The Economics of AI Development
Why tokens get cheaper while budgets grow—and why accepted work is the unit that matters
Slide contents
1. The Economics of AI Development
Why tokens get cheaper while budgets grow—and why accepted work is the unit that matters
2. Tokens fall. Budgets do not
Original forecast synthesized from market, FinOps, and agentic-workload evidence
3. Growth is outrunning outcomes
Three independent signals with different sample boundaries
4. The 2029 arrows diverge
A synthesis of historical trends, tariffs, and demand expansion
5. Only one multiplier gets cheaper
Spend expands through tasks, calls, context, and infrastructure
6. Bound the task, not the person
Money, steps, time, and authority are bounded across the workflow
7. Costs live at different levels
Seat, exploration, production, platform, and risk have different owners
8. Showback before chargeback
Expose cost and outcome before hard allocation
9. Count the accepted task
Tokens are a billing unit but a weak value unit
10. Quality first, price second
Only routes that pass one acceptance threshold are comparable
11. Self-estimates are not outcomes
Observed tension
METR: 19% slower
Participants expected +24%
Narrow 2025 sample
The useful conclusion
Not a universal penalty
Tools already changed
Build your own baseline
12. The deepest lock-in is behavioral
Prompts, tools, evals, and habits move more slowly than HTTP clients
13. Own the task contract
State, data, evals, and acceptance belong to the company
14. Contracts do not start with spend
SLA, capacity, data terms, and criticality matter before discounts
15. People need modes, not models
Auto, Fast, Deep, and Sensitive hide market volatility
16. Routing closes on outcomes
Policy, observability, and evals form independent control loops
17. Maturity starts with telemetry
A learned router is the last step, not the starting point
18. Local wins only when Q* exists
Align the quality gate before comparing fully loaded cost
19. Hybrid beats ideology
Local handles stable flow; frontier handles the difficult tail
20. Ninety days reveal the economics
Visibility, guardrails, and portfolio management arrive in sequence
21. One scorecard connects the system
The operating measurement model for every task class
22. Manage work, not tokens
Count accepted tasks
Put budgets on traces
Budget at natural ownership levels
Own state, evals, and contracts
Start routing with telemetry
Cheap intelligence expands demand faster than it shrinks the bill