Five Things We Learned in San Francisco about Tokenomics

KEY INSIGHT: On October 20, ~100 practitioners and suppliers met in San Francisco to debate Tokenomics. We learned that enterprises can see what they spend on AI but struggle to show what it returns. Suppliers measure tokens, engineers measure tasks, and finance measures margin and headcount, and those measures don’t yet connect. Attendees came from chipmakers, neoclouds, hyperscalers, financial institutions, consumer platforms, and universities. The half-day working session took place before the Tokenomics Foundation’s first in-person governing board and was held under the Chatham House Rule.

October 20, 2026 · San Francisco

Five Things We Learned at State of Tokenomics SF

1. AI is already reshaping the workforce. Labor came up in nearly every conversation, and few organizations have a way to plan headcount around AI that they can defend.
2. Saved time doesn’t reach the company. Task-level gains are real, but the time often goes to review queues and rework instead of more output.
3. AI budgets are moving to the business. When AI spending sits in the same budget as headcount and software, teams start making real trade-offs.
4. Nobody agrees where AI cost ends. Tokens are only part of the bill, and the hard cases (embedded vendor AI, capitalization, CPU costs) are unresolved.
5. Value comes fastest where output can be checked. The ability to verify AI output may matter more than the speed of producing it.

1. AI Is Already Reshaping the Workforce

The host described labor as one of the largest parts of the AI economics conversation. Organizations are seeing four changes at once: productivity gains, shifts in skills (more AI roles, fewer generalist and entry-level roles), budget trade-offs between hiring and AI, and layoffs. One attendee said they avoid discussing the last one because replacing people feels wrong. The host’s answer was that it is happening and has to be part of the conversation.

The State of Tokenomics survey found the same four themes. Outside data shows the pace. AI was cited in 101,743 U.S. job cuts through June 2026, nearly double the 2025 total, and Gartner found no correlation between AI-related reductions and better return on investment (ROI).

A technology chief financial officer (CFO) at a large regulated institution described a more measured approach. Teams now list AI as a driver in their workforce plans, alongside other efficiency levers. The company tracks net headcount change, weighing new investment against efficiencies, rather than counting cuts alone.

The Foundation’s value classification work separates labor claims into three categories, each with its own baseline. It also treats employee engagement and retention as outcomes to measure. Applying those categories to workforce planning is one of the clearest next steps for the community.

Capacity gain

The same people produce more.

Labor augmentation

AI does part of the task, and a person stays in the loop.

Labor automation

AI does the task end to end.

2. Saved Time Doesn’t Reach the Company

During Q&A, a delegate pointed out that developer productivity can rise 23 percent at the task level with no visible effect at the company level, because code still waits for review, approval, and release. The researcher on stage didn’t know of a formal study on this.

Outside data points the same way. The share of developers saving 10 or more hours a week with AI fell from 68 percent to 26 percent in a year, with the time going to meetings, searching for information, and checking AI output.

Teams using AI every day think in tasks. An agent may take dozens of steps, call tools, fail, and retry, so panelists recommended measuring cost per successful task, including the cost of a person fixing a run that didn’t finish. That measure is a start, but it stops at the task. The Foundation’s value classification work makes the next point explicit: freed capacity that isn’t put to other work has no realized value. The open question for leaders is how to redesign the workflow so the saved time goes somewhere useful.

3. AI Budgets Are Moving to the Business

The technology CFO said ROI is the wrong place to start, because AI is becoming part of nearly everything the company does. They compared it to electricity, which nobody evaluates on ROI anymore. Instead, they track long-term financial metrics such as pre-tax margin and revenue per employee, along with fully loaded unit costs by platform, such as cost per trade or cost per payment.

Next year, their company will move AI consumption budgets from a central pool into business units’ own budgets. A team that wants more AI will have to fund it by trading off labor or software spending. Cost dashboards have already led some teams to shut down initiatives after seeing what they cost.

Spending controls elsewhere are still basic. One supplier described a customer whose token bill rose about 20 percent after one new engineer joined the team. A researcher noted that flat per-person token limits lead some people to stop work until their allowance resets the next day. Attendees preferred soft caps combined with alerts when spending jumps.

The broader market is moving the same way. AI funding is shifting out of IT and into business-unit budgets (Futurum). In a recent webinar, Bain recommended separating predictable seat costs from variable consumption and matching controls to the risk of the work:

LSEG’s chief technology officer makes a similar case in a recent conversation with J.R. Storment: the leaders who own the P&L should also own the AI bill.

Watch: LSEG CTO: You’re Getting AI Accountability Wrong. Owning P&L Means Owning AI Bill.

4. Nobody Agrees Where AI Cost Ends

At recent energy and hardware conferences, the host said, vendor signage promised more tokens and more revenue. Asked where the revenue would come from, one vendor answered that the business would use the tokens to do things. A panel of chip, neocloud, and hyperscaler leaders took up that gap between making tokens and using them.

One panelist defended cost per token as a fair way to compare two infrastructure options running the same model and workload. The same panelist said it misleads when comparing models, since a cheaper model may use 10 times as many tokens for the same task.

On renting versus owning hardware, panelists named four factors: workload predictability, data sensitivity, sustained utilization, and the skills to operate it. At around 30 percent utilization, renting costs less.

A principal engineer added that open-source inference throughput roughly tripled in three months, so owned hardware can produce more over time if its software is kept current. Panelists also noted that agents run their tool calls on CPUs, which changes the hardware mix and adds a cost many estimates miss.

The technology CFO described the cases that remain unresolved:

The Foundation’s Total Cost of AI work estimates that tokens are 10 to 25 percent of total AI spend. FinOps practitioners supply the cost and usage data through the FinOps Open Cost and Usage Specification (FOCUS). The next step is agreeing on how to count these harder cases.

5. Value Comes Fastest Where Output Can Be Checked

The principal engineer argued that AI is advancing fastest in areas like software, where output can be tested against a known correct answer. In their view, a company’s ability to judge whether output is good will matter more than how quickly it can iterate. Sandboxes help, because they let agents run and test freely without putting production systems at risk.

Panelists described enterprises that now run automated evaluations on every new model within days of release, instead of spending a month on manual testing. Others pointed to the cost of failure: retries, failed runs, and the human time needed to fix them.

Outside data supports the point. Teams using coding agents now open far more pull requests than other teams, which has made code review the bottleneck (InfoQ). The Foundation’s value classification work already counts only output that clears a quality floor. The next question is how the ability to check output determines where AI can deliver value at all.

What Attendees Want Built

Asked what the industry should build together, panelists named telemetry that shows what tokens are spent on, benchmarks based on real enterprise workloads, and a common taxonomy so a token means the same thing across providers. One compared the benchmarks to the driving cycles behind fuel-economy ratings. Several also wanted a measure that links the value engineers see to the value a CFO will book.

The Foundation presented an early proposal. It would combine four kinds of data:

The target date is June 2027 at Tokenomicon, and the Foundation described the proposal as a starting point for discussion.