The pitch deck is a familiar sight in the current AI gold rush: a steep, upward-curving line representing Annual Recurring Revenue (ARR) that suggests a company has found a scalable vein of gold. For months, the venture capital community has watched as early-stage AI startups report explosive growth, jumping from zero to tens of millions in revenue in a matter of months. But behind the closed doors of due diligence, a different narrative is emerging. The numbers that look like a rocket ship on a slide are often a composite of optimistic projections, one-time windfalls, and creative accounting that would make a legacy SaaS founder blush.
The Metric Soup of AI Growth
A group of 15 venture capitalists has recently sounded the alarm on the opacity of ARR reporting within the AI sector. The core of the problem is a systemic blurring of definitions. In the traditional SaaS world, ARR is a sacred metric representing predictable, recurring revenue. In the AI world, however, ARR has become a catch-all term for any revenue-like figure. Joseph Floyd of Emergence Capital notes that he has reached a point of skepticism regarding almost every metric presented by AI founders. He points out that what is labeled as ARR is frequently just a run-rate—a snapshot of a single high-performing month or day multiplied by twelve or three hundred and sixty-five—rather than actual recurring contracts.
This lack of precision extends to the very nature of the money being tracked. Niko Bonatsos of Verdict Capital explains that investors now have to perform a forensic audit on every revenue claim. They must determine if the reported figure is actual revenue, Gross Merchandise Volume (GMV), or simply a quarterly figure multiplied by four. The volatility of AI adoption means a company might have a massive spike in usage for one month due to a single large pilot, which is then extrapolated into a multi-million dollar ARR figure that has no basis in future reality. Jeff Weinstein of FJ Labs observes that while the ability of AI companies to scale from zero to 10 or 50 million in a year is unprecedented and exciting, it makes the verification of those numbers a critical necessity rather than a formality.
The Structural Deception of the AI Model
The tension deepens when moving from simple ARR to more complex metrics like Contracted ARR (CARR) and Net Revenue Retention (NRR). The industry is discovering that a signed contract in the AI era is not the same as a guaranteed payment. Kevin Tu of DFJ Growth highlights a specific trend among companies supplying data to frontier AI labs. These firms often present eight-figure Statements of Work (SOW) as proof of demand. However, these documents are often riddled with milestones and customer approval clauses. If the AI lab decides the data quality is insufficient or the project pivot occurs, that eight-figure SOW never converts into actual cash. The gap between a signed letter and a realized payment is wider in AI than in any previous software cycle.
This instability is compounded by the shift toward usage-based pricing. Francisco Gimenez of 8VC argues that the hybrid model—mixing fixed subscriptions with usage-based fees—makes annualization nearly impossible for early-stage firms. Usage revenue is inherently volatile; it fluctuates based on the customer's internal experimentation and the efficiency of the model. When startups annualize this volatility, they create a facade of stability that doesn't exist. The result is a valuation based on a projection of usage that may never materialize as a consistent habit.
Perhaps the most concerning distortion lies in the cost structure. Vivek Krishnamurthy of Commerce Ventures points to the role of Forward Deployed Engineers (FDEs). In many AI startups, the product is not truly plug-and-play; it requires a team of engineers to be embedded with the client to make the system work. These FDE costs are often categorized as operating expenses (OpEx) rather than Cost of Goods Sold (COGS). If these implementation costs were correctly attributed to the revenue they generate, the legendary 70 percent gross margins claimed by many AI firms would collapse. The business model is not a scalable software product, but rather a high-end consultancy disguised as a platform.
This pattern of over-extrapolation is not limited to software. In the realm of Physical AI, Jason Kalira of Westly Group warns that autonomy rates measured in controlled pilot environments rarely survive the transition to the chaos of a real factory floor. Industrial and energy startups frequently lean on Letters of Intent (LOI) or Memorandums of Understanding (MOU) to signal traction. Yet, these are non-binding documents. The real signal, Kalira argues, is not the LOI but the Purchase Order (PO) or a contract backed by a non-refundable deposit.
As the honeymoon phase of the AI boom fades, the heuristics of the SaaS era—growth rates and NRR—are proving insufficient. Shaun Lee of Mubadala Capital suggests that for capital-intensive AI infrastructure firms, looking at run-rates without analyzing cash conversion is a dangerous game. If revenue growth is fueled by massive upfront capital expenditure or equipment financing, the actual economic value accruing to the company is far lower than the top-line growth suggests. The new gold standard for due diligence is shifting toward Free Cash Flow (FCF) and the actual capital required to generate each dollar of revenue.
Ultimately, the industry is moving toward a three-part verification process: confirming the product actually works, measuring exactly how well it works in a non-ideal environment, and determining which parts of the current success can actually be extrapolated into the future. The era of accepting a multiplied run-rate as truth is ending, replaced by a demand for operational transparency.
