ARR loses sheen as AI startups shift to usage-led metrics

The trigger came in February when Emergent AI, an eight-month-old vibe coding startup, said it had crossed a $100 million annual recurring revenue, scaling from $15 million in three months to $50 million in seven months, before doubling within a month. The pace and the disclosure drew scepticism from founders and executives on X, who questioned both the legitimacy of the numbers and the methodology behind them. Founder Mukund Jha later shared a snapshot of weekly revenue to defend the claims.

The episode has since widened into a broader discussion on whether ARR– long the cornerstone metric in the SaaS era — remains relevant for AI startups where revenues are increasingly tied to usage rather than fixed subscriptions.

Death of Predictability

Traditionally, ARR, was anchored in predictable subscription contracts. “There was a consistency and normally these contracts stood for 12 months or at times 36 months or even longer. This is why the metric used to be called contracted annualised recurring revenue (CARR). But software is now based on usage,” said Alok Goyal, investor at Stellaris Venture Partners.

That shift is pronounced in AI-native products, where pricing is linked to token consumption or even outcomes. “With AI tools, the output depends on the prompt such that each prompt generates a different result. You have to spend more on tokens to obtain an output of a higher quality,” said an investor at an AI-focused VC firm. “Token consumption represents the margins. The more tokens a platform consumes the worse their margins will be. For example, Anthropic has the priciest tools and therefore also the worst margins.”

Lightspeed partner Hemant Mohapatra, an investor in Emergent, defended the emphasis on usage. “Token consumption is all that matters in AI,” he said. Tokens are the smallest units of text processed by AI models, covering words, parts of words or punctuation, and form the basis for how usage, costs and revenues are measured in AI systems.

New Scorecard

Others maintain that usage alone is an incomplete proxy. “Token consumption is directionally useful but not sufficient. It tells you about activity, not value delivered,” said Abhishek Srivastava, general partner at Kae Capital. “A customer could be burning tokens on failed queries. What you want to measure is outcomes per token.” He added that annualising trailing 90-day consumption alongside cohort-level retention could better capture both scale and durability.

Goyal said Stellaris evaluates delivered revenue, or invoiced income, rather than extrapolated ARR.

The lack of consensus reflects the early stage of AI business models, even as global firms such as Anthropic, OpenAI, Databricks and TCS highlight ARR to signal growth and justify investments. Investors say this has also created pressure on startups to showcase rapid revenue expansion.

“ARR is indicative of the quality of the product and the value that the customer is deriving,” said Pankaj Agarwal of Prime Venture Partners, but added that metrics such as margins and retention are equally critical.

Srivastava pointed to expansion revenue within accounts and time-to-value as more meaningful indicators, along with gross margins. “A high-ARR AI company with 40% gross margins is very different from a SaaS company at 80%,” he said.

For now, ARR remains widely used, partly due to ease of comparison. But investors expect a shift. “We haven’t matured in calculating metrics like churn or LTV/CAC using token costs,” said Ravi Metta of BAT VC. “ARR is still widely used but not ideal. AI startups deliver higher-value, customised outputs, which should demand a different framework.”

As AI business models evolve, investors say the focus will move from headline revenue run-rates to a closer examination of value delivered, cost structures and durability of demand.