Pratyusha Duvvuri demoed Cisco's tokenomics dashboard built from interviews with 100 companies on AI Engineer.
Token maxing is out. Efficiency is in. Duvvuri cited Jensen Huang's 250k token spend bar for 500k employees, per AI Engineer. She then cited Ramp's report of a 1.5M dollar bill for one week, per AI Engineer.
Duvvuri defines tokenomics as observability plus interpretability of token use and cost, per AI Engineer. Engineers watch a 10,000 dollar monthly budget and model choice. Managers compare team KPIs against spend and adoption. Finance wants ownership mapping and ROI proof. Duvvuri said, "You're riding or sailing a ship where you don't even know the destination, and it's a big ship."
The product pulls Workday hierarchy, provider API keys and a separate finance hierarchy, per Duvvuri. It refreshes provider data on a one day cadence, per Duvvuri. The leader view splits cost and adoption with a ghost graph for projected use. Costs roll up from staff to CEO Chuck Robbins and slice by job family, model and thinking level. The talk was delivered at AI Engineer NYC 2026, per DEV Community coverage.
Reconciliation is the hard part. Duvvuri said a 100,000 unit OpenAI bill can later correct to 120,000 units. Provider deals also vary by customer, with 4 cents per credit, 3 cents per credit or free use to 30 dollars then 10 cents, per Duvvuri. Internal tests showed a cheaper model at 96% on evals versus 99% on a flagship model, per Duvvuri. Duvvuri blurred live numbers because Cisco had not cleared them for release.
Finance needs strict role controls and separate trees for OPEX and COGS, per Duvvuri. The build covered OpenAI, Anthropic and GitHub Copilot plus other providers in one view. Cisco went from zero to GA in 10 weeks serving 90,000 people, per Duvvuri. Forecasting ships in the next couple of weeks, followed by ROI inference, per Duvvuri.
