ARR Doesn't Mean What It Used To – By Simon Wu

Venture investors increasingly distrust headline startup metrics because familiar labels now cover very different realities, especially in AI. ARR may mean recurring revenue, gross merchandise value, signed bookings, or a single strong month multiplied by twelve. Investors want founders to define the calculation and distinguish durable revenue from annualized estimates. CARR is even less reliable: signed contracts and statements of work can fail to convert when customers cancel, delay onboarding, reject milestones, or require delivery the company cannot scale. For usage-based products, consumption should not be annualized until it is predictable; infrastructure companies also need to show how financing and capital requirements affect cash generation.

Other metrics can conceal similar weaknesses. Compute-heavy businesses may report strong gross margins before usage drives costs up, while companies may classify implementation and forward-deployed engineering expenses as operating costs rather than cost of revenue. NRR can look unusually high when small proof-of-concept customers become larger contracts, making comparisons with conventional benchmarks misleading. Outside software, physical-AI autonomy rates measured in controlled pilots and sales pipelines built on nonbinding letters of intent can exaggerate real-world performance and demand. The broader problem is that old benchmarks - growth, retention, and burn - were designed for simpler pricing and steadier businesses. Investors now focus more on durability: whether a metric holds in normal operating conditions, what supports it, and how much can reasonably be extrapolated.