How AI Startups Hallucinate Their Revenue Metrics

Some revenue is more recurring than others

Good afternoon from St. Andrews, Scotland!

On 17th April 2026, Scott Stevenson, the founder of Spellbook, called Contracted ARR “a huge scam in AI startups.” He’s frustrated because some companies are taking future, back‑loaded contracts and reporting them as today’s recurring revenue.
That turns a probabilistic stream of cash into a hard number, it’s akin to an accounting version of an AI hallucination. In this piece, I’ll define ARR and CARR properly, show how they’re being abused in the current AI boom, and suggest a few questions investors should ask before they underwrite $100mm ARR.

Are founders (and investors) obfuscating the truth about their startups’ ARR numbers or do they just not understand the difference between CARR and ARR? Or is it both?

Contracted ARR is a scam | @scottastevenson (X)

What said AI startups are doing is that they’re presenting optimistic, back-loaded deals as today’s recurring revenue, and that can make growth look much stronger than the cash actually being collected.

Models hallucinate by turning probabilities into faux facts; AI startups hallucinate ARR by turning probabilistic future cash into today’s revenue.

Whilst this is happening, a lot of AI startups have been “hallucinating” their revenue. Whether that comes from CEOs of AI startups lying or not understanding the difference between how revenue is recorded on a GAAP or IFRS basis, ARR, run-rate revenue, contracted ARR (CARR), bookings, and billings.

ARR is calculated as the following:

Last Year’s Subscription Revenue + New Subscription Revenue + Expansion Subscription Revenue – Churned Subscription Revenue = Current ARR

The above formula is the recommended version as it considers both expansion and churn during the year, and is based on last year’s subscription revenue numbers. You can also do the following:

\(ARR = MRR ✕ 12\)

which will give you ARR if you have monthly recurring revenue (MRR) already.

Whilst it’s not “wrong” to annualize monthly recurring revenue via multiplying current MRR by 12, investors do not treat all monthly revenue as equally durable. When revenue has no minimum term commitment or elevated churn, investors may discount it because predictability is lower. The issue is not the billing cadence, but whether historical retention supports treating that revenue as truly recurring.

This matters because it gives founders a lot of room to choose the version of ARR that flatters them most.

Another thing that adds to the confusion (highlighted by Artificial Lawyer, covering the legal tech space) is that the definition of ARR can also potentially include contracted future income that the startup hasn’t received yet - this is the first problem. I disagree with including future income and would rather define it as the total predictable subscription-based revenue a company expects to earn each calendar year - this is the second problem, the definition of ARR isn’t universally agreed upon, so there’ll be some folks who’ll include future income as part of their ARR calculations.

ARR is meant to show the annualized value of recurring subscription contracts. Let’s say a startup invoices $100k in revenue in January, its ARR for the current year would be $1.2 million. That’s what it’s supposed to be.

CARR goes further. If the same customer has signed a contract that ramps from $100k this year to $300k in year three, CARR might count the $300k from day one, even though the company is only earning a third of that today.

What this does is that it creates a situation where you have multiple revenue metrics like GAAP revenue, run-rate revenue, ARR and then CARR telling people different things despite revenue being the commonality. As a result it creates performance issues, and what’s more is that gullible media outlets take the numbers they hear at face value and treat it as gospel without really questioning anything. Anyone - whether it’s a public or private company - that doesn’t report and recognise revenue on a GAAP/IFRS basis is screwing investors over.

For a sale to be counted towards ARR, it should:

• correspond to an annual contract or, even better, a multi-year contract without early cancellation options. If one is going to annualize monthly contracts, they should explicitly mention that.

• have fixed pricing set in advance based on users or company size, with the “Annualized” number based on the initial pricing, not the eventual or possible future pricing. Usage-based revenue should not be part of this because it is non-recurring.

• be based on a reasonable length of time, such as one month or one quarter, rather than one day or one hour during The Big Sale.

• not correspond to usage-based revenue, token revenue from marking up other providers’ fees, services, implementation fees, license fees, royalties, or transaction/processing fees.

In my post titled ‘Who Captures the Value If AI Inference Becomes Cheap?’, I stated:

Proprietary platforms are at extremely high risk of being ‘enshittified’. The ‘enshittification’ inculdes but is not limited to the implementation of adverts. A brief look and you’ll find endless threads already alleging this of OpenAl, Claude and Gemini. It’s almost guaranteed to happen as they seek to claw back their investment by monetising LLMs.

I still stand by this today. It’ll look something like this:

1. acquire funding from VC + other investors to subsidise LLM and token costs

2. give tons of services for free (and usually at a loss) to attract tons of users

3. raise prices overtime and/or move to some version of usage-based pricing (and implement ads), cut back free use via rate limits, reduce quality of services. And on top of that, count the “eventual” price towards the current ARR + claim higher growth rates based on future expected revenue that may or may not arrive.

Beyond the obvious issue of customers cancelling contracts and subscriptions, it is a problem because churn rates are higher in the AI space, and customers are usually trialling these products before they commit to anything long-term (unlike enterprise SaaS due to higher switching costs and deeper workflow integration). There’s a lack of stickiness but startups using CARR ignore that fact.

The concept of ARR being used by startups selling AI based on unprofitable subscriptions, token usage or other services doesn’t make sense to me. This implies a lack of durability in revenues. From the little I know, there’s not a single example of sustained, positive margin expansion and impact of AI inside a true corporate enterprise that’s not a small test or some pilot project. When you have thousands of companies paying $200 per month, it’s not hard to show up with revenue in the billions. On the software front (with Databricks and Snowflake), if you look at the companies that use that software, those companies generate enormous revenues and enormous margins. And these products are in critical production workflows that underlie those revenues and profits. That is just not true with AI today.

Since we’re talking about private companies, the actual numbers are hard to track beyond the little that’s publicly reported. On May 7th, FT Alphaville also highlighted how “other income” is boosting Big Tech earnings. The “other income” here is ASC 321 fair value movements - unrealised gains on minority stakes that must, under US GAAP, run through the P&L rather than on the balance sheet. Nothing improper, but entirely non-cash, and the circularity is entertaining: hyperscalers fund Anthropic, Anthropic’s valuation doubles, hyperscalers book income.¹ Public Big Tech is doing the mark‑to‑model version of the same trick AI startups do with ARR.

A significant concern is that many AI startups have adopted the metric ‘Contracted Annualized Recurring Revenue’ (CARR), which incorporates future upsells and expansions that customers have agreed to, even if these amounts have not yet been recognized as revenue.

CARR = Annualized Recurring Revenue (ARR) + future bookings, expansions, and downgrades agreed to within the current period but not yet recorded as revenue.

In some contexts, CARR can be acceptable, especially if it’s for an established company and the “future events” are only for the next month or quarter.

However, many AI startups have assumed that customers paying at heavily discounted “Year 1” contract rates will continue paying without cancelling and will agree to the much higher Year 3 rates. Then, they base their “ARR” numbers on these Year 3 contract values, even if the contracts are cancellable well before Year 3.

Source: BIWS

Just because it’s technically allowed, it doesn’t mean that startups should always annualize monthly contracts and call it ARR. If a company is going to do that, they should be clear about it.

Nikunj Kothari of FPV Ventures spoke on the (mis)use of ARR in December 2025, where he called ARR a “liar’s valuation”. In the article, he uses examples of a med-tech company leasing out a machine to a hospital for MRI scans and also a fintech company - both of which get revenues mainly from transactions/one-time uses. There’s nothing recurrent about a transaction that happens less than once a year especially in the med-tech case; unless something is severely wrong with your health, there’s no reason paying for an MRI scan should be recurring. In fintech, most of the revenue composition for some startups is at 80% card swipes and payment processing. It’s far too volatile because if consumer and/or business sentiment weakens, people will spend less meaning revenue from transaction fees decreases.

He goes on to say…

A vibe coding platform sells you a seat for $20 per month. That’s subscription revenue. You also pay for usage: tokens, compute, API calls. That’s consumption revenue. These two types of revenue have completely different economics, but companies blend them together.

Here’s the math. Say a company has 1,000 customers paying $20 per month for seats. That’s $20,000 in monthly subscription revenue, or $240,000 annually. Those same customers collectively spend $200,000 per month on usage, which is $2.4M annually. Total annual revenue is $2.64M.

The company reports “$2.6M ARR” as a single number.

But the subscription piece and the usage piece are fundamentally different. The subscription revenue is predictable with decent margins and probably good retention. The usage revenue swings wildly month to month, and margins are often brutal because you’re paying OpenAI or Anthropic for API calls and marking them up 20%.

Blending them into one ARR number hides everything important. A business that’s 90% usage-based gets compared to a business that’s 90% subscription-based as if they’re equivalent. They’re not.

Then he also speaks on ‘creative counting’:

Free trials counted as ARR by assuming they’ll all convert. They won’t.

Monthly customers multiplied by 12 as “ARR” even though they signed month-to-month because they’re not sure about the product yet. That’s not annual recurring revenue. That’s monthly-maybe-recurring revenue.

Pilots counted as ARR before conversion. Verbal commitments treated as signed contracts. “Pipeline ARR” presented as if it were actual revenue.

Imagine a startup telling a new employee it’s at “$100mm in ARR” and growing by 3x every year. That headline is doing some heavy lifting. It’s there to defend a punchy valuation, which in turn is used to justify the equity grant and the risk of walking away from a safer job.

Under the hood, that $100mm might be $65mm of low‑margin pass‑through transactions that disappear if one big customer churns, $25mm of usage revenue that swings ~30-40% month to month, and maybe $10mm of actual contracted subscriptions. The company reports one clean number, but the cash flows underneath behave like three different businesses. Someone’s equity will be priced based off the headline, not the mix that will decide whether there’s anything left for common.

The confusion of ARR in AI startups also happens with public companies too.

Verint’s July 2025 earnings report offers a clear example of SaaS ARR disclosure for public AI companies. The company reports total ARR and breaks out AI ARR, which it defines as recurring revenue from its AI-driven cloud solutions. Non-AI ARR includes the rest of its software and services subscriptions. Verint reports about $372mm in AI ARR, which is growing at over 20% year over year, and aims for AI ARR to make up most of its total ARR in the medium term. From a business perspective, this is still standard subscription ARR: multi-year contracts, high renewal rates, and little usage volatility. The “AI” label simply describes the product mix rather than changing how the metric is calculated.

Adobe does not label any revenue as “AI ARR,” but its Digital Media ARR, which includes Creative Cloud and Document Cloud subscriptions, is where generative AI is monetized. In its 2024 annual report, Adobe reported Digital Media ARR of about $17.33bn as of November 29, 2024, an increase of $2bn (13%) from the previous year. This figure represents the annualized recurring revenue of Creative Cloud and Document Cloud subscriptions at the end of the period. Financial press coverage of FY25 notes that Digital Media ARR growth, at around 11–12% year over year, was primarily driven by demand for AI-enhanced Creative Cloud and Acrobat services. Firefly and Acrobat AI Assistant together added over $250mm in new AI-focused ARR. These results are also influenced by price increases and higher-tier bundles that now include Firefly and other generative AI features. The revenue from AI is included in the overall subscription base, not reported separately. From an ARR vs CARR perspective, Adobe serves as a control group, showing annualised subscription revenue from well-established tools, with AI providing an added boost rather than prompting a reclassification of backlog or GMV as “AI ARR.”

C3.ai have shifted from using classic term subscriptions toward a Snowflake-style consumption model which muddies ARR. According to its SEC filings, C3.ai splits GAAP revenue into subscription revenue and professional services. Additionally, they disclosed remaining performance obligations (RPO) - the contracted but unrecognised portion of subscription deals - rather than leaning heavily on a single ARR number. Customers will buy C3 “units” tied to compute / storage and draw down usage over time, Snowflake do exactly the same thing albeit for software. This reflects the reality that annualizing variable vCPU‑hour usage at a point in time can overstate durability of revenues. C3.ai’s ARR leans closer to “annualised run-rate” than textbook ARR.

SoundHound AI’s investor relations defines its “subscription revenue” as follows:

SoundHound AI Investor Presentation

Me after looking at how SoundHound AI reports ARR

Pillar 2 is what I’m focusing on. I think there’s a misunderstanding here. “Revenue from usage-based fees” and “Revenue per query” aren’t recurring, so they don’t really count as subscriptions.

If we tried to value SoundHound by using different revenue multiples for each income stream, it would be hard to feel confident. The company’s data doesn’t seem to be clearly classified, which makes things tricky.

This is moreso a case of misunderstanding subscriptions than it is a cARR vs ARR issue.

Revenue from usage fees and queries is still worth something, but it is worth less than revenue from locked-in subscriptions that will take at least a year to deliver.

As the textbook-defined purpose of ARR is to approximate a SaaS company’s subscription revenue over the next yearbased on its most recent month or quarter, this gives people a better sense of current growth rates and its momentum. Some implicit assumptions will be baked into those numbers:

1. High gross margins - Since the costs of delivering and servicing enterprise SaaS are low, the Gross Margin of these products can be very high (often 70%+). The challenge is in retaining customers. While customer support and IT/infrastructure costs are present, they are far lower than those incurred by most AI services.

2. High and predictable retention rates - Enterprise SaaS products have traditionally had churn rates of ~10% per year or less (very low), which makes ARR meaningful since it’s more likely to persist in each period. But with most AI startups, it’s “easy come, easy go”: High growth, but also high cancellation rates because it’s so easy to switch.

For those analysing companies that sell mainly AI-based products/services, it’s better to focus on GAAP/IFRS revenue or run-rate revenue as opposed to ARR and be more conservative with multiples due to the aforementioned issues.

Run‑rate revenue is decent ‘back‑of‑the‑napkin’ math for how fast something’s growing, but multiplying one noisy month by twelve and calling it “next year’s revenue” is wishful thinking when customers can walk away at 30 days’ notice. If the business is mostly clean subscriptions with good margins and low churn, then ARR can be a fair shorthand but you still have to look under the hood and assess those numbers.²

For traditional subscription companies, people should be asking: What percentage is annual contracts versus monthly? What’s the gross + net revenue retention and the gross margin? This is why I said the first ARR formula is better than the second ARR formula as you can more easily determine how “sticky” the product is, how much new customer acquisition will be required for growth, and what the eventual cash flows might be.

Hybrid companies will be slightly different but it’s still worth asking questions such as: How is ARR calculated? What’s the composition of subscription versus usage-based revenue? What are the margins by revenue types? What does retention look like in 12 months cohorts, not 3 months cohorts?

Every metric can be gamed at the expense of one or another. ARR in isolation is basically Goodhart’s Law at work: "When a measure becomes a target, it ceases to be a good measure." You use ARR in conjunction with gross margin, churn, along with customer count and ACV. Nobody should look at ARR alone. This combination should allow you to reverse out any revenue data you need and then tie that to expenses and you have everything you need.

If you lie or manipulate ARR when raising money, that’s called fraud. Don’t do that. In the end many things require a quick assessment and almost any metric can be gamed, manipulated, or flat out lied about. That’s the road to fraud and it will end badly.

Raising money is a step along the path (unless you are just stealing it) and you actually have to build value to get any type of exit. So manipulating financial data is in no one’s best interests.

This isn’t a silver bullet to fixing revenue hallucinations, but it does help you make it eaiser to find the companies that are worth paying attention to.

Thanks for reading!

1. The cleaner number is operating cash flow minus capex, with the equity gains added back - which is what the cash flow reconciliation does anyway.

2. Run Rate = This months revenue 𝗑 12