Validiti presents
The token math

Two ways to let AI answer over your data.
Only one of them scales.

One way pours every record into the model — the bill grows with your data, and it stops working the moment your data won’t fit in the model’s context window. The other sends only the question; your data stays where it is, and the answer comes back computed over every record. Here is the difference, measured on real records.

483
tokens to answer — flat, whether you hold a thousand records or a hundred million
152,497×
more tokens the other way, at a million records — and climbing with every record you add
< 0.1s
to answer over a million records — every one of them, not a sample

Tokens to answer — across the size of your data

Your AI writes the query Read all your data into the AI a model’s context window
100 1K 10K 100K 1M 10M 100M 1B 10B 1K 10K 100K 1M 10M 100M RECORDS IN YOUR DATA → TOKENS TO ANSWER → 200K-token context (e.g. today's frontier chat models) 1M-token context (e.g. the largest windows) above a line: too big to send in one call — must be split into many Read all your data into the AI grows with every record you own Your AI writes the query — 483 tokens, flat the same, no matter how big your data is

The same, in numbers

Your dataRead it all inWrite the queryDifferenceCost / question*Fits a 200K-token model?
1,00074,079483153×$0.2222fits
100,0007,365,97948315,250×$22.10no — ~37 calls
1,000,00073,655,979483152,497×$220.97no — ~369 calls
100,000,0007,365,555,97948315,249,598×$22,096.67no — ~36828 calls

Reading it all in crosses a 200K-token window at just ~2,715 records, and even a 1M-token window at ~13,575 records. Past that it can’t run as one request at all — it’s split into hundreds or thousands of calls that each see only a slice. Writing the query never moves off 483 tokens, and gets cheaper still when the same question is asked again.

And the answer comes back in milliseconds

Measured, answering over every record — not a retrieved handful:

100,000 records
10 ms
1,000,000 records
0.10 s
reading that million in, instead
~4.1 hrs*
“Reading all your data burns tokens in proportion to your data, and stops working past the context window. Writing the query burns a flat handful of tokens no matter how big your data is — and the answer comes back over every record, in the time it takes to blink.”

*Measured on a real records dataset. Token counts use a deliberately conservative estimate — real business records tokenize denser than the estimate, so the gap in practice is larger, not smaller. Costs shown at a representative $3 per million input tokens; scale to your model’s price. The read-in time is an estimate from token volume. Every figure is reproducible.