KERALA RISING · FROM EXPORTING PEOPLE TO EXPORTING IDEAS This third table is the one that anchors the whole financial argument in arithmetic. At a million verified actions a month, a very large program, the routine classifier bill is a few thousand rupees; even escalating a tenth of those to a frontier model lands at one to two lakh rupees a year. Scale to several million actions a month across the full portfolio and the figure is still single-digit lakhs. There is no volume, within any realistic horizon for a state of Kerala’s size, at which the AI becomes a material line item. That is not optimism; it is the unit economics of inference multiplied out, and a reviewer can reproduce it from the per-inference figures in the References. On the direction of the assumptions A reviewer is right to ask which way the assumptions in these tables lean, because a model can be made to say anything by choosing its inputs. The plain answer is that the tables lean conservative on cost and conservative on revenue, in opposite directions, so that the central conclusion is robust rather than flattering. On cost, the inference figures assume a generous escalation rate to the expensive frontier model and ignore the steep annual price decline that will, in reality, reduce the bill every year; the true inference cost is therefore likely to be lower than shown. On revenue, the ramp assumes cautious sign-on and conservative per-unit figures, 259