ZiffyVolve
Back to insights
The Economics of AI at ScalePart II · The Scale Problem

ZiffyVolve Executive Insight

When Plumbing
Fights Back

As software becomes a source of demand, falling AI unit costs can make far more automated activity economical. The constraint shifts from price toward permission, identity and control.
By Vinay Prathy5 minute read

Demand no longer waits for a person

Part I examined a familiar corporate response to a visible meter: organizations optimize the unit, then shift attention toward what the technology produces. AI adds a complication. Software itself can now consume the meter.

A human request pauses for judgment between cycles. An automated workflow can call a model, invoke a tool, query a system and trigger the next action without waiting for another person. Budgets, rate limits and policy still impose boundaries; human attention no longer paces every unit of demand.

Exhibit 1 · The demand shift

One request ends. One event can keep producing activity

Human attention is no longer the only governor of demand

Demand forecasting therefore cannot stop at users or prompts. It must account for interactions per business event, event frequency and the likelihood that one interaction creates another.

A cheaper unit can produce a larger bill

Once software can create demand, lower unit cost can expand aggregate consumption by making more workflows, greater frequency and deeper call chains economical. A Jevons-style rebound is a useful lens, not a prediction that every efficiency gain will be consumed. The narrower point is sufficient: a unit-cost forecast cannot substitute for a demand forecast.

Exhibit 2 · The scale paradox

A cheaper unit can make a much larger field of activity economical

Direction, not causationThe platform-wide trajectory and customer count describe different scopes. They show magnitude, not that lower unit cost caused the growth.

Unit efficiency and aggregate economics can move in opposite directions

Google’s platform-wide figures establish magnitude, not causation. AT&T offers a separate enterprise signal: an average of 45 billion tokens per day in July 2026. The measures are not comparable, but they show token consumption at a scale where aggregate demand becomes a management issue in its own right.

For finance, a lower rate per token may coexist with higher total spend and still represent successful adoption. The question is whether the additional activity creates sufficient value after infrastructure, oversight and exception handling. Favorable unit economics and unfavorable aggregate economics can occupy the same dashboard.

Cheap automation changes what management has to manage

Text messaging offers a useful precedent. As message economics improved, software generated a larger share of traffic: authentication codes, order updates, appointment reminders and other application-to-person messages. Price per message became less revealing than who could send, what recipients had agreed to receive and which traffic should be delivered or stopped.

Governance emerged across industry practice and law. Mobile operators and messaging providers developed the 10DLC registration ecosystem to identify brands and declared campaigns; CTIA practices emphasize consent and consumer choice; FCC rules separately govern obligations such as honoring revocation. The architecture is specific to messaging. The progression from cheap capacity to control is the relevant precedent.

Exhibit 3 · The constraint moves

When automated volume scales, price becomes only the first management question

The meter remains in operations. The constraint moves into the operating model

AI will not reproduce the same architecture. The relevant lesson is economic: once software can initiate activity, control becomes part of the cost of scaling it. Orchestration, observability, evaluation and exception handling belong in the model alongside inference.

What leadership should do

Budget the activity. Govern the authority.

Model the volume each workflow can generate, the authority it carries and the controls required to contain it. Unit cost is an input to that decision—not the decision itself.

Discuss the implications

Selected sources

Evidence and context

Primary and authoritative sources support the factual claims. The interpretation is ZiffyVolve’s.
Google / AlphabetAI scale across Google surfaces3.2 quadrillion tokens per month in 2026; 9.7 trillion two years earlier.AT&TThe tokenomics equationAn average of 45 billion tokens per day reported in July 2026.FinOps FoundationToken economics: the atomic unit of AI valueConsumption, orchestration and the limits of unit-cost analysis.CTIAMessaging principles and best practicesIndustry practices for consumer and non-consumer messaging.The Campaign Registry10DLC ecosystem overviewCarrier-led registration for brands and application-to-person campaigns.Federal Communications CommissionRules governing revocation of consentLegal requirements for honoring requests to stop robocalls and robotexts.Yale Energy HistoryW. Stanley Jevons and The Coal QuestionHistorical context for the rebound-effect analogy.