Gartner's number, as the column relays it, implies its own base: $2.52 trillion at 44% growth means 2025 sat near $1.75 trillion, so this year's increment is roughly $770 billion [2][1]. That increment is the money actually in motion, and the argument worth having is which ledger it lands on.
The token bill wins internal arguments for a reason that has little to do with its size. It arrives metered from a vendor, monthly, and a lower cost per million tokens can be put on one slide. The other bill, what Bijlani calls datanomics, covers discovery across every system data lives in, deduplication, cleaning, privacy masking and the pipeline engineering that has to happen before a model touches the result, and it sits with teams that already have other work [8][11]. It is also defended as money spent so that something does not happen, which is the hardest case to carry through a planning cycle. The legible cost gets managed; the illegible one is absorbed into schedule, and then into the stalled pilots and quietly retired projects he describes [8].
The board-deck version says unit economics are improving because inference keeps getting cheaper. True, and incomplete, because the eras stack rather than succeed one another. An enterprise running agents under a sovereignty mandate is paying the token bill, the agent bill and the data bill at once, and by the column's account most boards track only the first [9]. IBM's figures as cited put an AI-driven incident near $6 million, about a fifth above the $4.99 million average, with the organizations hit disproportionately those that never extended existing access controls to their AI systems [6][2].
A skeptic has an obvious objection: Bijlani leads an AI practice at Hewlett Packard Enterprise, and the remedy he prescribes runs through governance and private cloud work of the kind his employer sells [1]. Worth holding onto. The load-bearing figures are not his, though. The adoption and scaling counts are Stanford's, and the eight-in-ten data-limitation figure is enterprises naming their own roadblock rather than a supplier naming it for them [3][5]. What we have is the column's rendering of those reports rather than the reports, and the supplied text stops partway through the first of four datanomics parameters, so how the other three would be measured is not on this record [11].
None of this forces a reallocation this month, and token prices will keep falling while pilots keep clearing approval. The decade claim is IDC's: that by 2030 half of new economic value from digital businesses in Asia Pacific will come from organizations governing AI capabilities today rather than merely deploying them [7]. The choice in front of a committee this quarter is narrower than that, and it is whether data readiness is priced into a programme's approved cost or left to surface later as slippage nobody budgeted.
