Kerala Rising: From Exporting People to Exporting Ideas · Page 88 of 353

KERALA RISING · FROM EXPORTING PEOPLE TO EXPORTING IDEAS

KERALA RISING · FROM EXPORTING PEOPLE TO EXPORTING IDEAS
inference on commodity infrastructure, which is
about two-hundredths of one Indian paisa. A
general-purpose frontier vision model, reserved for
hard or contested cases, costs on the order of
US$0.002 per image, a little under two rupees. And
the trend is steep: equivalent the system capability
has been getting roughly ten times cheaper each
year. A program designed today will run on cheaper
inference every year it operates.
[3]
[4]
[5]
Put that against volume. Suppose a mature program
verifies one million citizen actions a month; a very
large number, far beyond early-stage activity. If
every single one were checked by the routine
classifier, the monthly inference bill would be on the
order of a few hundred rupees. Even if a tenth of
them were escalated to a frontier model for a harder
judgement, the monthly bill would be in the low
lakhs of rupees at most. These are amounts that
disappear inside the rounding of a government
budget line.
To see how unusual this is, contrast it with how AI
costs are usually discussed. The reason enterprises
worry about inference bills is that they run frontier
models, with long prompts and long outputs, on
every request, and chain many such requests
together, a workload whose per-call cost is cents to
tens of cents and whose volume is enormous. This
program is the opposite workload: a single, short,
image-classification call per civic action, on the
88