Claim
When AI supply constraints drive frontier model prices up, labs preserve total GPU-hour consumption by segmenting models into premium, mid-market, and value tiers. Workloads that can no longer afford frontier routing route to cheaper tiers rather than being abandoned, sustaining Jevons' Paradox even as per-token costs at the top rise.
Mechanism
Jevons' Paradox holds that efficiency gains in resource use increase total consumption rather than reducing it. AI labs face a version of this in reverse: supply constraints push frontier prices up, which should reduce total demand. The tier segmentation strategy converts that risk into a demand routing problem. Premium models cost 13x more than value-tier models for 20% additional capability. Mid-market models deliver 96% of frontier capability at 40% of frontier cost. Value models reach 84% of frontier capability at 1 to 5% of cost. This spread means most workloads have a cheaper substitute that is good enough for the task.
The critical infrastructure layer is routers. As price sensitivity grows, the intelligence layer routing requests to the right model becomes the margin-capturing asset. A lab that offers all three tiers with a router captures the full demand curve. A company building on AI inference needs to model which workloads require frontier capability and which can be rerouted down without quality loss.
Specific 2026 pricing anchors: Anthropic Fable 5 at $50 per million output tokens; Google Gemini at $12; GPT-5.6 Luna at $0.20/$1.20 per million tokens; DeepSeek V4 Flash at $0.03.
Conditions
Holds when: workloads are fungible across capability tiers, meaning mid-market and value outputs are acceptable substitutes for frontier outputs in that task. The tier-routing model requires that a meaningful portion of the workload is not frontier-only.
Fails when: the task genuinely requires frontier capability (autonomous research, novel reasoning, high-stakes judgment calls), and no cheaper tier delivers acceptable quality. In those cases, price increases reduce consumption rather than routing it. Also fails if open-source models commoditize the value tier to zero marginal cost, collapsing the segment economics for commercial providers.
Evidence
Supply constraint confirmed by three hyperscaler CEOs in Q2 2026 earnings:
"We continue to be supply constrained, a sign of momentum & rapid adoption." (Sundar Pichai, Alphabet)
"The demand we already have for 2028 is striking." (Andy Jassy, Amazon)
"Cloud GPUs are sold out." (Jensen Huang, NVIDIA)
Pricing data: premium models cost 13x more than value tier for only 20% additional capability. Mid-market tier delivers 96% of frontier capability at 40% of premium cost. Value tier provides 84% of intelligence at 1 to 5% of premium pricing.
Signals
- Your AI inference bill rises as the frontier price increases but overall token consumption stays flat or grows, suggesting routing substitution is working.
- Mid-market models (GPT-5.6 Sol, Kimi K3) begin handling tasks previously sent to frontier-only endpoints with no perceptible quality drop.
- The share of your inference spend on frontier models declines quarter over quarter even as your total AI budget grows.
Counter-evidence
Jevons' Paradox historically applies to efficiency gains reducing per-unit cost; the supply-constraint scenario driving prices up is structurally different and may not sustain the same dynamic. If frontier price increases outpace what routing can absorb, net GPU-hour consumption could fall. Open-source models (DeepSeek V4 Flash at $0.03) may commoditize the value tier to near-zero, breaking the segment economics for closed providers and concentrating value entirely in the frontier tier. The 13x cost premium for 20% capability gain assumes commodity task distribution; enterprise workloads skewed toward high-judgment tasks will find the substitution rate lower.
Cross-references
- Reselling inference at cost-plus yields zero margin as model prices compress: value-based pricing for inference; the tier segmentation model is the demand-side complement.
- Model substitution savings are reinvested in more tokens, not returned to the budget: how model cost savings get reinvested into capability; the tier model shapes where that reinvestment goes.
- AI-native companies invert the traditional software cost structure, spending more on compute than on payroll: cost inversion between model and compute layers; the tier strategy interacts with where margins are held.