Anthropic Says $65 Billion. OpenAI Says $40 Billion. Only One of Them Is Counting the Same Way.

Bloomberg reported on Monday, August 17, 2026 that Anthropic told investors its annualized revenue run rate passed $65 billion at the end of July. That is more than seven times where the company sat at the close of 2025. Two days earlier, preliminary second quarter revenue landed above $11.5 billion against $787 million in the same quarter a year ago. The company has filed confidentially for an IPO with Morgan Stanley, Goldman Sachs and JPMorgan on the cover, and could list as early as this fall.

Every outlet ran the same comparison: Anthropic at $65 billion, OpenAI at $40 billion. Case closed, one lab is lapping the other.

I do not think that comparison survives contact with the S-1, and I want to be careful about why. This is not a claim that Anthropic is inflating anything or doing anything improper. The growth is extraordinary under any convention. The problem is narrower and more boring: the two companies are not counting the same dollar the same way, and nobody has been forced to reconcile that in an audited document yet.

First, the Growth Curve, Because It Is Genuinely Absurd

Before the accounting argument, the numbers deserve their own paragraph. This is one of the steepest revenue ramps anyone has put on paper.

The quarterly line tells the same story without the annualization multiplier doing any work: $787 million in Q2 2025, $4.73 billion in Q1 2026, more than $11.5 billion preliminary in Q2 2026. That is above 14x year over year on a trailing quarter. Anthropic also says it posted positive adjusted operating income in the quarter, which would be a first for a frontier lab.

Hold that word "adjusted." We come back to it.

The Same Dollar, Counted Twice Differently

Here is the mechanism. A large share of both companies' enterprise revenue does not arrive through a direct API relationship. It arrives through a hyperscaler: Claude through AWS Bedrock and Google Vertex and Microsoft Foundry, and OpenAI models through Microsoft. The customer pays the cloud. The cloud keeps a cut. The lab gets the rest.

The question accountants have to answer is whether the lab is the principal in that transaction or the agent. Principal means you book the whole customer payment as revenue and the partner's cut as cost of revenue. Agent means you book only your cut.

Anthropic books gross. OpenAI books its Microsoft channel net. Same economics, two very different top lines.

Run the sensitivity yourself and the headline moves a real amount. Take the $65 billion run rate. Assume half of it flows through cloud channels, and assume the blended partner fee on that half is somewhere between 15 and 25 percent. That is $4.9 billion to $8.1 billion of revenue that a net reporter would never have put on the top line, putting a net-equivalent figure somewhere around $57 billion to $60 billion.

I want to flag those assumptions loudly. The channel mix is not public and neither is the blended fee. I picked a 50 percent mix because Anthropic's enterprise motion leans hard on Bedrock and Vertex, and a 15 to 25 percent band because that is the ordinary range for this kind of marketplace arrangement. If the real mix is 30 percent, the haircut is half the size. If it is 70 percent, it is larger. The point of the exercise is not the specific number. The point is that the gap between $65 billion and $40 billion is not entirely a gap in business performance, and right now nobody outside either company can tell you how much of it is.

The Case That Gross Is the Right Answer

The lazy version of this article treats gross reporting as a trick. It is not, and the argument for it is strong enough that I expect the auditors signed off without much drama.

Anthropic sets the model price. Anthropic controls the weights, the serving behavior, the rate limits, the deprecation schedule, and the safety policy that governs what the model will and will not do. The hyperscaler is running infrastructure and billing against a service it does not define. Under the ordinary principal versus agent test, controlling the service before it transfers to the customer is exactly what makes you the principal. Plenty of software companies book gross on far weaker facts.

And here is the part that cuts against my own framing: growth rate is invariant to the convention. If you haircut every period by the same percentage, 14x year over year is still 14x. The slope does not care. Whatever discount you apply to $65 billion, you apply to the $9 billion it grew from. The trajectory is real and it is not an artifact.

The convention matters for two things instead: the absolute number in a headline comparison against a peer using a different convention, and gross margin percentage, because gross reporting mechanically compresses margin by stuffing partner fees into cost of revenue.

Now Back to "Adjusted"

The second number in the release deserves the same treatment. Positive adjusted operating income in Q2 would be the first time a frontier lab cleared that bar, and it lands with a modifier doing load-bearing work.

To Anthropic's credit, the figure reportedly includes model training costs, which is the expense skeptics assumed would be quietly excluded. That is the hard one and it is in there. What it excludes is stock-based compensation, and critics have argued the adjustment also sidesteps training compute amortization and infrastructure capital expenditure.

Stock comp at a company that has raised this much private capital at these valuations is not a rounding error. It is plausibly large enough on its own to move a thin positive operating line back under zero on a GAAP basis. That is not an accusation. It is the reason the word "adjusted" is in the sentence, and it is the single most likely place a fall prospectus produces a headline that reads worse than the August press cycle did.

Why This Matters If You Are Just Buying Tokens

If you build on this stuff rather than trade it, there is still something practical here.

A vendor that books channel revenue gross and pays a 15 to 25 percent partner fee out of it does not see your Bedrock dollar and your direct API dollar as the same dollar. The direct relationship is materially better for them. That asymmetry is where enterprise discounting pressure actually comes from, and it predicts something specific: over the next several quarters, expect the aggressive commit pricing and the earliest access to new capability tiers to show up on the direct path first, with marketplace parity arriving later.

It also means the channel decision is a negotiating lever. Procurement teams default to buying through the hyperscaler they already have a contract with because it is easier. Easier is worth something. It is not always worth the spread, and it is worth knowing the vendor has a reason to want you off that path.

The other durable read is on price stability. A company approaching a public listing has an incentive to protect reported gross margin through the quarters that become the comparison base. That is a mild argument against expecting deep list-price cuts on the flagship tiers near term, in a market where almost every other headline rate has been moving. Introductory rates reverting on schedule is the more likely shape than a new cut.

Our Take

Anthropic went from $9 billion to $65 billion in seven months. That is the story and no accounting argument touches it. I do not want the rest of this piece read as debunking, because there is nothing here to debunk.

What I am arguing is narrower: the industry has spent eighteen months ranking labs by a self-reported metric that has no standard definition, no audit, and no requirement that any two companies compute it the same way. Run rate is a marketing unit. It got treated as a scoreboard because it was the only number available.

The interesting thing about an IPO is that it ends that. An S-1 forces a single company to state its revenue recognition policy in writing, get it audited, and live with the number. When Anthropic files publicly, we will get the first hard data point on what frontier lab revenue actually looks like when someone with liability exposure has to sign it. If OpenAI follows, we get a second, and the comparison becomes possible for the first time.

My honest expectation is that the audited numbers come in lower than the press-cycle numbers and that this changes very little, because a company growing at this rate can absorb a restatement in a way a company growing at 30 percent cannot. The risk is not the size of the haircut. The risk is the gap between the two figures being large enough that the market decides it was managed rather than merely inconsistent. Those are different problems and only one of them is fatal.

Three things I am watching:

One. Whether the S-1 breaks out channel revenue as a disclosed line item or buries the principal versus agent determination in a revenue recognition policy note. A number in a table means the comparison against OpenAI can finally be done. A policy note means it still cannot.

Two. Whether the first audited GAAP operating line for Q2 2026 stays positive once stock-based compensation is back in it. If it does, the profitability claim was conservative and the skeptics were wrong. If it does not, the gap between the two figures becomes the story for a quarter.

Three. Whether OpenAI clarifies or changes its own convention before its own listing. Net reporting understates OpenAI in exactly the way gross reporting flatters Anthropic, and the first company to file sets the disclosure standard the second one gets measured against.

Until then, treat $65 billion and $40 billion as two numbers from two different measuring systems, and be suspicious of anyone who subtracts one from the other.