The Problem With ARR: AI Didn't Change What Growth Is Worth. It Changed What Growth Means.

Somewhere in the last three years, an entire asset class kept using a word after the thing it described had largely stopped existing.

The word is ARR. The case against it is not that the number gets inflated. Inflation would be a manageable problem, because a level problem can be discounted for. The real issue is that ARR now covers four or five different economic phenomena that behave nothing alike, and the industry compares them to each other anyway, ranks them, sets benchmarks off them, and prices rounds against the result.

What the number was built to do

Annual recurring revenue became standard around 2010, when software moved from upfront licenses to subscriptions. Sean Barrett, founder and managing partner at Counter Global, told Inc. in April that investors wanted shorthand for the annualized value of recurring revenue that had not yet shown up in the financials. It was a good solution to a real problem, because in that world the word "recurring" carried weight. A contract renewed or it did not. Churn was an event with a date attached. The annualization rested on a legal document.

Consider what the same three letters now have to cover.

Consumption pricing, where revenue tracks token spend and can halve in a month without a cancellation or a churn event. Enterprise pilot budgets, or what Greenfield Partners' Itay Inbar calls experimental budgets: money that exists specifically because a CIO wants to find out whether the category is real. Free credits handed out at volume, which Upfront Ventures' Zhang, quoted in the same Inc. piece, has flagged as badly clouding demand signals. And run rate math, where a company takes its best month, multiplies by twelve, and reports the product as though it were a subscription base.

Those four things carry wildly different durability profiles. In the reported number they look identical.

The benchmark nobody will decompose

Meanwhile the bar keeps climbing, and it climbs in a way that gets quoted everywhere. ICONIQ's Enterprise Five scorecard puts top quartile growth for companies between $1 million and $10 million in ARR at 515% year over year in its 2025 edition, up from 485% the year before, and the firm expects the trend to continue as AI native companies scale. SaaStr's read of what growth funds are seeing puts the trip from $1M to $100M at 8 to 11 quarters for AI native companies, against a prior top quartile benchmark of 19 to 20. Stripe's 2025 annual report, as summarized by Lobster Capital, counted twice as many startups reaching $10 million ARR within three months as the year before.

So here is the question the industry has agreed not to ask. How much of that increase is companies actually growing faster, and how much of it is companies counting differently?

Nobody knows. Not within a range anyone would defend in a partner meeting. The composition of the numerator shifted at the same moment the numerator got bigger, and none of the major published benchmarks control for that. The 515% figure measures a population whose reporting conventions moved while the measurement was being taken.

Plenty of it is real. Software really does distribute faster now, time to first dollar really did collapse, and the strongest companies in this cohort are building things that will still be around in a decade. None of that is in dispute. What is in dispute is whether one blended percentile can tell you which companies those are, and the industry currently behaves as though it can, mostly because the alternative involves more work.

The tell

If the growth were what it appears to be, retention would look roughly normal. It does not.

A ChartMogul study covering 3,500 software companies, cited in Forbes in April, found median gross revenue retention of 40% among AI native companies as of 2025, with net revenue retention at 48%. The median company in the fastest growing software cohort on record loses roughly six of every ten revenue dollars it starts the year holding.

Those two facts cannot both describe durable businesses. They can both be true of a population where a large share of reported revenue is experimental, consumption linked, credit subsidized, or annualized off a spike. That reading is the simplest one available, and also the one nobody wants.

Why the industry uses it anyway

The interesting question is not whether investors understand this. Most of the ones quoted in the trade press understand it perfectly well, and several will say so on the record in the same interview where they cite an ARR figure. Inbar's own framing is that what his firm underwrites is the go to market motion behind the growth rather than the growth itself, and that founders who cannot explain the difference lose credibility quickly.

They keep using the number because it is liquid. It compares across companies, it fits in an email, it survives an investment committee, and it gives a partner something specific to tell an LP. A cohort table does none of that. The cost of being wrong also runs in one direction. A fund that underwrote a shaky ARR figure and lost money has plenty of company. A fund that passed on the category leader because it wanted three more quarters of retention data will be hearing about that decision for a decade. When the professional risk of skepticism exceeds the professional risk of credulity, you get the behavior we currently have.

This is an incentive problem rather than an intelligence problem, but it should be named accurately instead of dressed up as a new growth paradigm.

What the more careful investors are doing

The better end of the market has already moved, and where it moved is worth noting.

Net revenue retention above 120% has become the working Series A threshold for AI companies, according to an analysis of Carta and CRV data by Value Add VC, meaning the existing customer base grows on its own before a single new logo lands. The same analysis notes that investors increasingly ask for cohort level detail rather than blended averages, because a company at $4 million ARR with 90% NRR is backfilling churn with new logos, and a cohort table makes that obvious in about ninety seconds. Gross margin expectations have been rebuilt from scratch around inference costs sitting in COGS: ICONIQ's January 2026 State of AI snapshot projects average AI product gross margins near 52% this year, well below the 75% that used to define a healthy software business.

Cowboy Ventures' Aileen Lee, speaking at TechCrunch Disrupt in November, described modern evaluation as an algorithm with different variables and different coefficients, weighing data generation, moat, founder history and technical depth differently depending on the company. That makes for a much worse headline than "grow 5x." It is also the only description of the job that holds up against the retention data.

The thing that did not change

Growth is worth what it has always been worth. A company that acquires customers who stay and spend more remains the best asset in private markets, and AI did nothing to that.

What AI changed is that the word stopped pointing at one thing. It now points at a bundle: partly the real asset, partly a CIO's curiosity budget, partly a good month multiplied by twelve. The industry is still pricing the bundle as though it were the asset.

The fix is not a better benchmark. It is a willingness to ask what sits inside a number before repeating it, which is slow and unglamorous and also the job. Some funds are doing it. The rest will learn what kind of revenue they bought the same way everyone else does, at renewal.