• Monday Momentum
  • Posts
  • We Said Anthropic Passed OpenAI. It Depends Who Is Counting.

We Said Anthropic Passed OpenAI. It Depends Who Is Counting.

How two AI companies record the same dollar of revenue in different ways, and why the scoreboard this newsletter published in August needs a correction six weeks before a $2 trillion IPO

Author

Justin Wright
October 12, 2026 • Est. Reading Time: 6 minutes

Happy Monday!

Both companies are GAAP compliant. They simply count sales made through cloud partners differently. (Source: Axios)

In August, this newsletter ran a piece titled "Anthropic Passed OpenAI by Ignoring the Customer Everyone Wanted." The argument was that Anthropic's enterprise strategy had overtaken OpenAI's consumer strategy, and the headline evidence was that Anthropic had surpassed OpenAI in quarterly revenue for the first time.

This week we learned that comparison was not as clean as we initially presented it. Axios, the Financial Times, and Reuters reported that OpenAI's annualized revenue is about $50 billion, not the roughly $70 billion that circulated in September. Nothing was lost and no customers left. The $20 billion gap exists because Anthropic and OpenAI record sales made through cloud partners in different ways, and the $70 billion figure was OpenAI's revenue recalculated using Anthropic's method so investors could compare the two.

That means the August comparison set one company's revenue, counted generously, against the other's, counted conservatively. This issue explains what happened, what it changes, and why it matters more now that Anthropic is asking public investors for $2 trillion.

Anthropic books the full value of sales made through AWS and Google Cloud as revenue and records the cloud provider's cut as an expense. OpenAI books only its own share of certain partner sales. Both approaches comply with accounting rules. Counted Anthropic's way, OpenAI is at roughly $70 billion; counted OpenAI's way, about $50 billion. Cloud marketplaces drove 47% of Anthropic's 2025 revenue. Who leads the AI revenue race depends on the convention, and our August scoreboard did not account for it.

TL;DR

Two Ways to Count the Same Dollar

Imagine a company pays $100 to use Claude through Amazon Bedrock. Amazon keeps a cut and passes the rest to Anthropic. How much revenue did Anthropic earn?

Under Anthropic's approach, the answer is $100. It records the full amount as top-line revenue and lists Amazon's share as a sales and partnerships expense. Under OpenAI's approach for certain partner sales, the answer would be only the portion it actually keeps.

Francine McKenna, an accounting professor who writes the newsletter The Dig, told Axios the difference turns on a standard accounting question: who is the principal and who is the agent? The company that controls the customer relationship and is responsible for delivering the product can book the gross amount and both conclusions can be defensible. As McKenna has noted, two companies with the same business model can reach different conclusions depending on their contracts.

So this is not misconduct, and it would be wrong to frame it that way. It is a presentation choice with large consequences for anyone comparing headline numbers, which is exactly what this newsletter did.

What This Does to the Scoreboard

Below is the same race measured both ways. The figures come from different months and one is an estimate, so treat this as an illustration of the gap rather than a precise ranking.

The AI Revenue Race, Counted Two Ways

Basis

Anthropic

OpenAI

Gross (partner sales at full value)

~$65B run rate (July)

~$70B (investor figure, September)

Net (only the company's own share)

~$60B (estimate*)

~$50B (October)

Method used in its own headline numbers

Gross

Net

*Our estimate, assuming Anthropic's 2025 cloud channel mix of about 47% and the roughly 16% platform take rate Reuters derived from the prospectus. Anthropic has not disclosed a net figure for 2026.

On a gross basis, the two companies look roughly even, with OpenAI slightly ahead. On a net basis, Anthropic still appears to lead. What does not hold up is the version we published, which compared Anthropic's gross number with OpenAI's lower figure and presented the crossover as a settled fact. In August we cited OpenAI's run rate at about $40 billion against Anthropic's $65 billion. That framing made the gap look far wider than any consistent comparison supports.

Why It Matters Six Weeks Before the IPO

The channel question is not a rounding error for Anthropic. Prospectus-based reporting shows sales through cloud marketplaces reached about $2.16 billion in 2025, roughly 47% of revenue. That share has climbed every year, from 11% in 2023 to 32% in 2024 to nearly half last year. Reuters estimated Anthropic paid the platforms about $351 million in distribution fees, which works out to roughly 16 cents of every dollar sold through them.

That matters for valuation because multiples are applied to revenue. When investors price Anthropic at a multiple of its top line, nearly half of that top line includes money that passes through to Amazon and Google. The revenue is real, but the question is whether a dollar that Anthropic shares with a distributor deserves the same multiple as a dollar it keeps outright.

Axios added a sharper point: annualized run rate, the figure both companies and this newsletter have leaned on all year, is a Silicon Valley construct suited to early-stage startups. Public market investors will want actual reported revenue. The prospectus will force that discipline on Anthropic, and the eventual OpenAI filing will do the same.

The market has not settled on a number either. A Hyperliquid prediction contract implied roughly $2.09 trillion on October 5. Chamath Palihapitiya predicted the deal prices at "$1 trillion or less." Research firm New Constructs put fair value near $144 billion. A range that wide usually means investors disagree about the inputs, not just the outlook.

What We Still Think

Correcting the record does not mean reversing every conclusion, so it is worth being specific about what survives. Anthropic's growth is real under any convention. Revenue grew twelvefold in 2025, and quarterly revenue more than doubled from Q1 to Q2 2026. The enterprise strategy is working, and on a consistent net basis Anthropic still appears to be ahead. Its Q2 adjusted operating profit, which we already noted last week excludes stock-based compensation, is a meaningful milestone.

What changes is the certainty; OpenAI is not trailing by the wide margin our August piece implied. It says it expects to exit the year above $70 billion, and the price war is visibly shifting spend: OpenRouter reported in September that its users spent more on OpenAI models than Anthropic's for the first time in two and a half years. Anthropic answered this week with Haiku 5.5 at $0.10 per million input tokens, exactly matching GPT-6 Luna. This is a closer race than "the scoreboard flipped" suggested.

What This Means for Practitioners

For anyone reading AI company financials, ask three questions before comparing numbers: Is the figure annualized or actual? Is it gross or net of partner sales? Which month is it from? A surprising share of AI coverage this year, ours included, skipped at least one of them.

For enterprise buyers, the channel data has a practical implication. Nearly half of Anthropic's revenue flows through AWS and Google, which gives those platforms real leverage over pricing and terms. If you buy Claude through a cloud marketplace, you are part of the most strategically important revenue stream Anthropic has, and that is worth remembering at renewal.

For investors watching the IPO, the prospectus footnotes on revenue recognition will matter more than the headline. As McKenna has pointed out, the audited statements and footnotes are where the treatment of cloud partner sales gets spelled out.

The Bottom Line

The AI revenue race has been reported all year as a single scoreboard. It was really two scoreboards using different rules, and the company with the more generous rule is the one going public first. Neither company did anything improper, the problem was the comparison.

The useful lesson is not that Anthropic's numbers are inflated or that OpenAI's are understated. It is that "who is winning" was always partly an accounting question, and the honest answer for now is that the race is closer than the headlines suggested. When the audited numbers arrive, we will rerun the comparison on a consistent basis and report what we find.

In motion,
Justin Wright

If two companies can sell the same product through the same cloud partners and report revenue that differs by billions of dollars, both under the same accounting rules, how much of what we think we know about which AI company is winning is really a measure of business performance, and how much is a measure of accounting judgment?

Food for Thought

Quick Hits

  • Anthropic disclosed that its models took unintended real-world actions during testing, including Claude Haiku 4.5 submitting a false tip to Philadelphia's unsolved-murders site in July and a testing model filing 20 incomplete visa applications through the State Department's public form. Police called Anthropic's two-month notification delay "unacceptable," and the White House demanded faster incident reporting. (TechCrunch)

  • Anthropic launched Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens, a 90% cut in list price from Haiku 4.5 and the same rate as GPT-6 Luna. (Technology Org)

  • Google paused new product vulnerability submissions to its open-source bug bounty until at least early 2027, citing "a significant rise in automated submissions, the vast majority of which are not valid." (TechCrunch)

If you haven’t listened to my podcast Mostly Humans: An AI and business podcast for everyone yet, new episodes drop every week!

Episodes can be found below - please like, subscribe, and comment!