The residual-scarcity test
From residual scarcity to scarcity migration
Technological progress does not eliminate economics. It changes where the binding bottleneck sits. Coding skill may yield to compute; compute may yield to energy; energy may yield to grid connection, land or legal permission. Software functionality may yield to trusted state; a professional directory may yield to identity and access rights.
S₀ → S₁ → S₂ → … → Sₜ
Each S is the binding scarcity at a stage of the system. The valuation question is not merely what remains scarce, but whether the provider being valued owns the new scarcity, can follow it through vertical integration, or depends on another actor that captures the rent.
Only after all three attacks have been evaluated should durable scarcity rent be assigned.
| Candidate residual | AI attack strength | Hypothesized durability | What must remain irreducible |
|---|---|---|---|
| Software implementation | Very high | Low | Little beyond execution |
| Routine expertise | Very high | Low | Unresolved judgment or liability |
| Data quantity | High | Low–medium | Unique information content |
| Compute requirement | High but contested | Medium | Demand growth exceeding efficiency and new supply |
| Distribution | Medium | Medium | Access that agents cannot route around |
| Trust | Medium | Medium–high | Judgment not replaceable by verification |
| Network counterparties | Medium | Medium–high | Actual participation, not information aggregation |
| Regulation and legal permission | Low–medium | High | Sovereign authorization |
| Unique locations and resources | Low | High | Physical non-substitutability |
| Sovereign authority | Very low | Very high | Legitimate coercive and legal power |
This table is a hypothesis map, not a forecast. Compute can be reproduced through better design and fabrication, substituted by alternative architectures, or reduced through algorithms, sparsity, caching and specialized execution. Energy faces new supply, substitute generation and efficiency. Data faces collection, simulation, synthetic generation and sample efficiency. Distribution faces cheaper acquisition, interface displacement and agent-mediated discovery. Regulation and physical resources face weaker attacks and may therefore become stronger destinations for rent.
The paper can now state a general test for any AI valuation. Start with the economic function and remove every layer AI can plausibly make cheap, reproducible or substitutable. What remains?
Define the human or institutional outcome, not the current product category.
Estimate the post-transformation resource cost rather than inheriting the historical cost pool.
Identify the cheapest sufficient alternative, including non-AI redesigns.
Separate recurring frontier inference from work that can migrate to specialized execution.
Estimate realistic purchase frequency, utilization and provider share.
Identify physical, legal, network, data, trust or distribution bottlenecks that remain scarce.
Ask whether the provider owns those bottlenecks or merely depends on them.
Estimate how quickly AI lowers the cost of attacking each bottleneck.
Only then assign margin, growth and duration assumptions.
Run the valuation with explicit terminal economics rather than a narrative multiple.
Durable extraordinary value requires invested capital × a positive (ROIC − WACC) spread × competitive-advantage period and fade
If the required invested-capital base, the ROIC−WACC spread or the competitive-advantage period tends toward zero, the product can remain extraordinarily useful while the provider's extraordinary economic profit disappears.