There are no new crimes, just new ways to calculate revenue

Every decision a company makes is a financial decision. So finance isn't just the function that reports on the business; it's where the business gets decided.

Which makes it worth asking where your finance team's hours actually go. Reconciling last month. Chasing receipts. Policing a $30 subscription. All of it backward-looking.

That's why I use Brex: corporate cards plus AI-native workflows that absorb the record-keeping. Zero-touch receipt matching, real-time policy enforcement, continuous reconciliation. Controls fire before money leaves the account. Close takes days, not weeks.

Defense handled. Go play offense. Thousands of companies — including Anthropic, Coinbase, and DoorDash — already run on Brex. Head to brex.com/metrics to join them.

The Mostly Metrics world tour continues 🌎

RSVP below, rumor is Walter our editor in chief may make an appearance

Join us in London on October 15th and Boston on October 20th

Welcome back to the newsletter where we talk about how companies talk about their numbers… and also talk about how the public talks about those companies talking about their numbers.

Well, I can’t believe I'm saying this, but I think everyone is completely over reacting to these leaked Anthropic numbers, while simultaneously under reacting to how these new AI companies are calculating their run rate revenue.

We’re focusing on the wrong thing.

TL;DR: Anthropic’s FY 25 numbers got out, showing

Why might this not be as bad as it seems?

1. You really think they aren't going to include 2026 when they file?

We are ten months into the year, people.

These are 2025 numbers, much of which we already knew, or at least sensed. And it excludes the period in which they rocketed from $9B in annualized revenue at the end of last year, to $30B in April, to $65B by the end of July (more on how that’s calculated in a sec).

That’s the fastest growing period in the history of fast growing companies… ever!

Yea, a lot of this is "not good" but it's also "not complete".

2. The $42 billion number overstates what they lost actually running the biz.

It's a shitty headline number to lead with, meant to scare people.

Yes, AI might kill us all (minor point), but that’s not the operating loss from a functional business perspective.

Because $34 billion was a noncash revaluation of financing that may convert into shares.

OK, what's that mean?

It's due to financing choices, likely convertible instruments that will cause dilution at some point as the value of the company goes up. Which is a big deal from a shareholder perspective; you do care about stock charges and warrants that will decrease the value of shares on a one for one basis. But it's a very different thing than paying people who work there, or buying compute (which they buy a lot of), or settling your AWS bill.

If we're talking about operating losses, you've gotta compare $8 billion to the $4.6 billion in revenue. In other words, it cost them roughly $12.6 billion to generate $4.6 billion in topline last year.

Again, not “good” from a unit economics perspective! But if you're assessing the costs of actually running Anthropic, look at the operating loss. If you're assessing what existing shareholders ultimately own, account for the convertibles too.

3. What this data seems to corroborate is some fugazi around the revenue numbers.

Around the same time these numbers came out, a fascinating interview dropped where Harry Stebbings spoke with Higgsfield founder and CEO Alex Mashrabov. Alex made a comment, almost in passing, that they calculate annual revenue "the same way OpenAI and Anthropic do."

They take the last 28 days of revenue and multiply by 13, which gets you to 364 days.

(Because that’s totally a standard Roman Gregorian calendar cut.)

When I heard that, it made me go back and look at Anthropic's reported revenue numbers over time. And it seems to check out.

Anthropic reported $4.6B of actual revenue for 2025. At the end of the same year, we were talking about them being at roughly $9B of annualized revenue. A bit of a gappppp.

One view tells you how much revenue you actually generated over the year. The other looks at what you did over the last 28 days and says, "OK, what if we did that (looking for random number to use) 13 times?" This view is an expression of your theoretical capacity.

(I like to sprint a mile as fast as I can, multiply by 13, and brag to people that’s my half marathon time)

And given how fast these companies are growing, those numbers are going to be VERY different. And the 2026 revenue Anthropic eventually reports is def not going to match the $65B we've been quoting in the OpenAI vs Anthropic revenue arms race.

Anthropic‘s 2025 numbers and what we know about 2026 so far did not make me feel like the sky was falling. Before we start reassembling the chairs on the deck of the Titanic, I'd like to see what the 2026 figures print. Are the unit economics getting any better? Because if this isn’t “at scale” then what is?

Hopefully they aren't still selling a dollar for sixty cents… Maybe they’re selling a dollar for 75 cents now?

Time will tell, and that time looks like a late November IPO.

Here's how Higgsfield described the math:

I've written at length about the crimes of CARR, or Contracted Annual Recurring Revenue.

But with each new pricing model we get new revenue crimes. Pricing necessity is the mother of all revenue invention.

I posted this Lead Edge hierarchy last week because it remains one of the greatest pieces of finance literature ever produced.

The premise is that if the number at the top looks like shit, keep moving down the list until you find one that looks good.

I think we need to make an addition.

If companies don't have good TTM revenue, they report the last 28 days of revenue multiplied by 13.

To be fair, this is WAY less criminal than most of the stuff further down the hierarchy. They did not report that they were the 13th best place to work in Topeka, Kansas. At least it's live revenue.

They're also not taking some three-year ramped enterprise contract, grabbing the biggest number exiting year three, and pretending they're servicing all of it today.

You can’t live off CARR. If I fed my dog CARR, he'd die.

Now, I’m no stranger to “run rate revenue.” Back in the day, we'd actually call it that - take your last quarter and multiply it by four.

I like the quarterly version better than monthly (or 13) because you have three months in there to smooth things out. Twenty-eight days is where it starts to get suspect.

When a really good month gets multiplied by 13 you look like a rock star. If you run a big promotion, launch a new product, sign a whale or see a giant spike in consumption, every incremental dollar generated during that window gets 13 little friends when you report your annualized number. Suddenly a good month is a great year.

And the shorter your measurement window gets, the more sensitive the number becomes to whatever weird stuff happened recently.

We've already seen public companies wrestle with this in usage-based models. Some use a month. Some use 90 days. MongoDB has even used different windows depending on the type of customer (enterprise vs self serve).

There's no Revenue Police coming to your house to tell you which one is correct (although I will be running for deputy commissioner in 2028).

But the window matters a lot.

The other fun part is that the math eventually works against you.

If you're growing 10% month over month, using your latest 28 days is awesome. You're constantly throwing away older, smaller revenue periods and replacing them with the newest, biggest one.

Now stop growing, or worse, decline 10%.

That same math you loved on the way up will very efficiently annualize your shitty month on the way down.

I bet you dollars to donuts we start hearing a lot more about 90-day averages when that happens.

There's also a question of what you actually use this number for.

For an investor update where you're trying to show how fast a business is growing, I get it. You almost have the opposite problem with GAAP revenue. A trailing twelve-month figure for a company growing this quickly is stale AF.

But I wouldn't want to run the company exclusively off a 28-day annualized number. You could make some whacky operating decisions… Like locking in compute based on a promotional month or under hiring customer support.

AI companies already have the ability to juice demand through credits, free usage, and pricing. Much of their GTM strategy is giving away compute for free.

And this is really the bigger problem with ARR now. We took a metric that used to mean something fairly specific and tried to stuff every new business model into it, taking creative liberties along the way:

I wrote last year that usage-based pricing had turned ARR into the Wild West.

We are now in Westworld.

There are no new crimes, just new ways to calculate revenue.

And Lead Edge is going to need a bigger piece of paper.

Wishing you trade at a high revenue and EBITDA multiple,

CJ