On this pageWhy the average lies
Runway is cash divided by burn. The arithmetic is trivial; the hard part is which burn. For AI companies the honest answer is “next month’s”, and next month looks nothing like the last three averaged together.
Why the average lies#
Inference spend grows with usage, and usage at a healthy AI company grows every week. A trailing three-month average lags that curve by six weeks by construction. When usage is growing 8% a week, a model built on the average overstates runway by roughly two months.
That is not a rounding error. It is the difference between raising from strength and raising from a deadline.
Modelling the slope#
Nebula splits burn into two parts and models them separately:
- Fixed — payroll, rent, committed cloud reservations. Projected flat, with known step changes such as a new hire’s start date.
- Usage-driven — inference, on-demand compute, per-call APIs. Projected from the slope of the last 28 days of daily spend, weighted toward recent days.
The model recalculates every hour from the live ledger, so a new customer’s traffic shows up in runway the day it starts, not at month end.
A range, not a date#
A single runway date implies a precision nobody has. We show a band — the date at the current slope, and the dates if growth halves or doubles — and we tell you which vendor lines drive the spread.
Across accounts with at least six months of history, the central estimate has landed within ±4 days of the actual cash-out date in back-tests.
Acting on it#
The point of a better runway number is an earlier decision. Set an alert for when the low end of the band crosses a threshold — nine months is common — and Nebula will tell you, with the three vendor lines that moved it most.