TLDR

Builders already talk about AI leverage. 75 percent cite productivity. Finance still cannot book most of it. Tokenomics SF is the translation problem. Tokens, tasks, and margin are three ledgers. Until they share IDs, the valuation story floats.

Key Takeaways
  • ICONIQ's GenAI survey (April 2025, N=258) is a leverage story. 75% productivity, 51% cost savings, 20% revenue, 20% retention, and a claimed 35% valuation premium.
  • Tokenomics Foundation's San Francisco session found the opposite problem. Spend is visible. Return is not.
  • Cost per token compares infrastructure. It does not compare models. Tokens are only 10 to 25% of total AI cost.
  • Saved task time that dies in review queues has no realized value. Freed capacity has to be redirected.
  • Enterprise value is revenue growth times operating leverage times defensibility times AI readiness. Adoption alone does not raise the multiple.

On 20 October 2026 about 100 practitioners met in San Francisco to debate tokenomics. J.R. Storment's recap for the Tokenomics Foundation is blunt. Enterprises can see AI spend. They cannot show the return. Suppliers measure tokens. Engineers measure tasks. Finance measures margin and headcount. Read the full note, Engineers and CFOs Aren't Speaking the Same Language Yet.

That finding collides with a slide most operator decks now carry. AI is leverage: a small change in input, a large change in output. ICONIQ's GenAI survey of 258 builders (April 2025), as used in a TechCrunch deck and later McKinsey State of AI 2026 press summaries, is the chart. 75 percent cite productivity. 51 percent cite cost savings. Only 20 percent cite revenue. Only 20 percent cite retention. The same briefing stamps a 35 percent valuation premium.

Tokenomics SF is the corrective. Leverage that dies in a review queue is not a P&L event.

Five things the room actually agreed on

Workforce. AI is already in headcount plans. The Foundation splits labor into capacity gain, augmentation, and automation.

Saved time. A delegate described a 23 percent task-level developer gain with no company-level effect, because code still waits for review. Developers reporting 10-plus hours saved a week fell from 68 percent to 26 percent in a year. The hours went to meetings, search, and checking AI output. Freed capacity that is not redirected has no realized value.

Budgets. A technology CFO at a large regulated firm will move AI consumption into business-unit budgets, next to labor and software. They watch pre-tax margin, revenue per employee, and fully loaded unit cost, not a standalone ROI slide. One customer token bill rose about 20 percent after a single new engineer joined.

Cost boundary. Cost per token compares two stacks running the same model. It misleads across models, because a cheaper model may burn ten times the tokens on the same task. Same lesson as value per inference dollar. Tokens are 10 to 25 percent of total AI spend.

Checkable output. Value arrives fastest where you can test the result. Coding agents open more pull requests. Review is the bottleneck.

Adoption does not raise the multiple

A second operator slide writes enterprise value as a product: revenue growth times operating leverage times defensibility times AI readiness. Subject to market conditions, risk, and valuation multiples. One M&A comparison cited there is AI-native names near 11.5 times EV to revenue versus legacy SaaS near 3.8 times.

That spread is not a trophy for rolling out copilots. Adoption alone does not raise multiples. Buyers pay for durable cash flows, defensible advantages, and proven productivity gains. A 75 percent productivity score in a builder survey is a leading indicator. It is not booked margin, not quality-gated output, and not a named cost owner on the P&L.

Translate the two ledgers

The Foundation's early proposal, aimed at Tokenomicon in June 2027, combines four records: the metered bill in FOCUS, runtime telemetry with shared workload IDs, a verified result, and a value claim finance can book. Same point as how to measure AI ROI and how AI turns text into math.

A desk example

Illustrative numbers, not a survey result. Ten engineers at $180k fully loaded is $1.8M labor. Tokens at $2k a month are $24k a year. If tokens are 20% of total AI cost, the fully loaded AI bill (tokens, tools, retries, review) is $120k. A 23% faster ticket cycle is the Tokenomics SF task-level gain. If 40% of saved hours go back to review, realized capacity is 23% times 0.60, about 14%, or $250k of the labor base.

  • Hours return to meetings: booked value $0.
  • Six extra closed tickets a month at $5k contribution: $360k. Minus $120k AI, booked $240k. Value per AI dollar is 2.0.

The 23% lives on the engineer ledger. The $240k lives on the CFO ledger. Only the second number is a P&L event.

Sources