Kerala Rising: From Exporting People to Exporting Ideas · Page 259 of 353

KERALA RISING · FROM EXPORTING PEOPLE TO EXPORTING IDEAS

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