# Why No One Agrees on What Their AI Tokens Cost

Pratyusha Duvvuri, Cisco | AI Engineer NYC 2026 | 18:57

Source: https://www.youtube.com/watch?v=KdG9OSOtJOA
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/why-no-one-agrees-on-what-their-ai-tokens-cost
Published: 2026-11-09
Tags: cost, enterprise, finance, observability

## TL;DR
- Tokenomics is the observability and interpretability of token usage and cost.
- AI spending now needs to account for efficiency, custom provider pricing, and different needs across engineering and finance.
- A useful tokenomics product combines provider data with company and finance hierarchies, while handling changing bills, users, and permissions.

## Summary
Pratyusha Duvvuri defines tokenomics as the observability and interpretability of token usage and cost. She explains why the term means different things to engineers, engineering managers, and finance teams. Engineers need model and budget detail. Managers need adoption, trends, and comparisons between teams. Finance needs budgets, alerts, return on investment, and spend mapped to its own ownership structure. Provider dashboards do not combine data across vendors, and their initial bills can change after reconciliation. Duvvuri walks through a dashboard with cost, adoption, forecasting, organizational rollups, job-family views, model filters, and a separate finance hierarchy with access controls. Her team took the product from customer conversations to general availability in 10 weeks. She is honest about the difficult parts: provider pricing varies by customer, data changes over time, employees move, and one product must stay simple for very different audiences.

## Key ideas
### Tokenomics has different meanings across companies
[00:01](https://www.youtube.com/watch?v=KdG9OSOtJOA&t=1s)
Duvvuri asked about 100 companies and spoke with engineers, product managers, engineering managers, and CFOs. Their answers ranged from basic observability, such as knowing what tokens cost, to a full financial breakdown of return on investment for every team. She does not claim there is one universal definition. Instead, she offers a working version that can guide product decisions.

### AI spending is moving from status to efficiency
[01:49](https://www.youtube.com/watch?v=KdG9OSOtJOA&t=109s)
Duvvuri describes a change from treating high token usage as a status symbol toward asking whether model calls are efficient. In an internal Cisco use case, a cheaper model scored 96% in evaluations while a more expensive model scored 99%. The team kept the cheaper model for internal work. For customer-facing use, she says they might accept the extra cost for better customer satisfaction.

### Provider pricing is increasingly customized
[04:03](https://www.youtube.com/watch?v=KdG9OSOtJOA&t=243s)
Model providers no longer have one simple pricing arrangement for every customer. Duvvuri describes agreements with different per-credit prices and thresholds, including contracts that charge nothing up to a limit and then apply a higher rate. This variation makes a shared view of token cost more useful than relying on a generic provider price list.

### Different users need different cost information
[04:42](https://www.youtube.com/watch?v=KdG9OSOtJOA&t=282s)
Engineers want to know which model to use, how much of a budget they have spent, and which skills or tools consume tokens. Engineering managers need adoption and trend data across teams, including cases where one team spends more while producing weaker results. Finance teams need budgets, alerts, return-on-investment information, and spend mapped to their own ownership structure.

### Provider dashboards do not provide one accountable view
[06:46](https://www.youtube.com/watch?v=KdG9OSOtJOA&t=406s)
A company may have separate dashboards for OpenAI, Anthropic, GitHub Copilot, and other providers. Duvvuri says users must otherwise move between these systems, while the dashboards do not serve finance and management needs in the same way. Provider bills can also change after reconciliation. A bill initially shown as $100,000 might later become $120,000, so the combined system must update its records.

### Tokenomics combines usage visibility with interpretation
[08:38](https://www.youtube.com/watch?v=KdG9OSOtJOA&t=518s)
Duvvuri defines tokenomics as "the observability and interpretability of token usage and cost." Interpretation includes return on investment, the set of models being used, and the right categories for a particular use case. The definition is deliberately broad because the useful product depends on what each audience needs to understand.

### The product depends on organizational and provider data
[09:18](https://www.youtube.com/watch?v=KdG9OSOtJOA&t=558s)
The dashboard needs a company hierarchy from Workday, ADP, or a similar system, along with API keys and organization IDs for the model providers. It also needs the finance hierarchy, which differs from the management hierarchy. To limit sensitive workforce data, customers can provide a CSV with names, IDs, job functions, and other basic fields instead of their complete employee history.

### The dashboard joins cost, adoption, forecasting, and ownership
[10:51](https://www.youtube.com/watch?v=KdG9OSOtJOA&t=651s)
The product lets users filter by time range, including a fiscal calendar. Its leader view contains separate cost and adoption dashboards, trend information, projected usage, organizational rollups, and consumption by job family. Users can inspect individual and team usage, compare groups such as sales and engineering, and filter by person, job function, model, or thinking pattern.

### Changing data and multiple audiences make the system hard to build
[16:44](https://www.youtube.com/watch?v=KdG9OSOtJOA&t=1004s)
Duvvuri names reconciliation as a major challenge because provider data can change after an earlier figure was presented. Employees and managers also change, which affects attribution. The system must handle large organizations and new product users while remaining simple for engineers, managers, and finance teams. Her team shipped the product to general availability in 10 weeks, after a demanding development sprint.

## Notable quotes
- "Tokenomics is the observability and interpretability of token usage and cost." (08:38)
- "Everyone has their custom way of defining costs for tokens, which is especially why tokenomics starts to become important." (04:23)
- "We went from zero, talking to customers, to building a full-fledged product and shipping it to GA in 10 weeks." (15:48)
- "The true challenge is how do you make it so simple and dumb and easy to use that everybody just gets what they want from it." (17:47)

## Tools & references mentioned
- Galileo
- Splunk
- Cisco
- OpenAI
- Anthropic
- GitHub Copilot
- Workday
- ADP
- Jensen Huang
- Chuck Robbins

## Who should watch
- You are building internal controls for AI usage across several model or coding-tool providers.
- Engineering leaders need to compare adoption and spend across teams, while finance needs a separate ownership and reporting structure.
- Your provider bills, employee ownership, or model pricing changes after the initial usage data is collected.

## Related talks

- [The AI Bill Is Too Damn High: Tokenomics 101](https://aietalks.com/talks/the-ai-bill-is-too-damn-high-tokenomics-101) (Eugene Cheah, Featherless AI, 55:25)
- [Tokens Are the Cheap Part](https://aietalks.com/talks/tokens-are-the-cheap-part) (Steve Jamison, Jellyfish, 17:32)
- [Notion's Token Town](https://aietalks.com/talks/notions-token-town) (Sarah Sachs, Notion, 23:55)
- [FinOps for AI Agents: Who Spent All the Tokens?](https://aietalks.com/talks/finops-for-ai-agents-who-spent-all-the-tokens) (Tisha Chawla & Susheem Koul, Microsoft, 21:24)
- [From Tokenmaxxing to Trusted Throughput](https://aietalks.com/talks/from-tokenmaxxing-to-trusted-throughput) (Mingsheng Hong, Ironclad, 23:04)
