The SAVRN Scarcity Tracker
Eleven primary-sourced indicators, one composite index, read against what a GPU-hour actually bills as tokens. Supply tightness, equipment and power lead time, and demand and price confirmation, each scored 0 to 100 against a Balanced-to-Acute scale.
What the same scarcity reads against
The token index is the market's blended asking price. SAVRN's own basis and the frontier ceiling are read live from the SAVRN Index, not duplicated here. The scarcity-implied price is where the index says the token index is headed.
What is actually scarce, group by group
Each indicator is scored 0 to 100 against a calibrated Balanced-to-Acute anchor, weighted, and rolled into the composite index above. Every value carries its own source.
Supply tightness
Group score 87.9| Indicator | Value | As of | Score | Weight | Source |
|---|---|---|---|---|---|
| Primary-market vacancy | 1.4% | 2026-06-30 | 94.3 | 20% | Source ↗ |
| Under-construction capacity preleased | 80.4% | 2026-06-30 | 80.8 | 15% | Source ↗ |
| Months of available future capacity | 6months | 2026-06-30 | 85.7 | 10% | Source ↗ |
Equipment & power lead time
Group score 82.7| Indicator | Value | As of | Score | Weight | Source |
|---|---|---|---|---|---|
| Gas turbine order-to-delivery lead time | 5years | 2026-03-23 | 75.0 | 10% | Source ↗ |
| Gas turbine plant price | 3000$/kW | 2026-03-23 | 75.0 | 5% | Source ↗ |
| Power transformer unit cost vs 2019 | 77% | 2025-08-31 | 77.0 | 5% | Source ↗ |
| GE Vernova gas backlog + slot reservations | 116GW | 2026-06-30 | 90.5 | 8% | Source ↗ |
| ERCOT large loads seeking interconnection | 474GW | 2026-06-30 | 94.2 | 7% | Source ↗ |
What share of a megawatt's annual output pays the rent
Two disclosed wholesale contracts, three throughput points, five price bases. Each cell is contract rent ÷ (price × tokens a megawatt produces in a year). Above 100% means the contract's own rent cannot be paid back by that basis at that throughput.
| Contract | Throughput point | Token index | Scarcity-implied | SAVRN | Retail median | Frontier |
|---|---|---|---|---|---|---|
| Oracle – OpenAI | GB300 NVL72 @ 100 tok/s/user | 76% | 32% | 20% | 29% | 2% |
| Oracle – OpenAI | Vera Rubin NVL72 @ 100 tok/s/user | 35% | 15% | 9% | 13% | 1% |
| Oracle – OpenAI | GB300 NVL72, MLPerf throughput | 12% | 5% | 3% | 5% | 0% |
| Monarch – Anthropic | GB300 NVL72 @ 100 tok/s/user | 92% | 39% | 24% | 35% | 2% |
| Monarch – Anthropic | Vera Rubin NVL72 @ 100 tok/s/user | 43% | 18% | 11% | 16% | 1% |
| Monarch – Anthropic | GB300 NVL72, MLPerf throughput | 15% | 6% | 4% | 6% | 0% |
How the index is built. Each indicator scores linearly between a Balanced anchor and an Acute anchor, clipped 0 to 100. The composite index is the weighted average of every indicator actually observed this reading; a missing series drops out rather than counting as zero. Regime bands: Balanced under 40, Tight from 40, Scarce from 70, Acute from 85.
Where SAVRN's own numbers come from. The SAVRN and Frontier price bases are read live from the SAVRN Index's own tracked data, the same figures used throughout What AI Costs, not duplicated or separately estimated here.
Refreshed on a weekly schedule. Every value traces to a dated source with a verbatim excerpt; nothing is entered without one.
Seven references, one system
Capital, delay, policy, water, grid, permits and scarcity are one system seen from seven sides. Each reference is primary-sourced, refreshed on a schedule, and links into the others, so a project can be followed from the money committed to the docket that stopped it.