The AI funding machine runs on momentum, and momentum runs on numbers. But what happens when the numbers don’t mean what they used to mean?

Some AI startups are stretching the definition of annual recurring revenue beyond recognition, and their investors know exactly what’s happening. ARR used to be a straightforward metric: take your monthly recurring revenue, multiply by 12, and you’ve got a number that tells investors how much predictable income you’re generating. Clean, comparable, useful.

Not anymore.

The new playbook works like this: a startup signs a deal where a customer commits to $1 million in API credits over the next year. There’s no recurring subscription, no guarantee the customer will renew, and usage might taper off after the first few months. But the startup books it as $1 million in ARR anyway. Or they count pilot programs that are explicitly time-limited. Or they roll up a bunch of one-time implementation fees and enterprise services into the recurring revenue bucket.

This isn’t creative accounting in the illegal sense. It’s creative accounting in the “everyone involved understands what’s actually happening but chooses to describe it differently” sense. The founders do it because inflated ARR helps them raise at higher valuations. The VCs accept it because they need portfolio companies with impressive growth trajectories to raise their own next funds. And the later-stage investors go along with it because everyone else is doing the same thing and you can’t afford to sit out the AI wave over accounting principles.

The problem is that these numbers are becoming the foundation for real business decisions. When a startup says it’s doing $10 million in ARR, that figure determines how much runway they have, what kind of valuation they can command, and whether they’re on track to IPO. If half that revenue is actually usage-based consumption that could drop to zero next quarter, those decisions are built on sand.

Why this matters now

We’re watching this play out in real time with SpaceX’s IPO filing. The company is targeting a valuation that would make it the largest public offering in American history, with a total addressable market estimate of $28 trillion and risk factors that run 36 pages. That kind of ambition requires numbers that tell a very specific story, and the market is starting to ask harder questions about what those numbers actually represent.

The SpaceX filing is notable for what it includes: a pay package tied to establishing a Mars colony, an explicit acknowledgment that the business model depends on things that haven’t been built yet, and projections that require a certain amount of faith. That’s honest, at least. The risk is spelled out in writing.

The AI revenue game is less transparent. When companies talk about ARR in pitch decks and press releases, they’re usually not including footnotes about how much of that revenue is actually recurring, how much is consumption-based, or how much depends on pilots that might not convert. The number just sits there, looking solid and comparable to every other ARR figure in the market, even though it might mean something completely different.

The downstream effects

This matters because it distorts the entire market. If one startup counts pilot programs as ARR and raises at a $500 million valuation, every other startup in that space needs to do the same thing or risk looking like they’re losing. Investors start making decisions based on metrics that can’t actually be compared. Employees join companies thinking they’re on rocket ships when the revenue might evaporate in six months.

And when the market eventually corrects, which it always does, the companies that played the game hardest will have the most painful reckonings. Revenue that was never really recurring will disappear. Valuations that were never really justified will collapse. Investors who knew the numbers were inflated but chose to look the other way will claim they were deceived.

We’ve seen this before. During the last tech bubble, companies counted barter deals as revenue. During the SaaS boom, some startups front-loaded annual contracts and called it recurring. The playbook changes but the incentives stay the same: show growth at any cost, worry about sustainability later.

The difference this time is that AI is moving faster and the amounts involved are larger. When companies are raising $100 million seed rounds and achieving billion-dollar valuations before they have a finished product, the stakes for getting the metrics right are higher than ever.

What comes next

The market will eventually force clarity. Public investors don’t tolerate creative revenue definitions the way private investors do, which is why companies have to clean up their metrics before IPO. But by the time that happens, a lot of capital will have been allocated based on numbers that didn’t mean what people thought they meant.

The smarter move would be for someone, anyone, in the AI ecosystem to start demanding real definitions. What exactly are you counting as ARR? How much of it is genuinely recurring? What’s the actual retention rate once pilots end? Those questions aren’t hard to answer, but they require everyone to agree to stop playing the game at the same time.

Until that happens, expect more stories about AI startups with explosive growth that turns out to be less explosive once you read the fine print. And expect more investors who claim they were shocked, shocked to discover that the metrics they funded were inflated, even though everyone in the room knew exactly what was happening.