ABSTRACT
This memorandum forecasts the AI economy by elimination rather than projection. For each question it enumerates the candidate answers, computes for each the time its binding constraint would need to deliver the required expansion, and compares that against the horizon remaining. Six constraints are derived from a question that never mentions artificial intelligence: what does it take to obtain one more unit of a limiting input? Only the first, physical law, licenses deductive elimination. The other five reprice, which is why every verdict carries a probability rather than a proof. The engine decides eight of the thirty questions here, and that ratio is reported as a result: it is a test for which questions have physically determinate answers, and most do not. Turned on the field's own forecasts, the method produces its most consequential result. Every published estimate of AI's contribution to output is a bet on an unmeasured elasticity between deployed compute and attributable output. Its observed values span a range that reverses all of them, including this memorandum's own GDP band, which the engine eliminates at the low end. The aggregate question is therefore underdetermined on current public data. Two further findings cost the paper more than they gave it. The Residual Ratio, this framework's coined unit, is withdrawn: built from primary sources, none of its four complement components is constructible, so the ratio never had a numerator. And a pre-registered backtest records a maximum-confidence miss, on the one constraint tier the framework had already named as most exposed. What survives is smaller and sourced: the price of a fixed capability falls four to fifty-six times faster than the price of the best available model, a binding constraint attracts no supply response where a cycle does, and one dated call is staked with its adjudication rules specified in full.
KEY FINDINGS
As intelligence commoditizes, value migrates to whatever intelligence cannot make abundant. The binding constraint on the 2026 buildout is a transformer, not a chip and not a dollar.
Every published forecast of AI's contribution to output is a bet on one unmeasured elasticity. Its observed range reverses all of them, including this paper's own.
The Residual Ratio, this framework's coined unit, is withdrawn. None of its four complement components is constructible from disclosed data, so it never had a numerator.
A pre-registered backtest records four hits and one maximum-confidence miss, on the tier the framework had already named as most exposed to substitution. Mean Brier 0.228.
Reliability, not capability, remains the binding constraint on agents through 2027. On 31 December 2027 the best generally available model's METR 80% task-completion horizon is under eight hours. P = 0.85.
Abstract
This memorandum forecasts the AI economy by elimination rather than projection. For each question it enumerates the candidate answers, computes for each the time its binding constraint would need to deliver the required expansion, and compares that against the horizon remaining.
Six constraints are derived from a question that never mentions artificial intelligence: what does it take to obtain one more unit of a limiting input? Only the first, physical law, licenses deductive elimination. The other five reprice, which is why every verdict carries a probability rather than a proof. The engine decides eight of the thirty questions here, and that ratio is reported as a result: it is a test for which questions have physically determinate answers, and most do not.
Turned on the field's own forecasts, the method produces its most consequential result. Every published estimate of AI's contribution to output is a bet on an unmeasured elasticity between deployed compute and attributable output. Its observed values span a range that reverses all of them, including this memorandum's own GDP band, which the engine eliminates at the low end. The aggregate question is therefore underdetermined on current public data.
Two further findings cost the paper more than they gave it. The Residual Ratio, this framework's coined unit, is withdrawn: built from primary sources, none of its four complement components is constructible, so the ratio never had a numerator. And a pre-registered backtest records a maximum-confidence miss, on the one constraint tier the framework had already named as most exposed. What survives is smaller and sourced: the price of a fixed capability falls four to fifty-six times faster than the price of the best available model, a binding constraint attracts no supply response where a cycle does, and one dated call is staked with its adjudication rules specified in full.
"Superintelligence may or may not arrive on schedule; the electricity bill, the memory, and the trust will arrive regardless."
What This Paper Withdrew
The Residual Ratio, its own coined unit, withdrawn for having no constructible numerator. A published maximum-confidence backtest miss, kept on the register with its Brier score. Four fabricated citation identifiers, removed. Eight verdicts flagged where the arithmetic does not discriminate and the stated probability is substantially a prior. The corrections are the record, not an appendix to it.
Method: The Six Immovables
The via negativa method does not project. It eliminates. For each question, every candidate answer is tested against six constraints derived from a single question that never mentions artificial intelligence: what does it take to obtain one more unit of a limiting input? The answers partition exhaustively, and they order themselves by how fast each one can supply that unit.
Tier 0, Thermodynamics. Physical law. Never relaxes. Tier 1, Talent and Absorption. Generational time. Ten to forty years. Tier 2, Data. An accumulated stock, drawn down and not replenished. Tier 3, Matter. Industrial production. Two to seven years. Tier 4, Law and Legitimacy. A collective decision. Three to twenty-four months. Tier 5, Capital. A price. Days to weeks.
Only Tier 0 bars a candidate outright. The other five reprice it as late, expensive, or contingent on a decision not yet taken, which is why every verdict below carries a probability rather than a proof. The gap between what a candidate requires and what the binding constraint permits is the slack ratio; below one, the constraint must expand faster than it has ever expanded for the candidate to survive.
This has a structural advantage over projection: it is falsifiable by construction. Each surviving answer carries a signpost, an observable that would move the verdict if triggered. The register below tracks all twenty on a published schedule.
"The slack ratio: what the binding constraint permits, over what a candidate requires. Below one, and the answer needs a wall to move faster than walls move."
The Call
ONE DATED, FALSIFIABLE PREDICTION -- P = 0.85
Reliability, not capability, remains the binding constraint on AI agents through 2027. On 31 December 2027, the best generally available AI model's METR 80%-reliability task-completion time horizon is under eight hours.
Adjudicated on METR's public time-horizon leaderboard. Falsified if, on or before 31 December 2027, any generally available model posts an 80%-reliability horizon of eight hours or more. The 80% horizon, not the widely cited 50% horizon, is the standard, because it is the level at which work can actually be delegated.
Resolves 31 December 2027 | Metric: METR 80% horizon | Falsifies at: 8 hours or more
Signature bet: By 31 December 2027 the combined market value of the three leading high-bandwidth-memory makers exceeds the combined valuation of the two leading frontier labs. Value accruing to the complement rather than the intelligence.
The Live Signpost Register
Every prediction is operationalized as a tracked trigger. A forecast that cannot fail is not a forecast. The register below tracks all twenty signposts on a published schedule.
| ID | Signpost | What is measured | Triggers when | Source | Cadence |
|---|---|---|---|---|---|
| S1 | Data wall | Frontier dataset size vs. Epoch stock; AI-content share of web | Datasets exceed ~100T tokens, or contamination passes ~90% of new pages | Epoch; Ahrefs | Semiannual |
| S2 | Agent reliability | METR 80% horizon; production single-task success | 80% horizon reaches multi-hour AND production success above 90% | METR; enterprise data | Quarterly |
| S3 | Financing cascade | AI-linked credit spread; correlated defaults | Forced refinancing failure at a top-5 buildout, or spread above 150bp | BIS; issuer filings | Monthly |
| S4 | Compute control | Incumbent accelerator revenue share | Share falls below ~70% as custom silicon scales | Earnings; analysts | Quarterly |
| S5 | Model commoditization | Open-weight vs closed frontier gap; price per token | Open-weight reaches frontier parity, or price decline halts | Epoch; API price sheets | Quarterly |
| S6 | Liability regime | AI-agent liability doctrine and statute | A bespoke AI-agent liability statute, or ruling shifting liability to developers | Official Journal; case law | Semiannual |
| S7 | SaaS repricing | Seat counts; outcome-pricing share; AI-native ARR | Incumbent seats fall >15% YoY, or AI-native ARR growth stalls >50% | Earnings; private-market | Quarterly |
| S8 | Humanoid economics | Unit cost; deployed unit count | Build cost below ~$30k AND deployed base above ~100k units | Manufacturer disclosures | Semiannual |
| S9 | AV mainstreaming | Paid rides/week; number of open metros | Above ~2M weekly rides across 30+ metros, or vision-only unsupervised launch at scale | Operator disclosures | Quarterly |
| S10 | Capital gap | Attributable AI revenue run-rate vs. Bain path | Annual AI revenue run-rate above $400B by 2027 | Earnings; Bain | Quarterly |
| S11 | Power ceiling | US interconnection median wait | Median wait falls below 3 years | LBNL Queued Up | Annual |
| S12 | Distributional backlash | Policy: displacement tax, moratoria | Any G7 AI-displacement tax or deployment moratorium enacted | Legislative trackers | Quarterly |
| S13 | The Call | METR 80%-reliability time horizon | 80% horizon reaches 8 hours by 31 Dec 2027 (falsifies the call) | METR | Quarterly |
| S14 | Thermodynamic headroom | Joules per unit output against the practical CMOS ceiling | Headroom falls below one order of magnitude | Published efficiency benchmarks | Annual |
| S15 | Talent stock and flow | Inbound AI researchers to the US | Inbound flow recovers above its 2017 level | NSF; visa data | Annual |
| S16 | Tier stability | Whether observed expansion rates preserve the tier ordering | Any tier's rate crosses the tier above it for two consecutive years | Cross-tier rate tracking | Annual |
| S17 | Correlated stress | Simultaneous movement in two or more Tier 3 sub-constraints | Two sub-constraints tighten in the same quarter | Supply chain indices | Quarterly |
| S18 | Absorption | Realized productivity growth against the electrification precedent | Annual gain exceeds 1.0pp for two consecutive years | BLS; OECD | Annual |
| S19 | The Elasticity Gap | Implied elasticity of attributable output to deployed capacity | Alpha stabilizes outside 1.2 to 1.5 for two consecutive years | Earnings; academic lit | Annual |
| S20 | Tier 3 classification of power | New-entry response in capacity auctions | A market clears with substantial new entry at or below the cap | PJM; FERC filings | Annual |
Citation
Scott, Sidney. "Via Negativa: The AI Economy by Elimination." The Ashby Institute, 2026. arXiv:XXXX.XXXXX.
Published by The Ashby Institute. This memorandum is analysis, not investment advice. Forward-looking statements are uncertain and may prove wrong; that is the purpose of the register above.
CITATION
Sidney Scott, Via Negativa: The AI Economy by Elimination. The Ashby Institute, July 2026. TAI-WP-2026-02. arXiv:XXXX.XXXXX (preprint).