Skip to content
← All insights

[AID | AI Infrastructure] The AI cost narrative has flipped: from 'tokenmaxxing' to 'cost per outcome'

Tokenmaxxing is over. Between October 2025 and July 2026, calls to spend more on AI gave way to spending caps, model routing, and demands for proof of return. The question is no longer how many tokens a company uses but what those tokens bought.

[AID | AI Infrastructure] The AI cost narrative has flipped: from 'tokenmaxxing' to 'cost per outcome'

Between October 2025 and July 2026, the AI cost conversation moved from "how much are you using" to "what did the spend actually buy". Connectionary tracked public statements on AI cost across that period with AID.

Advocacy · 'spend more'
14%5%
share of statements, first vs. last period
Control & pushback · efficiency and ROI
28%69%
share of statements, first vs. last period
The narrative turned
May 2026
budget shocks made cost the story
Advocacy (use more)NeutralControl and pushback (efficiency · ROI)
Oct 2025 – Feb 2026
n = 36

Infrastructure anxiety

Cost meant GPUs, data centers, power, capex: "is the industry spending too much?" Enterprise token budgets were not yet the issue.

Mar – Apr 2026
n = 77

Tokenmaxxing

Usage read as a productivity signal. Companies ran internal leaderboards and pushed employees to consume more.

"Spend at least half of salaries on AI."

ADVOCACY Jensen Huang, NVIDIA

Apr – May 2026
n = 121

Budget shock

Uber's full-year AI budget gone in four months. Microsoft narrowed internal Claude Code access. The bills arrived.

"We used up the annual AI budget by April."

CONTROL Praveen Neppalli Naga, Uber CTO

Jun – Jul 2026
n = 168

Cost per outcome

Model routing, spending caps, usage monitoring. The question is no longer "how many tokens" but "what did they buy."

"The issue never came up. Then suddenly it did."

CONTROL Sam Altman, OpenAI

Public statements on AI cost, Oct 2025 to Jul 2026 · 47 from Korea · attitudes coded from speakers' own public remarks · Analysis by AID · Connectionary

One company, two answers: the split inside the sellers

SELLING TOKENS

"Spend at least half of salaries on AI."

Jensen Huang, CEO, NVIDIA

PAYING FOR TOKENS

"My team's compute costs now exceed its payroll."

Bryan Catanzaro, VP, NVIDIA

The input stage

JoongAng Ilbo's Factpl covered this shift and how enterprises are responding. In the piece, Connectionary pointed to one of the biggest sources of waste: the input stage, where AI reads everything on the page, then pays for it again on every query. Structure information once, and every query becomes cheaper, faster, and more reliable.

What matters

🔴 RISK: AI spending has outrun annual budgets and internal access controls. Uber exhausted its full-year AI budget by April, while Microsoft narrowed internal access to Claude Code as usage bills arrived.

🟡 WATCH: Enterprises are moving from usage targets to outcome-based controls. Model routing, spending caps, and usage monitoring are replacing token consumption as the operating focus. The unresolved question is what AI spending actually produces.

🔵 SIGNAL: The AI cost divide now runs inside the companies selling consumption. NVIDIA's CEO advocates spending heavily on AI, while an NVIDIA vice president says his team's compute costs exceed payroll. Sellers are also confronting the economics of paying for tokens.


The AI cost debate has moved from maximizing consumption to proving outcomes, and the split now runs through the companies selling the compute.

This analysis was featured in JoongAng Ilbo Factpl, "ChatGPT billed 10 million won: pay for what you use, the AI bill shock" (Jul 28, 2026).