The $700 billion is a haircut on a specific pile, and the pile is what makes it legible. Against the $5.84 trillion of asset-side rate exposure the Dallas Fed attributes to the duration of deposits other than large time deposits [5], the price-sensitivity case removes roughly 12% [1]. The separate weighted-average-life case removes about 9.9% [2], which is exactly what the arithmetic demands: the paper approximates effective deposit duration as weighted average life multiplied by one minus the deposit beta [3], so trimming life by a tenth trims duration by a tenth.
The beta case bites harder than proportionally, and that leaks a parameter the write-up does not print. Read the shock as a 10% proportional increase in beta, and the decline implies a starting beta somewhere around 0.5 to 0.55, depending on whether you measure the $700 billion against the deposit-supported $5.84 trillion or the $7 trillion total [5]. In other words, the model already assumes banks pass through about half of any rate move. The scenario is not a world where sticky deposits stop being sticky; it is a world where a moderately responsive deposit book becomes slightly more responsive. That is the modest end of what agentic account management could do [11], and it still costs $700 billion of ten-year-equivalent capacity [6].
Scale matters here too. About 83% of the modelled asset-side rate exposure rides on the duration characteristics of ordinary deposits [6]. The two shocks are presented separately, but if both landed, the combined $1.28 trillion is close to 22% of that base [4].
What the paper will not say is where the loss lands, because the credit effect depends on how banks adjust assets and funding [8]. The two adjustment routes point at different victims. Issue more term debt and lending composition stays closer to unchanged, but the Dallas Fed says wholesale funding would likely raise borrowing costs for consumers and businesses [9]. Hold more reserves and Treasuries against faster, less predictable outflows and the composition changes instead, leaving less room for illiquid credit [10]. Either way the depositor with an AI agent chasing yield does not pay for the switching; the borrower who cannot shop the loan does.
The empirical support is one step removed. A 2025 Central Bank of Brazil paper found heavier use of the Pix instant-payment system raised liquid-asset holdings and reduced liquidity transformation, and the source is explicit that Pix is not a direct comparison with US tokenized deposits [13]. On the US side, the analysis leans on an initiative that the source dates to June 2026, when The Clearing House announced work to develop 24/7 interoperable tokenized commercial-bank money including automated and agentic-commerce uses [12]. An initiative to develop is not volume. The authors themselves say tokenized deposits remain early, the magnitude is uncertain, and their views should not be attributed to the Dallas Fed or the Federal Reserve System [14].
