses are doing hundreds of millions.

That distribution matters for anyone reading market commentary. Statements about the AI industry's revenue are, in practice, statements about two companies, and the experience of everyone else is not captured by them.

What concentration means for everyone else

If you are building an AI company, this distribution is the market you are actually in. The comparisons that matter are not with the top two, whose economics and capital access are unavailable to anyone else.

They are with the long tail, where a $50 million run rate is a genuine achievement and where the funding environment described in the analysis of venture concentration applies with full force.

The reporting convention of grouping all of these companies into a single category called AI does real damage here, because it sets expectations against two outliers rather than against the median.

Commitments are not revenue and rarely become it in full

The largest numbers in this sector are purchase commitments, and they are the least meaningful when quoted without a term.

The pattern of announced figures shrinking on execution is already documented. Nvidia's announced framework to invest up to $100 billion in OpenAI was finalised at $30 billion. That is a 70% reduction between announcement and definitive agreement, and it is covered in the breakdown of AI circular deals.

Apply the same discipline to any commitment. Divide by the term. Compare the annual figure to the buyer's current revenue. If the annual commitment exceeds what the buyer earns in a year, you are looking at a plan rather than a purchase.

There is a second reason commitments deserve suspicion in this sector specifically. Several of the largest are between parties who also hold equity in each other, which means the buyer's ability to pay is partly a function of capital the seller provided. That structure is examined separately, and it makes commitment figures harder to read than they would be between unrelated counterparties.

Where this ranking is weak

Several problems, and they are not minor.

Every figure here is company-reported or press-reported. None comes from an audited public filing, because none of these companies files publicly yet. Leaked financials have been widely reported and not disputed, which is not the same as verified.

Run rates at companies growing this fast go stale within weeks. Every previous monthly snapshot in this sector has been outdated by the time it was published, and this one will be too.

Revenue per employee, a favourite metric in this category, has been omitted deliberately. Reliable, dated headcount figures exist for almost none of these companies, and computing a ratio from an unreliable denominator produces a confident number that means nothing.

How to use this

When you seeAsk
A revenue figure with no dateAs of when? Run rates in this sector are stale within a quarter.
Two companies compared directlyIs the reporting basis the same? Gross and net are not comparable.
A very large commitment numberOver how many years, and how does the annual figure compare to the buyer's revenue?
"Fastest growing company ever"Growing on which measure, from what base, and reported by whom?
A market sizing claimDoes it describe the industry, or does it describe two companies?

None of this requires financial training. It requires refusing to accept a number without a date, a basis and a source, which is a habit rather than a skill.

Why this will get easier soon

Both major labs have filed confidentially for public listings. When public documents appear, several of the questions above become answerable from primary sources for the first time.

The accounting policy note will settle the gross-versus-net question. The revenue recognition disclosure will settle what counts as booked. And the risk factors section will state, under legal obligation, what each company believes could go wrong.

Until then, every leaderboard including this one is an assembly of company statements and press reports. Treating it as anything more is the specific error this post exists to discourage.

It is worth noting what public filing will not settle. Private companies further down the list have no listing plans and no disclosure obligation, so the median figure and the shape of the long tail will remain estimates even after the top two become verifiable. The concentration finding will get sharper at the top and stay fuzzy everywhere else.

Frequently asked questions

Which AI company has the most revenue?

On reported annualised run rate, Anthropic led as of August 2026 at approximately $47 billion, ahead of OpenAI at around $25 billion. The comparison is not clean, because Anthropic reports on a gross basis counting end-customer spend through cloud resellers as revenue while OpenAI reports closer to net. The gap between them is partly accounting convention.

What is the difference between run rate and revenue?

Booked revenue is what a company earned in a completed period, ideally audited. Annualised run rate takes a recent period, usually a month, and multiplies it out to a yearly figure. Run rate is a forward-looking projection that flatters a business in a steep ramp and would equally exaggerate a slowdown. It is not revenue.

Why is gross versus net reporting important?

Because it changes the headline number materially for the same underlying business. Gross reporting counts total end-customer spend flowing through resellers as revenue, with partner payments booked as an expense. Net reporting does not. Placing figures reported on different bases side by side in a ranking produces a comparison that means nothing.

How concentrated is AI revenue?

Extremely. Across the leading AI companies, the top two account for the large majority of combined run rate, with the median sitting around $1 billion. Several companies routinely described as major AI businesses are doing hundreds of millions. Statements about industry revenue are, in practice, statements about two companies.

Are announced AI commitments the same as revenue?

No, and they are frequently quoted as though they were. Purchase commitments often span a decade, so a headline figure should be divided by the term before it means anything. Announced arrangements also shrink on execution: Nvidia's announced framework of up to $100 billion for OpenAI was finalised at $30 billion.

Can these revenue figures be verified?

Not currently. None of these companies files publicly, so every figure is company-reported or press-reported. Some derive from leaked financials that have been widely reported and not disputed, which is not the same as audited. Both major labs have filed confidentially for listings, so public documents should eventually resolve this.

Where to start this week

One habit, applied to the next AI revenue figure you read.

Before accepting it, find three things: the date it refers to, whether it is booked revenue or a run rate, and whether the basis is gross or net. If any of the three is missing from the article, the number is decoration rather than information.

You will find that most published figures fail this test, including several that circulate as settled facts. That is the useful discovery, and it takes about thirty seconds per number.

Apply it to this post as well. Every figure here carries a date and a basis note for exactly that reason, and where a number could not be given both, it was left out.

References

  1. Anthropic Series H announcement and reported run rate figures, 29 May 2026, compiled from company announcements and public reporting.
  2. Forbes, OpenAI and Anthropic are testing two very different AI business models, 21 May 2026. Used for reporting basis and revenue mix.
  3. Leaked audited 2025 financials for OpenAI, reported by the Financial Times and The Information. Used for booked revenue and operating loss.
  4. PYMNTS, Nvidia signals final investments in OpenAI and Anthropic, 4 March 2026. Used for the commitment reduction example.

No company in this ranking files publicly. Every figure is company-reported or press-reported, reporting bases differ between companies, and run rates in this sector have historically gone stale within weeks of publication.

SK
Shubhi K
Founding Member, Zan Digital. Writes about AI product economics, B2B software markets and what the numbers behind vendor claims actually say.

Related reading