Tokens are the new currency of enterprise AI, and the bills are coming due. After companies blew through annual AI budgets in months, leaders at Splunk's .conf26 shared hard-won lessons for the agentic era.
"Tokens have squarely become the new currency," Jeetu Patel, president and chief product officer at Cisco, told the audience from the keynote stage at Splunk's .conf26 user conference in Denver.
A year ago, when many organizations were embracing "tokenmaxxing" tactics like consumption-incentivizing leaderboards, such a statement might have drawn a shrug. But after companies like Uber announced earlier this year that they had blown through their entire annual AI budget in just a few months, tokenomics has become a hot topic.
Agentic tokenomics is the practice of managing the cost, consumption and return on investment of tokens used by autonomous AI agents that plan and execute multi-step tasks across enterprise systems. Here are the most important cost-control lessons shared by leaders during the Splunk conference.
Is the Tokenmaxxing Era Over?
The blank-check approach to AI consumption ended almost as quickly as it began, as agentic workloads push token volumes far past what human-driven queries ever generated.
If today's generative AI bills are not enough to motivate a more conservative approach to consumption, tomorrow's agentic AI bills almost certainly will be. Patel said agents already consume more tokens than humans.
"In the past, the human would ask a question," Patel said. "There was a certain amount of cost that you would typically have that was incurred for tokens. Now, what happens is, rather than having a little bit of infrastructure to consume, you start to see a huge amount of volume back and forth. These agents tend to be chatty."
It’s time to tighten up, according to Greg Leffler, director of developer evangelism for Splunk.
"The blank check era ended very quickly,"Leffler told The Forecast in an interview. "We were in that era for a couple of months, and then the bills started coming in."
Tokenomics is a sign AI has matured, Kamal Hathi, senior vice president and general manager for Splunk at Cisco, told journalists at the event.
"We went from 'let's start leveraging AI' to 'let's get the best possible return on investment,'" Hathi said.
How Should Organizations Measure AI Value?
Cost controls must go beyond spending caps to calculate the return on every token, tying consumption to business outcomes rather than raw usage.
Cost controls must be more nuanced than simply setting a spending cap, explained Vikram Chatterji, senior director of agent resilience at Splunk and former CEO of Galileo. Instead, he told a press panel, organizations must try to calculate the return on their AI spending.
"When we think of tokenomics, it's not just, what's the cost, and how do I keep tabs on it?" Chatterji said. "It has to be dynamic pricing based on the ROI."
In the same panel, Hao Yang, head of AI at Splunk, said organizations can track simple metrics such as task completion rate to ensure agents are performing the work organizations ask of them. "It drives the conversation to business outcomes," he said.
What Costs Lurk Beyond Token Spend?
Agents can trigger third-party services and cascade costs that dwarf the token bill, making total agent spend the metric leaders need to track.
During the press panel, John Morgan, senior vice president and general manager of Splunk Security, said the total cost of agentic AI includes factors beyond tokens.
"An agent spend is not the same as token spend," Morgan said. "An agent might trigger off another third-party service. The standards need to get to the point where we're looking at agents, agent spend, and can quickly figure out the token spend, which is a prerequisite for that.
“Token spend might right now be the highest portion of that, but agents could easily spin off and create costs that are much, much higher."
When Does a Frontier Model Cost Too Much?
Routing every task to the most powerful model inflates bills, but underpowered models trigger retries that can cost more in the long run.
"I see a lot of people that use frontier models for literally everything they do, and they're like, 'Why is our AI bill so expensive?'" Leffler said. "Well, you're calling the most expensive luxury model for every single thing you're doing. You have to learn that there are different models and that they have different price tiers, and that those price tiers change."
Yang said choosing an underpowered model may lead to large numbers of retries and higher overall costs.
"The optimal strategy isn't always to route the problem to the cheapest model," he said. "The metric you're trying to optimize is the cost for you to complete that task."
Some organizations are spending money on AI for use cases that are a better fit for a simple automation script, explained Ronak Desai, founder and CEO of the AI site reliability engineering (SRE) startup Ciroos.
"If it is a very predictable workflow, where you do A and then B and then X and Y, you shouldn't be using AI," Desai said.
Why Build Cost Controls Before Deployment?
Agents left unchecked can rack up tens of thousands of dollars in days, making proactive guardrails essential during development, not after.
Proactive cost controls are particularly important for agentic AI, experts and practitioners said. Karun Subramanian, senior director of AI for infrastructure and operations at UnitedHealth Group, pointed to a company that ran up a $45,000 bill after agents "went down a loop" and made millions of calls to external services in just a couple of days. "You need to make sure that the token economics is considered during the development stage," he said during a panel discussion.
How Do You Budget for Unpredictable Consumption?
Token use defies deterministic budgeting, forcing organizations to combine forecasting, testing, observation and guardrails.
Experts and practitioners stressed that token consumption is inherently less predictable than conventional compute, requiring organizations to replace deterministic budgeting practices with a combination of forecasting, testing, observation and guardrails.
"How can you manage something you cannot predict?" Subramanian said.
"It's not just going to your OpenAI dashboard to look at how many tokens you are using. You've got to go deeper than that. You have to look at the trend. You have to do some pattern analysis."
Why Is Standardization Lagging on Token Telemetry?
Vendors expose cost data in inconsistent ways, leaving IT teams to reconcile dozens of formats before they can understand their spending.
Chatterji said vendors are not routinely making telemetry data available to their customers, resulting in gaps in standardization and transparency.
"When the IT teams ask those vendors, 'Give me the ability to actually track cost,' it's not done in a standardized way," Chatterji said.
He explained that the Linux Foundation is pushing toward standardization.
"Unless that actually happens, IT teams are going to struggle with having 20 different ways by which they're getting cost brought in," he said. "It's going to be very difficult for them to standardize and understand how much they are spending. That is a pretty big problem that needs to be solved."
Related:
What Is an AI Agent Gateway? How Enterprise IT Governs Agentic AI at Scale
Token Era Reckoning: 2026 AI Economics and Infrastructure Readiness
Own Your Token Machine: How Composable GPU Infrastructure Rewrites AI Economics
AI's Defining Decade: Nutanix Leaders on Inference, Agents, and the Road Ahead
Calvin Hennick is a contributing writer. His work appears in BizTech, Engineering Inc., The Boston Globe Magazine and elsewhere. He is also the author of Once More to the Rodeo: A Memoir. Find him on LinkedIn.
© 2026 Nutanix, Inc. All rights reserved. For additional information and important legal disclaimers, please go here.
Related Articles
